Operationalizing AI for Scale and Sovereignty

Companies are taking control of their own data to tailor AI for their needs. The challenge lies in balancing ownership with the safe, trusted flow of high‑quality data needed to power reliable insights. This conversation from MIT Technology Review’s EmTech AI conference examines how AI factories unlock new levels of scale, sustainability, and governance—positioning data control as a strategic imperative for governments and enterprises.

About the speakers

Chris Davidson, Vice President, HPC & AI Customer Solutions, HPE

Chris Davidson is Vice President of HPC & AI Customer Solutions at Hewlett Packard Enterprise. He leads HPE’s global strategy for AI Factory solutions and Sovereign AI, working with governments, enterprises, and research institutions to build secure, scalable national- and enterprise-grade AI capabilities.

He also directs Product Management and Performance Engineering across HPE’s HPC and AI portfolio, including large-model training platforms and Cray exascale systems. His teams define product strategy, performance architecture, and deployment models that position HPE at the forefront of high-performance and AI computing.

During his nine years at HPE, Chris has led key initiatives across Performance Engineering, AI Cloud, and Professional Services, shaping how HPE delivers optimized, cloud-native, and globally deployed high-performance systems. He previously held technical and leadership roles in the biotech and medical diagnostics sectors.

Chris holds an M.B.A. in Entrepreneurship and Finance and a B.S. in Biology from Loyola University Chicago.

Arjun Shankar, Division Director, National Center for Computational Science, Oak Ridge National Laboratory

Mallikarjun (Arjun) Shankar is the Division Director for the National Center for Computational Science at the Oak Ridge National Laboratory. His research focuses on the interdisciplinary bridge between computer science and large-scale scientific discovery campaigns that rely on scalable computing and data science. He is a joint faculty appointee at the University of Tennessee’s Bredesen Center, a senior member of the IEEE and a senior member of the ACM.

Cyber-Insecurity in the AI Era

Cybersecurity was already under strain before AI entered the stack. Now, as AI expands the attack surface and adds new complexity, the limits of legacy approaches are becoming harder to ignore. This session from MIT Technology Review’s EmTech AI conference explores why security must be rethought with AI at its core, not layered on after the fact.

About the speaker

Tarique Mustafa, Cofounder, CEO, and CTO, GC Cybersecurity

Tarique Mustafa is Cofounder and CEO/CTO of two AI-powered cybersecurity companies: GCCybersecurity, Inc. and its data compliance spinout, Chorology, Inc. A prolific inventor and internationally recognized authority in knowledge representation, inference calculus, and AI planning, Tarique has spent his career applying autonomously collaborative AI to solve complex, ultra-high-scale challenges across cybersecurity, data security, and compliance — with deep expertise spanning Data Classification, DLP, and DSPM industries. His groundbreaking innovations and multiple USPTO patents have earned him global recognition, including frequent invitations to deliver keynote addresses at prestigious international security conferences and forums.

At GCCybersecurity, Tarique architected the core AI algorithms powering the company’s 4th and 5th generation fully autonomous data leak protection and exfiltration platform — among the most advanced platform of its kind. Prior to founding GCCybersecurity and Chorology, he served as founding CEO/CTO of NexTier Networks, a Silicon Valley provider of award-winning Data Leak Prevention solutions. With over 20 years of technical leadership experience, Tarique has held senior roles at Symantec, DHL Airways IT, MCI WorldCom, EDS, Andes Networks, and Nevis Networks, where he served as Principal Architect and built industry-leading security products leveraging next-generation security monitoring, event correlation, IDS/IPS, and SSL/IPSec technologies.

Tarique holds multiple approved and pending patents with the USPTO and has authored numerous research publications spanning Information & Data Security, Computer & Network Security, Software Architecture, Database Technologies, and Artificial Intelligence. A recipient of the prestigious Rotary International Scholarship for doctoral studies in Computer Science at the University of Southern California (USC), Tarique also holds master’s degrees in engineering and computer science from USC, and a bachelor’s degree in mechanical engineering from NED University of Engineering & Technology.

Rebuilding the data stack for AI

Artificial intelligence may be dominating boardroom agendas, but many enterprises are discovering that the biggest obstacle to meaningful adoption is the state of their data. While consumer-facing AI tools have dazzled users with speed and ease, enterprise leaders are discovering that deploying AI at scale requires something far less glamorous but far more consequential: data infrastructure that is unified, governed, and fit for purpose.

That gap between AI ambition and enterprise readiness is becoming one of the defining challenges of this next phase of digital transformation. As Bavesh Patel, senior vice president of Databricks, puts it, “the quality of that AI and how effective that AI is, is really dependent on information in your organization.” Yet in many companies, that information remains fragmented across legacy systems, siloed applications, and disconnected formats, making it nearly impossible for AI systems to generate trustworthy, context-rich outputs.

“Really, the big competitive differentiator for most organizations is their own data and then their third-party data that they can add to it,” says Patel.

For enterprise AI to deliver value, data must be consolidated into open formats, governed with precision, and made accessible across functions. Without that foundation, businesses risk “terrible AI,” as Patel bluntly describes it. That means moving beyond siloed SaaS platforms and disconnected dashboards toward a unified, open data architecture capable of combining structured and unstructured data, preserving real-time context, and enforcing rigorous access controls. When the groundwork is laid correctly, organizations can move toward measurable outcomes, unlocking efficiencies, automating complex workflows, and even launching entirely new lines of business.

That value focus is critical, says Rajan Padmanabhan, unit technology officer at Infosys, especially as enterprises seek precision in the outputs driving business decisions. Rather than treating AI initiatives as isolated innovation projects, leading companies are tying AI deployment directly to business metrics, using governance frameworks to determine what delivers results and what should be abandoned quickly.

“We see this big opportunity just with AI literacy with business users, where they’re very eager to understand how they should be thinking about AI,” adds Patel. “What does AI mean when you peel the covers? What are the pieces and the building blocks that you need to put in place, both from a technology and a training and an enablement standpoint?”

The possibilities ahead are substantial. As AI agents evolve from copilots into autonomous operators capable of managing workflows and transactions, the organizations that win will be those that build the right foundation now.

“What we are seeing as a new way of thinking is moving from a system of execution or a system of engagement to a system of action,” notes  Padmanabhan. “That is the new way we see the road ahead.”

The future of AI in the enterprise will be determined by whether businesses can turn fragmented information into a strategic asset capable of powering both smarter decisions and entirely new ways of operating.

This episode of Business Lab is produced in partnership with Infosys Topaz.

Full Transcript:

Megan Tatum: From MIT Technology Review, I’m Megan Tatum, and this is Business Lab, the show that helps business leaders make sense of new technologies coming out of the lab and into the marketplace.

This episode is produced in partnership with Infosys Topaz.

Now, recent advancements in AI may have unlocked some compelling new industrial applications, but a reliance on inadequate data models means that many enterprises are hitting a brick wall. AI and agentic AI in particular place a whole new set of demands on data. The technology requires greater access, context, and guardrails to operate effectively. Existing data models often fall short. They’re too fragmented or siloed. Data itself often lacks quality. To bridge the gap, they require an AI-ready upgrade.

Two words for you: data reconfigured.

My guest today, are Bavesh Patel, senior vice president for Go-to-Market at Databricks, and Rajan Padmanabhan, unit technology officer for data analytics and AI at Infosys.

Welcome, Bavesh and Rajan.

Rajan Padmanabhan: Thank you. Thanks for having us.

Bavesh Patel: Thanks for having us.

Megan: Fantastic. Thank you both so much for joining us today. Bavesh, if I could come to you first, when we talk about AI-ready data, what exactly do we mean? What new demands does AI place on data, and how does this impact the way it needs to be structured and used?

Bavesh: Yeah. Great question. Appreciate you hosting us today. I think that obviously the whole world is enamored with AI because of all of the power that we can all see as users. AI is now democratized across hundreds of millions of users. And when we think about enterprises and businesses using AI, the quality of that AI and how effective that AI is really dependent on information in your organization, and that’s data. And what we found is that most enterprises, their data is kind of locked away in these different applications and different systems. And it’s very difficult to get a good view of, what is all my data? How trustworthy is it? How recent and fresh is it? And all of that is being injected into the AI. Unless you have a proper understanding of your data, the ability to ensure that it’s data that’s accurate and that can be used so that the AI can take advantage of it, you’re actually going to end up having terrible AI.

We see a lot of customers spend time on cleansing their data, organizing their data, making sure it’s access controlled correctly, and that tends to be the fuel of good AI.

Megan: Yeah. It’s such a foundational thing, isn’t it? But it can be missed, I think, quite easily. Rajan, what difference can having AI-ready data really make for enterprises as they unlock that full potential of AI and its applications?

Rajan: First and foremost, thanks for having us. It’s a pleasure. I think in continuation of what Bavesh talked about, see, data and AI is pretty synonymous. And similarly, the consumer AI and enterprise AI and enterprise agentic AI are different because first and foremost, the business needs to have the context. That context from your enterprise information, which is not only structured, both structured and unstructured and user-generated contents and all forms of data is going to be very, very critical to really get the context right, and really get any model that you pick. That’s where the platforms like Databricks really help with the plethora of models or whether you want to build your own models or whether you want to ground the model based on your data. That is going to be very, very critical. That is where getting the data for AI is going to be very, very critical.

The third critical part, and this actually will be one of the roadblocks for adoption of AI. That’s why if you see the AI adoption on the consumer side is skyrocketing, but on the enterprise side, the enterprises are struggling is primarily around the precision of their output, because you are taking a business decisions where you are taking a buy decision, you are taking a sell decision, or you are trying to recommend something, recommend the content. It could be 20 different use cases. For that, the precision is going to be very critical. We are seeing our customers, the successful customers, definitely for the precision to be more than 92% is not aspiration, that is a must-have. If you have that, definitely being that AI data is going to be the entrepreneur right now for that.

Megan: And I suppose if we’ve outlined there how critical this is, where should enterprises start then, professional perhaps, the level, what are the foundations when it comes to building an AI-ready data model?

Bavesh: Yeah. And I think Rajan hit the nail on the head. I mean, enterprises are grappling with a different set of problems than consumer AI. The first thing is that you’ve got to get a handle on your data. As I mentioned, a lot of the data is locked in. Ensuring that you have ability to put your data in a place where you can understand the holistic view of as much of your data as possible. That kind of starts with putting your data in open formats. A lot of the valuable data today in an organization is locked away in some proprietary SaaS app or some system, and all the datasets aren’t connected together to form that context. The first step is to really do an analysis of what is your data estate? What are the critical pieces of data that need to be put into a place where you can start to understand them and how they’re connected to one another?

Thinking about how do you set up your data catalog, thinking about how do the relationships between the data assets work, putting data governance around it, that seems to be the first step. And if you think about how ChatGPT was built, it took all the data on the internet and then aggregated it, synthesized it, and then built these transformer models, while enterprises, they don’t really have a handle of all their data within the organization. That’s the first foundation that you really want to think about. The second thing is that you don’t want to just go ad hoc, go and do random AI projects. You really need to be thinking about business value. A lot of our customers are looking at AI much more strategically in that they want to be able to get projects on the board with wins and then generate business value.

Building an AI value roadmap, which is connected to how well your data is organized, those two things seem to be foundational to how do you launch AI successfully in your organization.

Megan: That value piece is so important, isn’t it? And as I understand it, Infosys and Databricks have worked closely together to guide organizations through this transformation. I wondered, can you share some examples of the impact you’ve seen enterprises you’ve worked with, Rajan, what difference has it made to the ways in which they can integrate more sophisticated AI and agentic AI applications?

Rajan: Well, that’s a very, very good question. What both Databricks and Infosys has done is we have come up with, a kind of a framework first. First and foremost, it all needs to start with the value. One of the largest food products company where we collaborated together, what we have done is we have applied this framework. The framework consists of six different things. First and foremost, very critical is the value management, which Bavesh touched upon. We have worked together to come up with a 3M measurement framework, what we call adaptability, business value, and then responsible. You can’t just go and do a garage project. It has to be measurable. It should be responsible, follow all those things. That is going to be very critical. And we helped this client to prioritize, which will give them the most value for money, the investments that they are making.

The second critical part here is it is not like most of the enterprises today are not everybody’s AI-born companies. Most of them were born during analog days; most of them were born in digital days. There are companies which are applying AI for modernization, because a lot of your historical information, which is actually helping you to build that long-term context. And that is where we have worked closely with some of the native tools of Databricks, like Lakebridge or the AI assistants that are there, and then create composable services on top of it to help the clients unlock the value bringing into Databricks. And then the second part where we help the client is exactly to the point, the readying of data. Now you brought in the data, now you have to bring both the structured, unstructured, analytical and all these aspects.

And that is where the third layer, we closely work with the Databricks, which is part of leveraging all the great capabilities within the Databricks, be it Unity Catalog, be it the open formats, or be it the gateways and other aspects. We were able to make the data available for this client. What has really helped our client, the third part, is Agent Bricks, which is one of the differentiatiors. It gives you the flavor for the enterprise. That is where we have closely worked, and we built some of our industry-specific agents, be it CPG, be it energy, be it FS. And for this client, what we have done is we have taken some of those CPG-specific use cases. Either it could be on the HR space or the procurement space or on the marketing space. And this has really helped our client be able to build a business capability surrounding this and unlock eight to nine use cases, we call it as a products, agentic AI products, which can really drive more value for them, solving the real business problems.

And this kind of a comprehensive set of frameworks plus set of suites of services, plus our solution assets, Infosys solution assets, as well asunlocking the value from Databricks has really helped these clients. And we see similar patents for a lot of these successful engagements where we were able to continuously drive the value by applying this framework actually.

Megan: Right. Sounds like it made a real material difference. Rajan mentioned a few of the tools in Databricks catalog there, Bavesh. I know you’ve recently worked to launch an operational database for AI agents and apps. I wonder how does a platform like that help organizations in this journey? What makes it different from some of the other platforms out there right now?

Bavesh: Databricks has come to market with a new offering called Lakebase, which is really an OLTP database where you can build your AI apps. And if you think about it, there’s really two main types of data in an enterprise. There’s all the historical data, which is all the things that have happened, and that’s really what your analytics is based on. You have an old app system where you have put all your historical data and Databricks has come to market with what we call the Lakehouse, which is essentially a data warehouse with all of your data that is not operational in nature. It’s historical data. And I think that Lakehouse concept is really pushing forward with AI because a lot of our customers have thousands of users within their business and they need to get data. And what they’ve done is they’ve actually gone down the BI route, which is really building a dashboard or a report.

Most organizations have had thousands of these dashboards and reports proliferate across the organization and then they need to be customized. It just takes a long time for users inside of the business to actually get access to the data. AI now is really making that a lot easier from just the analytics perspective where we can now democratize access to the data, which has really been the holy grail for most data teams. They really want to get out of the way and just give the right data to the right people inside of the business with the right access.

With a product like Genie at Databricks, you can just use English language or whatever your language is to ask questions of the data. And it’ll give you back data that answers your questions in context. It’ll give you not just what ChatGPT will give you, which is information about a topic that’s on the internet, but it will actually tell you, “Well, why did my sales numbers not reflect what I expected in the month of April?”

It’ll give you some root cause analysis based on your enterprise data. Genie is going to be one of these things that’s really important where it’s going to truly kind of democratize data inside of the business. That’s kind of this OLAP world, which is what the Lakehouse is. More recently, we’ve come to market with what we call the Lakebase, which is the OLTP world. What we’re finding is that agents are now being deployed in these organizations, and those agents need a place to keep all of their orchestration, all of the context of what’s happening in that particular workflow. On the one hand, you’ve got users just asking questions. On the other hand, the next chapter is going to be around automating an entire business process. If you’re taking a function like generating a campaign in marketing, right? There are a lot of tools you use and a lot of steps you use.

An agent can come in and really automate a lot of that. But on the back end of that agent, you’re going to need to stand up a real-time database to keep track of all the things that the agent is doing. That’s what Databricks has brought to market, which is this OLTP Lakebase solution. The innovation that we have brought to market is that it’s a modern kind of Postgres database where we have separated the compute and storage, very much like what we did with the data Lakehouse with the data warehouse. But on the Lakebase, the data is on one copy inside of your cloud storage, and then the compute is separated and it’s serverless. You can do things like branching and you can start up the OLTP database really quickly. What we found is that agents are actually starting these Lakebases because they can very quickly go start one up, keep it running, put it down when it needs to, make a copy of it.

Agents are doing this, then they need the velocity, they need a cost-effective solution. And the beauty of all this is when you take the OLTP, which is all around the Lakebase and the real time, and you take the OLAP, you now have one system for all your data. You don’t have to copy the data around, you don’t have to manage all the permissions, you can set the context against it. We see these AI apps being really the future of how businesses run, where they’re going to take away all of the bottlenecks that humans are having to do repetitive work and automate these using LLMs and all these new technologies. We want to be the default for powering all that because we believe that our Lakebase technology is going to be faster, cheaper, and more secure for an AI database.

Megan: Sounds like a real game changer. And we’ve touched on this a couple of times already, I mean, this idea of value. We know that engaging the commercial value of investments into AI is really high on the priorities right now for senior leaders. How important is this value measure piece when it comes to creating AI-ready data systems, Rajan? How can organizations ensure they’re monitoring what is delivering and what isn’t?

Rajan: This is the paramount importance and most of the successful AI implementations or agentic AI implementations really required this value measurement. I’ll just extend the client example that I talked about, the large food products company, the global products company, to explain this question. I just want to create a metaphor. When the initial digital world came, we have a lot of these analytics primarily around defining those performance management KPIs, fact-based decisioning and other things were evolving over a period of time. Typically, a lot of these metrics are going to be very critical for them to measure how a function, how a business is doing. On a similar line for the value measurement, if I take the same example of the client, what is very critical for an organization is actually to map your outcome that you are expecting.

Iin this case, how do I optimize my spend on direct and indirect purchases? So by applying AI, I would like to identify the areas where I can optimize the spend. That means one of the critical measures that you have is, what is your indirect expense classification and what spends you have been classified and how much you are able to reduce by bringing in this. Establishing these measures and the metrics is going to be very, very critical. And once you establish these base metrics and the measurement, and the beauty of it is some of these metrics, to just extend what Bavesh was talking about, the capabilities that Databricks gives you, like metrics view, features, tools, and other things would actually help you to translate those AI telemetries, business telemetries that is coming from your applications into a measurable metrics in terms of an outcome, which you can actually measure using the Genie room for value management measurement.

Then what happens is two things that you can take, the use case, the products that as I said for this client, the products that we build either on the procurement side or on the marketing research side, if you find there is a value either because of VAC, they identify that they’re able to optimize or it is able to reachability, what is the reach, you can either accelerate that use case and further fine tune that product to expand it. Or there are, if you find it is not really driving the value or I’m not able to see the value that it is going to deliver, you can very well do a fast failure method rather than trying to make it work, you can understand and then you can take a call to pivot it to something else different.

There are three aspects here. What we see from our experience, not only with this client across some of our other clients from industrial manufacturing or FS or in the energy, is by setting up this metrics-driven valuation method upfront and then leveraging the capabilities to establish, transform these telemetries, signals into a measurement, what we call an AI compass room so that you really measure the business stakeholders, whether it is coming from a marketing office or whether it is coming from supply chain office or whether it is coming from a CFO office where they can say, “Hey, this is what it is intended to do, this is what the current measurement, and this is where it’s failing that can help them to pivot.” And this will actually drive and democratize AI, all the agent decay across the enterprise, and that really drives the value.

This is going to be one of the critical part that enterprise needs to do it. And that is where the six part framework that I talked about, applying that framework like value office, applying the ready for AI, applying the transformation fabric. Then the third part is the governance, which is going to be the entrepreneur of this. Then running your operations, not based on SLA, based on the experience level agreements and business metrics for you to continually measure, bringing all these six layers is going to be very critical. That’s when we see the organizations are very successful, and some of our proven examples exactly do the same that this is going to be very critical for organizations from a measurement standpoint.

Megan: Lots of tangible ways there that you can actually gauge value here. And you touched on governance and the impact of AI on governance is another huge talking point among senior leaders and interactions with data are a core part of that. To what extent is having the right governance and security protocols an integral part of having AI-ready data? To Bavesh, what scenarios do these systems need to handle? What does that mean for data models?

Bavesh: This is becoming kind of the prerequisite to deploying a successful AI project. I think MIT produced a report that said 95% of these new AI projects fail to actually generate business value. A big reason for that is you can go and prototype and stand up and vibe code a pilot, but when you’re actually moving a workload into production, you realize that governance becomes so critical.

So what do we really mean by governance? I think the first thing is getting your data in order, like I said, in open formats. Most companies realize now that the way they engage with their customers, the way they develop a drug, the way they approve a person for a credit limit increase, all of that enterprise information is actually their competitive advantage. Because you can go and use a frontier model like ChatGPT or Claude that everybody has access to.

Really the big competitive differentiator for most organizations is their own data and then their third-party data that they can add to it. Getting your data into an open format so you can understand your data and understanding your data is where governance comes in. Because when you think about governance, you really want to be able to find the data.

If I’m an end user or if I’m building an AI product, I want to know what data’s available to me. Can I trust the data? How fresh is the data? Is it coming from my analytics world or do I need a real-time system like a OLTP system? I need to find the data. I also need to make sure that access is controlled in a way that doesn’t cause any huge headaches from my organization. This becomes critical. If I have a whole bunch of PDFs that have purchase orders in them, who actually has access to all that data?

In a clinical trial, for example, in healthcare, you really want to ensure that people across trials don’t have visibility to patient data. Maybe the model that was used to build that was running across trial. Who has access to all the data? Who has access to only parts of the data? You really have to think about this. We also look at semantics of the data. Rajan brought this up right at the beginning of this, which is what is the context? How do we think about the metrics and all the things that the business users know in their head? We need to start codifying that somewhere. We have a product at Databricks called Unity Catalog where you can do the discovery, the access and the business semantics. You also want to share the data.

And in the world of agents, what we see is something called agent sprawl. In a very short order, just like how SaaS applications became very prevalent within any organization where they really solved a business problem. You go to a line of business and you say, “I need to be able to do credit underwriting” or “I am doing a prior authorization use case or pick thousands of use cases.” There’s a SaaS app for that. Much like that, there’s going to be this world in which agents are going to come into play, and most organizations are going to have lots of agents running all the time, but the reality of it is that how did that agent perform? What was the feedback loop from the user? What was the cost of running that workload and is it going up dramatically? And if you don’t have a way to monitor, to understand, and trace all the questions and answers and responses at scale, you’re going to find yourself in a big pickle. This actually could hurt your organization because users will be very confused about what to do.

When you look at governance, most organizations are recognizing that they have to start to understand what is it that they have put in place from a systems, from a process, from a tooling standpoint, focus on one use case, build out the governance for that, but build it in a way that’s going to allow you to become repeatable. AI is not going to be about one use case or two use cases. It’s whoever builds the flywheel of building many use cases in a safe, secure way, in a cost-effective way that’s driving a business outcome. If you don’t apply governance, it’s going to be very hard.

At Databricks, we made a big bet on governance four or five years ago. This is one of the main reasons our company’s growing right now because we can ensure that there’s quality data that’s going into all of your AI. You can use things like Genie and you can use things like Agent Bricks and you can build apps using Lakebase. None of that really works without governance. It’s really what we call the brain inside of Databricks.

Most of our customers spend a lot of time inside of Unity Catalog. And the great news is that AI is helping governance get set up much more quickly. We have a customer that three years ago, they were trying to get all of the data assets across all their domains from the customer, from the loyalty app, from the e-commerce engine. They had to go and map out all this data assets. AI is now doing a lot of their work for them. The human in the loop is just checking things.

We’ve made this much easier with AI. We always think about AI as a business use case and an outcome, which I think is going to be where the biggest value is. But at Databricks, we’re using AI inside of our platform to make it much easier to operate and to make it much easier to provide all the right things for your business. This is a super critical part of how we plan to innovate as AI takes fruition in the market.

Megan: And Rajan, Bavesh touched on this a little bit there, but does the integration of Agentic AI add another layer of complexity here too? What new consideration around governance does that raise?

Rajan: That’s a very, very valid question. I would like to take a metaphor to really explain. We are getting into the world of self-driving cars, robotaxis, and other things. While that takes us to the autonomous world, but still there are rules that you need to adhere to when you are driving on a road. The reason I’m bringing this metaphor is because what is actually required is actually adhering to the rules and different topographies, different things, depends upon where you are driving is going to be very, very critical. The complexity that agents are going to add is basically how you operate with those constraints.

For example, as a UTO, I can do 10 things, but say if I cannot approve a discount for more than 70% or I cannot give something as a bonus for someone because that is a part of the CFO, which an agent should be aware of.

That is one aspect, applying the constraints around it and making sure that the agents are adhering to the constraints. The second set of complexity that it builds is the tools to access. As a business, in today’s world, when you define a process, certain processes need a certain set of tools to really actionize it. There are certain entitlements, only people entitled to do certain things based on their identity, based on the need or the situation need, you need to govern. The third is information sharing. While MCP and other aspects are great, UCP and other aspects are great, but one critical thing is what you need to share, what you don’t need to share. And those are the critical considerations.

The last part is learning and relearning. Sometimes when you learn good things, you should keep something. Sometimes it is better for you to completely remove it and reevaluate in a newer way, relearn it in a newer way. These are all the critical things that are required. On the similar line for agents, it is going to be paramount, because when you are operating agents for an enterprise, you need to know, learn, and adhere to certain compliance related rules, business related constraints, and then the entitlement identity, and then sharing whatever that apply to a physical human will also start applying to an agent. That is where this is going to be very critical. This requires a new set of operating systems. That doesn’t really mean now get out of a new thing. That is where I’m just interpreting how Bavesh touched upon the Unity Catalog.

The best part that which we see and some of our clients that which are implementing is extending the Unity Catalog and the capabilities like now you can catalog the tools, catalog the MCP as well as catalog these agents, and then govern those agents based on the constraints, ground them based on the constraints.

It’s going to be very, very critical. Doing it not later, but starting that as part of your strategy and enforcing this as one of the critical dimensions of when you measure the value is also going to be very critical for an organization. It is like making sure that not only building the autonomous car, but as well as making sure that the car drives as per the rules of the road, not going rogue.

Megan: Lots to think about there. Fascinating stuff. Thank you. Just to close, with a quick look ahead, we all know the pace of development in AI and Agentic AI is so rapid. For those organizations that can prioritize AI-ready data now, what are the most compelling use cases for the technology that you can see coming to the fore in the next few years, Bavesh?

Bavesh: I think the excitement level is at its peak. We’ve seen so much investment in AI. I think the reason why there’s a lot of excitement is because you can look at the early adopters and you can see massive amounts of gains that these organizations are seeing. The one thing I will tell you is that the companies that there’s really three categories and the companies that I think are doing well, a lot of them started out with just copilots and things that are just giving people quick answers. Think about it as making an individual productive. That is the first phase. And the ROI on that has been somewhat questionable. With something like Genie, it makes it a lot more effective because it’s actually on your data and your data is contextualized in your organization. I think that’s one level of area that we’re going to see a lot of innovation. We’ll see most organizations just start to get the right information to the right person at the right time. And that has been a dream for a lot of organizations.

The second one is around automating entire business processes. We see functions within marketing, like I described earlier, or whether you’re going through a process of rebates for a company. There’s a whole bunch of steps involved where you have to go into three different apps and export data from Excel and put it over here. There’s thousands of people doing very laborious, monotonous, repeatable work. These agents are really going to help get an immense amount of not only productivity for the business process, but it’s just going to make things faster. Processes that took weeks are now going to take days. Processes that took days are going to take hours and minutes now.

One trend we’ve seen is that the AI world is so dynamic. In a world where you got lots of different players, you want to think about first principles, what are the foundations? You want to think about owning your data, making sure you have a handle on your structured and unstructured data. You want to put governance on that. But the other thing that you want to make sure that you don’t do is lock yourself in.

Today, if you think about it, Gemini is really good with multimodal. Anytime you have pictures or videos or things like that, Gemini just is super good. Whereas if you’re writing code, Claude is really good. If you’re just doing certain types of questions around introspection, ChatGPT is really good. What you really want is an open data platform where you can build your open AI on multiple clouds, which is what we built at Databricks.

I think that’ll help with the second piece, which is you can pick and choose because when you build these agents, you don’t have to be locked into just one. You should be picking the best quality and the best security and the best ROI and cost for a particular workload. One workload may use multiple of these models, and they might be even specific industry models. You need a system and a platform that can really handle this complexity.

I think the third category is business reimagination. A lot of people talk about this where, yes, you’re going to go and take the data and make it available and give everybody access to the data. You’re going to make existing processes much more efficient. But the third thing is there’s going to be brand new things that come out of it.

We have a very large customer who’s a bank and they have built a product that they didn’t have a year ago. Essentially, it’s machine learning and LLMs helping treasury departments forecast what their balances are going to be because they have more data at their fingertips. Historically, it took a long time for the data to get to the bankers. They were not able to really predict what a balance would be for a treasury department. Think about this for a big enterprise company, they have now built a brand new data AI solution that they’re monetizing and it’s generated hundreds of millions of dollars in the first six months. We’re seeing brand new lines of business open up and that is going to be really exciting because that’s where a lot of the transformation is going to happen. There’s going to be productivity. There’s going to be kind of automation at the business process level. Then there’s going to be these big new things that we didn’t even imagine that people are going to come up with.

We are actually seeing the early signals of this in every industry. We see retailers getting data at the hourly and the minute level so that they can integrate much more closely with their supply chains. We’re seeing much more targeted customer 360-degree use cases where as retailers or as consumers, we get annoyed by ads, but now it’s so contextualized and you have so much information about what really matters to your target customer, you’re giving them value added kind of information and that’s engaging them more. There’s a whole bunch of innovation happening with agentic commerce and things like concierge and virtualized shopping.

You look at any industry, there’s definitely new ways of doing things. This is what’s really exciting about AI, but you really have to not get too far ahead without thinking about what are the foundational things. You mentioned this earlier, which is open data platform, making sure you have governance correctly, making sure you think about your historical analytical data and your application data that’s going to be real time, having a good foundation to build on, that’s going to allow you to scale and move more quickly and compete in this new world.

We’re very excited about what we’re seeing with our customers and what they’re building. And honestly, that’s the best part about being in my role at Databricks, which is our teams really go to customers and say, “What are the outcomes you’re driving?” The early signals have been super positive. We’re seeing companies that get serious about all the foundational elements and really are methodical about building really outcome-based AI solutions, that 5% of projects that are being successful, those are wildly successful. That’s why we’re growing as a company because once you get a good project under your belt, that gets visibility within executives.

The last thing is that historically, a lot of tech has been in the IT department. You get the business designing how they want to go to market and how they’re going to compete and what products and services they want to offer. IT was the enabler and in many cases became the cost center and was relegated to rationalizing the portfolio of spend and tools.

But now we’re seeing the business kind of take the lead with AI where they want to understand, they want to know, “Hey, what can I be doing now that was not possible before?” We see this big opportunity just with AI literacy with business users where they’re very eager to understand how they should be thinking about AI. What does AI mean when you peel the covers? What are the pieces and the building blocks that you need to put in place, both from a technology and a training and an enablement standpoint? We’re spending a lot of time with executives helping them along this journey. We definitely see a lot of amazing opportunities ahead.

Megan: Yeah. So much innovation going on. And finally, how about yourself, Rajan? What on the horizon is exciting you the most?

Rajan: I think Bavesh covered quite a bit, but I think the way I’m seeing is today predominantly we are talking about labor shift. That means unlocking the potential of human or shifting the current way of working to the new way of working with the more efficiency game. It’s predominantly more of an efficiency game. I think that is what we are seeing now and the majority of the successful use cases around the labor shift. But what is pretty promising is the two kinds of shift, the business shifts.

What we are seeing as a new way of thinking or the new thing that is coming up is moving from system of execution or a system of engagement to system of action. That is the new way we see the road ahead. That is where some of the points that I touched upon. The business wants to have access to it, but how does it really make the real difference for it?

One classical example that I could clearly see which we have implemented for one of our customers primarily in the manufacturing space, is around the lifecycle of creation of a product and then publishing the content around the product in line with their different B2B marketplaces. Some of those, you are not just talking about recommending, creating, but actually you are able to reimagine this process, which used to involve five different departments, now can be done much faster, but at the same time gives you that veracity in terms of the decisioning that you are able to do and as far as how you’re able to actionize. That is the second thing which we are seeing.

The third part I think is also going to be is the way how the commerce has evolved. There is also not beyond that agentic commerce, but I think what we are seeing is that agent to agent commerce, agent to human commerce and agent to agent payments, agent to human payments, and then the content monetization.

These are the new set of business opportunities like building new business agentic products. It could be for family techs, it could be for on the consumer side, or it could be on the industrial technology side. These are going to be what I’m calling the economy shift, labor shift, business shift, because that is going to bring a new set of system of actions, moving them from the system of executions or the typical SaaS application with the bolt-on agentic, the so called agentic application. That is going to be a major transformation, and we are underway. But on the technology side, what is very critical for entrepreneuring is in today’s world you have data, analytical data, operational data, and then there is intelligence, there are different facets of it.

I think both this analytical core and operational core is going to really come into one. That’s why we are so gung-ho about the releases of Lakebase and other things because that is the way the future is going to drive. When they are really thinking about being ready for AI technology use cases, they should really think, how do you really create this unified core for the newer world?

The second part is people have to reimagine today, if I take SAP as an example, you do hundreds of edge applications, business applications needed to integrate another thing. Typically, we create sprawl of these integrations. One technology use case, people can say, “Hey, how do I really create a domain-based service mesh on top of this unified core and how do I make it more agentic integration ready?” That is one of the technology use cases that we are advising to the client.

I think now with a lot of the new areas that are coming around SAP, BDC with the Databricks, and this zero-based integration, that makes them rethink the way they need to integrate, the way they need to do things.

The third part, I think from a technology investment and technology, the use cases that most come for the technology that I would talk about is don’t just talk about now. This is the time that you have to, the way you own the people, the FTEs for your organizations. Agents are going to be your new FTEs.

That means that some of the new technology paradigm is going to be you will end up creating these co-intellects within your organization. That means you need to invest on what we call this agentic grid, where it becomes like a unified agentic fabric where every other agents can really collaborate and integrate and building on top of the same, the unified operational analytical core, the unified agentic integration on top of it, which is going to create a new set of experiences, agentic experiences rather than the traditional experiences or conversational experiences.

Then the new collaboration methods are going to be some of the critical aspects from a technology side that people have to really think from a technology standpoint. To start with, I would say you start looking at it from a data standpoint, building that unified core, building that unified integration and building that collaboration layer for both sharing and collaborating with intelligence as well as the agentic collaboration all governed under single umbrella. That is going to be the one critical use case which no one will feel bad about, and they are going to get really a 100X of their investments out of it.

Megan: Certainly no shortage of exciting developments on the horizon. Thank you both so much for that conversation. That was Bavesh Patel, senior vice president for Go-to-Market at Databricks and Rajan Padmanabhan, unit technology officer for data analytics and AI at Infosys, whom I spoke with from Brighton, England.

That’s it for this episode of Business Lab. I’m your host, Megan Tatum. I’m a contributing editor and host for Insights, the custom publishing division of MIT Technology Review. We were founded in 1899 at the Massachusetts Institute of Technology, and you can find us in print, on the web and at events each year around the world. For more information about us and the show, please check out our website at technologyreview.com.

This show is available wherever you get your podcasts, and if you enjoyed this episode, we hope you’ll take a moment to rate and review us. Business Lab is a production of MIT Technology Review, and this episode was produced by Giro Studios. Thanks for listening.

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

AI needs a strong data fabric to deliver business value

Artificial intelligence is moving quickly in the enterprise, from experimentation to everyday use. Organizations are deploying copilots, agents, and predictive systems across finance, supply chains, human resources, and customer operations. By the end of 2025, half of companies used AI in at least three business functions, according to a recent survey.

But as AI becomes embedded in core workflows, business leaders are discovering that the biggest obstacle is not model performance or computing power but the quality and the context of the data on which those systems rely. AI essentially introduces a new requirement: Systems must not only access data — they must understand the business context behind it. 

Without that context, AI can generate answers quickly but still make the wrong decision, says Irfan Khan, president and chief product officer of SAP Data & Analytics. 

“AI is incredibly good at producing results,” he says. “It moves fast, but without context it can’t exercise good judgment, and good judgment is what creates a return on investment for the business. Speed without judgment doesn’t help. It can actually hurt us.”

In the emerging era of autonomous systems and intelligent applications, that context layer is becoming essential. To provide context, companies need a well-designed data fabric that does more than just integrate data, Khan says. The right data fabric allows organizations to scale AI safely, coordinate decisions across systems and agents, and ensure that automation reflects real business priorities rather than making decisions in isolation. 

Recognizing this, many organizations are rethinking their data architecture. Instead of simply moving data into a single repository, they are looking for ways to connect information across applications, clouds, and operational systems while preserving the semantics that describe how the business works. That shift is driving growing interest in data fabric as a foundation for AI infrastructure.

Losing context is a critical AI problem

Traditional data strategies have largely focused on aggregation. Over the past two decades, organizations have invested heavily in extracting information from operational systems and loading it into centralized warehouses, lakes, and dashboards. This approach makes it easier to run reports, monitor performance, and generate insights across the business, but in the process, much of the meaning attached to that data — how it relates to policies, processes, and real-world decisions — is lost. 

Take two companies using AI to manage supply-chain disruptions. If one uses raw signals such as inventory levels, lead times, and supply scores, while the other adds context across business processes, policies, and metadata, both systems will rapidly analyze the data but likely come up with different conclusions. 

Information such as which customers are strategic accounts, what tradeoffs are acceptable during shortages, and the status of extended supply chains will allow one AI system to make strategic decisions, while the other will not have the proper context, Khan says. 

“Both systems move very quickly, but only one moves in the right direction,” he says. “This is the context premium and the advantage you gain when your data foundation preserves context across processes, policies and data by design.”

In the past, companies implicitly managed a lack of context because human experts provided the missing information, but with AI, there is a shortfall and that creates serious limitations. AI systems do not just display information; they act on it. If a system does not explain why data matters, an AI model may optimize for the wrong outcome. Inventory numbers, payment histories, or demand signals might be accurate, but they do not necessarily reveal which customers must be prioritized, which contractual obligations apply, or which products are strategically important. As a result, the system can produce answers that are technically correct but operationally flawed.

This realization is changing how companies think about AI readiness. Most acknowledge that they do not have the mature data processes and infrastructure in place to trust their data and their AI systems. Only one in five organizations consider their approach to data to be highly mature, and only 9% feel fully prepared to integrate and interoperate with their data systems.

Don’t consolidate, integrate

The emerging solution is a data fabric: An abstraction layer that spans infrastructure, architecture, and logical organization. For agentic AI, the fabric becomes the primary interface, allowing agents to interact with business knowledge rather than raw storage systems. Knowledge graphs play a central role, enabling agents to query enterprise data using natural language and business logic.

The value of the data fabric relies on three components: Intelligent compute to provide speed, a knowledge pool to provide business understanding and context, and agents to provide autonomous action are grounded in that understanding. What makes this powerful is how these capabilities work together, says Khan. 

The technology provides the architecture — a foundation that makes agent-to-agent communication and coordination possible. The process will define how businesses and IT share ownership, and establish governance and a culture in which people trust enough to adopt it. Now all three things must work together for a business data fabric to truly be successful.

“It empowers confident, consistent decisions, and when these elements all come together, AI just doesn’t analyze and interpret the data — it drives smarter, faster decisions that really create business impact,” he says. “This is the promise of a thoughtfully designed business data fabric, where every part reinforces the other, and every insight is grounded in trust and clarity.”

Technically, building a data-fabric layer requires several capabilities. Data must be accessible across multiple environments through federation rather than forced consolidation. A semantic or knowledge layer is needed to harmonize meaning across systems, often supported by knowledge graphs and catalog-driven metadata. Governance and policy enforcement must also operate across the fabric so that AI systems can access data securely and consistently.

Together, these elements create a foundation where AI interacts with business knowledge instead of raw storage systems — an essential step for moving from experimentation to real enterprise automation.

Beyond data isolation and dashboards

In the emerging era of agentic AI, the responsibility for monitoring, analyzing, and making decisions based on data increasingly shifts to software. AI agents can monitor events, trigger workflows, and make decisions in real time, often without direct human intervention. That speed creates new opportunities, but it also raises the stakes. When multiple agents operate across finance, supply chain, procurement, or customer operations, they must be guided by the same understanding of business priorities.

Without a common knowledge layer connecting disparate data together, coordination between systems quickly breaks down. One system might optimize for margin, another for liquidity, and another for compliance, each working from a different slice of data. 

Importantly, most enterprises already possess much of the knowledge needed to make this work, says Khan. Years of operational data, master data, workflows, and policy logic already exist across business applications — companies just need to make it accessible. Companies that deploy data fabrics gain greater trust in their data, with more than two thirds of enterprises seeing improved data accessibility, data visibility, and exerting more control over their data. 

“The opportunity isn’t just inventing context from scratch, it’s activating and connecting the context across your business that already exists,” he continues, adding that a data fabric is the “architecture that ensures data semantics, business processes and policies are connected as a unified system across all the clouds.”

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Making AI operational in constrained public sector environments

The AI boom has hit across industries, and public sector organizations are facing pressure to accelerate adoption. At the same time, government institutions face distinct constraints around security, governance, and operations that set them apart from their business counterparts. For this reason, purpose-built small language models (SLMs) offer a promising path to operationalize AI in these environments.  

A Capgemini study found that 79 percent of public sector executives globally are wary about AI’s data security, an understandable figure given the heightened sensitivity of government data and the legal obligations surrounding its use. As Han Xiao, vice president of AI at Elastic, says, “Government agencies must be very restricted about what kind of data they send to the network. This sets a lot of boundaries on how they think about and manage their data.”

The fundamental need for control over sensitive information is one of many factors complicating AI deployment, particularly when compared against the private sector’s standard operational assumptions.

Unique operational challenges

When private-sector entities expand AI, they typically assume certain conditions will be in place, including continuous connectivity to the cloud, reliance on centralized infrastructure, acceptance of incomplete model transparency, and limited restrictions on data movement. For many state institutions, however, accepting these conditions could be anything from dangerous to impossible. 

Government agencies must ensure that their data stays under their control, that information can be checked and verified, and that operational disruptions are kept to an absolute minimum. At the same time, they often have to run their systems in environments where internet connectivity is limited, unreliable, or unavailable. These complexities prevent many promising public sector AI pilots from moving beyond experimentation. “Many people undervalue the operating challenge of AI,” Xiao says. “The public sector needs AI to perform reliably on all kinds of data, and then to be able to grow without breaking. Continuity of operations is often underestimated.” An Elastic survey of public sector leaders found that 65 percent struggle to use data continuously in real time and at scale. 

Infrastructure constraints compound the problem. Government organizations may also struggle to obtain the graphics processing units (GPUs) used to train and access complex AI models. As Xiao points out, “Government doesn’t often purchase GPUs, unlike the private sector—they’re not used to managing GPU infrastructure. So accessing a GPU to run the model is a bottleneck for much of the public sector.” 

A smaller, more practical model

The many nonnegotiable requirements in the public sector make large language models (LLMs) untenable. But SLMs can be housed locally, offering greater security and control. SLMs are specialized AI models that typically use billions rather than hundreds of billions of parameters, making them far less computationally demanding than the largest LLMs.

The public sector does not need to build ever-larger models housed in offsite, centralized locations. An empirical study found that SLMs performed as well or better than LLMs. SLMs allow sensitive information to be used effectively and efficiently while avoiding the operational complexity of maintaining large models. Xiao puts it this way: “It is easy to use ChatGPT to do proofreading. It’s very difficult to run your own large language models just as smoothly in an environment with no network access.” 

SLMs are purpose-built for the needs of the department or agency that will use them. The data is stored securely outside the model, and is only accessed when queried. Carefully engineered prompts ensure that only the most relevant information is retrieved, providing more accurate responses. Using methods such as smart retrieval, vector search, and verifiable source grounding, AI systems can be built that cater to public sector needs. 

Thus, the next phase of AI adoption in the public sector may be to bring the AI tool to the data, rather than sending the data out into the cloud. Gartner predicts that by 2027, small, specialized AI models will be used three times more than LLMs.

Superior search capabilities

“When people in the public sector hear AI, they probably think about ChatGPT. But we can be much more ambitious,” says Xiao. “AI can revolutionize how the government searches and manages the large amounts of data they have.”

Looking beyond chatbots reveals one of AI’s most immediate opportunities: dramatically improved search. Like many organizations, the public sector has mountains of unstructured data—including technical reports, procurement documents, minutes, and invoices. Today’s AI, however, can deliver results sourced from mixed media, like readable PDFs, scans, images, spreadsheets, and recordings, and in multiple languages. All of this can be indexed by SLM-powered systems to provide tailored responses and to draft complex texts in any language, while ensuring outputs are legally compliant. “The public sector has a lot of data, and they don’t always know how to use this data. They don’t know what the possibilities are,” says Xiao.

Even more powerful, AI can help government employees interpret the data they access. “Today’s AI can provide you with a completely new view of how to harness that data,” says Xiao. A well-trained SLM can interpret legal norms, extract insights from public consultations, support data-driven executive decision-making, and improve public access to services and administrative information. This can contribute to dramatic improvements in how the public sector conducts its operations.

The small-language promise

Focusing on SLMs shifts the conversation from how comprehensive the model can be to how efficient it is. LLMs incur significant performance and computational costs and require specialized hardware that many public entities cannot afford. Despite requiring some capital expenses, SLMs are less resource-intensive than LLMs, so they tend to be cheaper and reduce environmental impact. 

Public sector agencies often face stringent audit requirements, and SLM algorithms can be documented and certified as transparent. Some countries, particularly in Europe, also have privacy regulations such as GDPR that SLMs can be designed to meet.

Tailored training data produces more targeted results, reducing errors, bias, and hallucinations that AI is prone to. As Xiao puts it, “Large language models generate text based on what they were trained on, so there is a cut-off date when they were trained. If you ask about anything after that, it will hallucinate. We can solve this by forcing the model to work from verified sources.”

Risks are also minimized by keeping data on local servers, or even on a specific device. This isn’t about isolation but about strategic autonomy to enable trust, resilience, and relevance.

By prioritizing task-specific models designed for environments that process data locally, and by continuously monitoring performance and impact, public sector organizations can build lasting AI capabilities that support real-world decisions. “Do not start with a chatbot; start with search,” Xiao advises. “Much of what we think of as AI intelligence is really about finding the right information.”

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Treating enterprise AI as an operating layer

There’s a fault line running through enterprise AI, and it’s not the one getting the most attention. The public conversation still tracks foundation models and benchmarks—GPT versus Gemini, reasoning scores, and marginal capability gains. But in practice, the more durable advantage is structural: who owns the operating layer where intelligence is applied, governed, and improved. One model treats AI as an on-demand utility; the other embeds it as an operating layer—the combination of operation software, data capture, feedback loops and governance that sits between models and real work—that compounds with use.

Model providers like OpenAI and Anthropic sell intelligence as a service: you have a problem, you call an API, you get an answer. That intelligence is general-purpose, largely stateless, and only loosely connected to the day-to-day operations where decisions are made. It’s highly capable and increasingly interchangeable. The distinction that matters is whether intelligence resets on every prompt or accumulates over time.

Incumbent organizations, by contrast, can treat AI as an operating layer: instrumentation across operations, feedback loops from human decisions, and governance that turns individual tasks into reusable policy. In that setup, every exception, correction, and approval becomes a chance to learn—and intelligence can improve as the platform absorbs more of the organization’s work. The organizations most likely to shape the enterprise AI era are those that can embed intelligence directly into operational platforms and instrument those platforms so work generates usable signals.

The prevailing narrative says nimble startups will out-innovate incumbents by building AI-native from scratch. If AI is primarily a model problem, that story holds. But in many enterprise domains, AI is a systems problem—integrations, permissions, evaluation, and change management—where advantage accrues to whomever already sits inside high-volume, high-stakes operations and converts that position into learning and automation.

The inversion: AI executes, humans adjudicate

Traditional services organizations are built on a simple architecture: humans use software to do expert work. Operators log into systems, navigate operations, make decisions, and process cases. Technology is the medium. Human judgment is the product.

An AI-native platform inverts this. It ingests a problem, applies accumulated domain knowledge, executes autonomously what it can with high confidence, and routes targeted sub-tasks to human experts when the situation demands judgment that the system can’t yet reliably provide.

But inverting human-AI interaction isn’t just a UI redesign—it requires raw material. It’s only possible when the platform is built on a foundation of domain expertise, behavioral data, and operational knowledge accumulated over years.

The three compounding assets incumbents already own

AI-native startups begin with a clean architectural slate and can move quickly. What they can’t easily manufacture is the raw material that makes domain AI defensible at scale:

• Proprietary operational data
• A large workforce of domain experts whose day-to-day decisions generate training signals
• Accumulated tacit knowledge about how complex work actually gets done

Services companies already have all three. But these ingredients aren’t moats on their own. They become an advantage only when a company can systematically convert messy operations into AI-ready signals and institutional knowledge—then feed the results back into operations so the system keeps improving.

Codifying expertise into reusable signals

In most services organizations, expertise is tacit and perishable. The best operators know things they cannot easily articulate: heuristics developed over the years, edge-case intuitions, and pattern recognition that operate below the level of conscious reasoning.

At Ensemble, the strategy for addressing this challenge is knowledge distillation. The systematic conversion of expert judgment and operational decisions into machine-readable training signals.

In health-care revenue cycle management, for example, systems can be seeded with explicit domain knowledge and then deepen their coverage through structured daily interaction with operators. In Ensemble’s implementation, the system identifies gaps, formulates targeted questions, and cross-checks answers across multiple experts to capture both consensus and edge-case nuance. It then synthesizes these inputs into a living knowledge base that reflects the situational reasoning behind expert-level performance.

Turning decisions into a learning flywheel

Once a system is constrained enough to be trusted, the next question is how it gets better without waiting for annual model upgrades. Every time a skilled operator makes a decision, they generate more than a completed task. They generate a potential labeled example—context paired with an expert action (and sometimes an outcome). At scale, across thousands of operators and millions of decisions, that stream can power supervised learning, evaluation, and targeted forms of reinforcement—teaching systems to behave more like experts in real conditions.

For example, if an organization processes 50,000 cases a week and captures just three high-quality decision points per case, that’s 150,000 labeled examples every week without creating a separate data-collection program.

A more advanced human-in-the-loop design places experts inside the decision process, so systems learn not just what the right answer was, but how ambiguity gets resolved. Practically, humans intervene at branch points—selecting from AI-generated options, correcting assumptions, and redirecting operations. Each intervention becomes a high-value training signal. When the platform detects an edge case or a deviation from the expected process, it can prompt for a brief, structured rationale, capturing decision factors without requiring lengthy free-form reasoning logs.

Building toward expertise amplification

The goal is to permanently embed the accumulated expertise of thousands of domain experts—their knowledge, decisions, and reasoning—into an AI platform that amplifies what every operator can accomplish. Done well, this produces a quality of execution that neither humans nor AI achieve independently: higher consistency, improved throughput, and measurable operational gains. Operators can focus on more consequential work, supported by an AI that has already completed the analytical groundwork across thousands of analogous prior cases.

The broader implication for enterprise leaders is straightforward. Advantages in AI won’t be determined by access to general-purpose models alone. It will come from an organization’s ability to capture, refine, and compound what it knows, its data, decisions, and operational judgment, while building the controls required for high-stakes environments. As AI shifts from experimentation to infrastructure, the most durable edge may belong to the companies that understand the work well enough to instrument it and can turn that understanding into systems that improve with use.

This content was produced by Ensemble. It was not written by MIT Technology Review’s editorial staff.

Building trust in the AI era with privacy-led UX

The practice of privacy-led user experience (UX) is a design philosophy that treats transparency around data collection and usage as an integral part of the customer relationship. An undertapped opportunity in digital marketing, privacy-led UX treats user consent not as a tick-box compliance exercise, but rather as the first overture in an ongoing customer relationship. For the companies that get it right, the payoff can bring something more intangible, valuable, and durable than simple consent rates: consumer trust.

The opportunities of privacy-led UX have only recently come into focus. Adelina Peltea, the chief marketing officer at Usercentrics, has seen enterprise sentiment shift: “Even just a few years ago, this space was viewed more as a trade-off between growth and compliance,” she says. “But as the market has matured, there’s been a greater focus on how to tie well-designed privacy experiences to business growth.”

And it turns out that well-designed, value-forward consent experiences routinely outperform initial estimates.
Touchpoints for privacy-led UX often include consent management platforms, terms and conditions, privacy policies, data subject access request (DSAR) tools, and, increasingly, AI data use disclosures.

This report examines how data transparency builds trust with customers; how this, in turn, can support business performance; and how organizations can maintain this trust even as AI systems add complexity to consent processes.

Key findings include the following:

• Privacy is evolving from a one-time consent transaction into an ongoing data relationship. Rather than asking users for broad permissions up front, leading organizations are introducing data-sharing decisions gradually, matching the depth of the ask to the stage of the customer relationship. Companies that take this tack tend to gather both a larger quantity and higher quality of consumer data, the value of which often compounds over time.
• Privacy-led UX is a prerequisite for AI growth. The consumer data that organizations gather is rapidly becoming a core foundation upon which AI-powered personalization is built. Organizations that establish clear, enforceable privacy and data transparency policies now are better positioned to deploy AI responsibly and at scale in the future. This starts with correctly configured consent mode across ad platforms.
• Agentic AI introduces new levels of both complexity and opportunity. As AI systems begin acting on users’ behalf, the traditional consent moment may never occur. Governing agent-generated data flows requires privacy infrastructure that goes well beyond the cookie banner.
• Realizing the advantages of privacy-led UX requires cross-functional collaboration and clear leadership. Privacy-led UX touches marketing, product, legal, and data teams—but someone must own the strategy and weave the threads together. Chief marketing officers
• (CMOs) are often best positioned for that role, given their visibility across brand, data, and customer experience.
• A practical framework can support businesses in getting it right. Organizations must define their data collection and usage strategies and ensure their UX incorporates data consent, including a focus on banner design. Following a blueprint for evaluating and improving privacy-led UX supports consistency at every consent touchpoint.

Download the report.

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Redefining the future of software engineering

Software engineering has experienced two seismic shifts this century. First was the rise of the open source movement, which gradually made code accessible to developers and engineers everywhere. Second, the adoption of development operations (DevOps) and agile methodologies took software from siloed to collaborative development and from batch to continuous delivery. Now, a third such shift looks to be taking shape with the adoption of agentic AI in software engineering.

Thus far, engineering teams have mainly used AI to assist with coding, testing, and other individual tasks, within tightly designed parameters. But with agentic capabilities, AI agents become reasoning, self-directing entities that can manage not just discrete tasks but entire software projects—and do so largely autonomously. If adopted and fully embraced by engineering teams, agentic AI will usher in end-to-end software process automation and, ultimately, agent-managed development and product lifecycle automation.

This report, which is based on a survey of 300 engineering and technology executives, finds that software engineering teams are seeing the potential in agentic AI and are beginning to put it to use, but so far in a mainly limited fashion. Their ambitions for it are high, but most realize it will take time and effort to reduce the barriers to its full diffusion in software operations. As with DevOps and agile, reaping the full benefits of agentic AI in engineering will require sometimes difficult organizational and process change to accompany technology adoption. But the gains to be won in speed, efficiency, and quality promise to make any such pain well worthwhile.

Key findings include the following:

Adoption momentum is building. While half of organizations deem agentic AI a top investment priority for software engineering today, it will be a leading investment for over four-fifths in two years. That spending is driving accelerated adoption. Agentic AI is in (mostly limited) use by 51% of software teams today, and 45% have plans to adopt it within the next 12 months.

Early gains will be incremental. It will take time for software teams’ investments in agentic AI to start bearing fruit. Over the next two years, most expect the improvements from agent use to be slight (14%) or at best moderate (52%). But around one-third (32%) have higher expectations, and 9% think the improvements will be game changing.

Agents will accelerate time-to-market. The chief gains from agentic AI use over that two-year time frame will come from greater speed. Nearly all respondents (98%) expect their teams’ delivery of software projects from pilot to production to accelerate, with the anticipated increase in speed averaging 37% across the group.

The goal for most is full agentic lifecycle management. Teams’ ambitions for scaling agentic AI are high. Most aim for AI agents to be managing the product development and software development lifecycles (PDLC and SDLC) end to end relatively quickly. At 41% of organizations, teams aim to achieve this for most or all products in 18 months. That figure will rise to 72% two years from now, if expectations are met.

Compute costs and integration pose key early challenges. For all survey respondents—but especially in early-adopter verticals such as media and entertainment and technology hardware—integrating agents with existing applications and the cost of computing resources are the main challenges they face with agentic AI in software engineering. The experts we interviewed, meanwhile, emphasize the bigger change management difficulties teams will face in changing workflows.

Download the report

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Enabling agent-first process redesign

Unlike static, rules-based systems, AI agents can learn, adapt, and optimize processes dynamically. As they interact with data, systems, people, and other agents in real time, AI agents can execute entire workflows autonomously.

But unlocking their potential requires redesigning processes around agents rather than bolting them onto fragmented legacy workflows using traditional optimization methods. Companies must become agent first.

In an agent-first enterprise, AI systems operate processes while humans set goals, define policy constraints, and handle exceptions.

“You need to shift the operating model to humans as governors and agents as operators,” says Scott Rodgers, global chief architect and U.S. CTO of the Deloitte Microsoft Technology Practice.

The agent-first imperative

With technology budgets for AI expected to increase more than 70% over the next two years, AI agents, powered by generative AI, are poised to fundamentally transform organizations and achieve results beyond traditional automation. These initiatives have the potential to produce significant performance gains, while shifting humans toward higher value work.

AI is advancing so quickly that static approaches to task automation will likely only produce incremental gains. Because legacy processes aren’t built for autonomous systems, AI agents require machine-readable process definitions, explicit policy constraints, and structured data flows, according to Rodgers.

Further complicating matters, many organizations don’t understand the full economic drivers of their business, such as cost to serve and per-transaction costs. As a result, they have trouble prioritizing agents that can create the most value and instead focus on flashy pilots. To achieve structural change, executives should think differently.

Companies must instead orchestrate outcomes faster than competitors. “The real risk isn’t that AI won’t work—it’s that competitors will redesign their operating models while you’re still piloting agents and copilots,” says Rodgers. “Nonlinear gains come when companies create agent-centric workflows with human governance and adaptive orchestration.”

Routine and repetitive tasks are increasingly handled automatically, freeing employees to focus on higher value, creative, and strategic work. This shift improves operational efficiency, fosters stronger collaboration, and generates faster decision-making—helping organizations modernize the workplace without sacrificing enterprise security.

Download the article.

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Shifting to AI model customization is an architectural imperative

In the early days of large language models (LLMs), we grew accustomed to massive 10x jumps in reasoning and coding capability with every new model iteration. Today, those jumps have flattened into incremental gains. The exception is domain-specialized intelligence, where true step-function improvements are still the norm.

When a model is fused with an organization’s proprietary data and internal logic, it encodes the company’s history into its future workflows. This alignment creates a compounding advantage: a competitive moat built on a model that understands the business intimately. This is more than fine-tuning; it is the institutionalization of expertise into an AI system. This is the power of customization.

Intelligence tuned to context

Every sector operates within its own specific lexicon. In automotive engineering, the “language” of the firm revolves around tolerance stacks, validation cycles, and revision control. In capital markets, reasoning is dictated by risk-weighted assets and liquidity buffers. In security operations, patterns are extracted from the noise of telemetry signals and identity anomalies.

Custom-adapted models internalize the nuances of the field. They recognize which variables dictate a “go/no-go” decision, and they think in the language of the industry.

Domain expertise in action

The transition from general-purpose to tailored AI centers on one goal: encoding an organization’s unique logic directly into a model’s weights.

Mistral AI partners with organizations to incorporate domain expertise into their training ecosystems. A few use cases illustrate customized implementations in practice:

Software engineering and assisting at scale: A network hardware company with proprietary languages and specialized codebases found that out-of-the-box models could not grasp their internal stack. By training a custom model on their own development patterns, they achieved a step function in fluency. Integrated into Mistral’s software development scaffolding, this customized model now supports the entire lifecycle—from maintaining legacy systems to autonomous code modernization via reinforcement learning. This turns once-opaque, niche code into a space where AI reliably assists at scale.

Automotive and the engineering copilot: A leading automotive company uses customization to revolutionize crash test simulations. Previously, specialists spent entire days manually comparing digital simulations with physical results to find divergences. By training a model on proprietary simulation data and internal analyses, they automated this visual inspection, flagging deformations in real time. Moving beyond detection, the model now acts as a copilot, proposing design adjustments to bring simulations closer to real-world behavior and radically accelerating the R&D loop.

Public sector and sovereign AI: In Southeast Asia, a government agency is building a sovereign AI layer to move beyond Western-centric models. By commissioning a foundation model tailored to regional languages, local idioms, and cultural contexts, they created a strategic infrastructure asset. This ensures sensitive data remains under local governance while powering inclusive citizen services and regulatory assistants. Here, customization is the key to deploying AI that is both technically effective and genuinely sovereign.

The blueprint for strategic customization

Moving from a general-purpose AI strategy to a domain-specific advantage requires a structural rethinking of the model’s role within the enterprise. Success is defined by three shifts in organizational logic.

1. Treat AI as infrastructure, not an experiment.  Historically, enterprises have treated model customization as an ad hoc experiment—a single fine-tuning run for a niche use case or a localized pilot. While these bespoke silos often yield promising results, they are rarely built to scale. They produce brittle pipelines, improvised governance, and limited portability. When the underlying base models evolve, the adaptation work must often be discarded and rebuilt from scratch.

In contrast, a durable strategy treats customization as foundational infrastructure. In this model, adaptation workflows are reproducible, version-controlled, and engineered for production. Success is measured against deterministic business outcomes. By decoupling the customization logic from the underlying model, firms ensure that their “digital nervous system” remains resilient, even as the frontier of base models shifts.

2. Retain control of your own data and models. As AI migrates from the periphery to core operations, the question of control becomes existential. Reliance on a single cloud provider or vendor for model alignment creates a dangerous asymmetry of power regarding data residency, pricing, and architectural updates.

Enterprises that retain control of their training pipelines and deployment environments preserve their strategic agency. By adapting models within controlled environments, organizations can enforce their own data residency requirements and dictate their own update cycles. This approach transforms AI from a service consumed into an asset governed, reducing structural dependency and allowing for cost and energy optimizations aligned with internal priorities rather than vendor roadmaps.

3. Design for continuous adaptation. The enterprise environment is never static: regulations shift, taxonomies evolve, and market conditions fluctuate. A common failure is treating a customized model as a finished artifact. In reality, a domain-aligned model is a living asset subject to model decay if left unmanaged.

Designing for continuous adaptation requires a disciplined approach to ModelOps. This includes automated drift detection, event-driven retraining, and incremental updates. By building the capacity for constant recalibration, the organization ensures that its AI does not just reflect its history, but it evolves in lockstep with its future. This is the stage where the competitive moat begins to compound: the model’s utility grows as it internalizes the organization’s ongoing response to change.

Control is the new leverage

We have entered an era where generic intelligence is a commodity, but contextual intelligence is a scarcity. While raw model power is now a baseline requirement, the true differentiator is alignment—AI calibrated to an organization’s unique data, mandates, and decision logic.

In the next decade, the most valuable AI won’t be the one that knows everything about the world; it will be the one that knows everything about you. The firms that own the model weights of that intelligence will own the market.

This content was produced by Mistral AI. It was not written by MIT Technology Review’s editorial staff.