Episode 31 — Databricks' State of AI Agents and Stanford's AI Index: What They Reveal About Enterprise AI in 2026

Episode 31: Two Industry Reports

Sunil: Okay, welcome, all, to Token Drop Episode 31. We were struggling today to figure out what to talk about. In the pre-meeting we were kicking around ideas. It's been a long week, and I think we're a bit tapped out, but we ended up landing on a couple of really interesting things. One is a report from Databricks that you came across, Sam, and the other, Sajjad, is one you came across from Stanford. Both are about the AI industry. It would be interesting to go through these reports and pick out a few things that stood out to each of you. Sam, why don't you go first? What is this Databricks report, and what would you take away from it?

Databricks' State of AI Agents Report

Sam: The Databricks report is about the state of AI agents. They're saying we're in a new era of enterprise AI: agentic architectures that produce outcomes and are curated for enterprise use cases, which is exactly where DaaX is today. These architectures often pair foundation models with enterprise context and tools so they can independently plan and take action.

Sam: The report has some key findings. Telemetry on the Databricks platform shows organizations thoughtfully using their data with multiple models and tools to build agentic systems. Investments in governance and evaluation are helping produce high-quality, accurate outputs. The report says multi-agent systems are becoming the new enterprise operating model, and AI agents are driving core database activities. About 80% of databases are now built by AI agents, would you believe it? A few years ago that was very different. AI is also now part of critical workflows across industries, and evaluation and governance are the building blocks of production today.

From Chatbots to Multi-Agent Systems

Sam: What they're claiming is that there's a significant shift: enterprises are moving from single chatbots, which we were used to, to multi-agent systems built on domain intelligence. That's a favorite topic of Sajjad's: domain intelligence, domain fusion, knowledge cartridges. That's where things are moving, toward agentic systems built on domain intelligence. And the report says growth in these systems has been 327% in just four months. So one thing we should be very clear about is that multi-agent systems are taking over.

AI Agents Are Building the Databases

Sam: Second, AI agents are driving core database activities. As I said, about 80% of databases are being built by AI agents, and 97% of database testing and dev environments are built by AI agents. That shift is driving the need for a new kind of database, something called Lakebase. It's a new database specifically targeted at AI.

Model Flexibility: Most Companies Use Multiple LLMs

Sam: AI is also now part of critical workflows across industries, with generative AI use cases like automating routine tasks. So model flexibility is very important in the new AI strategy. About 78% of companies use not just one LLM but two or more, which we already knew. We weren't tied to any one LLM, because we expected a combination of LLMs to be used for different use cases, and that's becoming critical across workflows in different industries.

Evaluation and Governance Get More Projects into Production

Sam: Finally, AI evaluation is critical to ensuring high-quality outputs. Companies that use evaluation tools get nearly 6x more AI projects into production. Companies using AI governance put around 12x more AI projects into production. So AI governance is becoming a top investment priority for companies right now.

Information Extraction: Turning Unstructured Text into Tables

Sunil: I'm scrolling through the report as we talk, because I hadn't had a chance to really look at it before. A couple of things I wanted to ask you about. It says information extraction represents 31% of usage, so it looks like 31% of agents are used for information extraction. Then it talks about converting unlabeled text into structured tables: product details, prices and descriptions from complex supplier PDFs. So I'm assuming they mean these agents are used to extract unstructured data from PDFs—

Sam: Yes.

Sunil: —into a structured database.

Sam: Yes. Data engineering into a structured database.

Sunil: Why would they want to do that? Not everything can be put into a relational database.

Sajjad: It's like our favorite A-to-D conversion, analog to digital. Right now, data that's out there in unstructured form, whether in PDFs, Word documents or elsewhere, is hard to process by machine. But when you convert unstructured data into structured data, into tables with rows and columns, you unlock the potential for machine processing. That's a big lift, so I guess that's one of the places this applies.

Sunil: That applies to numbers. You can crunch numbers and do mathematical computations on, say, specs or pricing. But descriptions are purely unstructured, so I'm not sure how that works.

Sajjad: You might be able to extract statistics. Say there are descriptions of school students and their performance over the years. You can extract all of that. Converting unstructured content into structured content has been the traditional way of processing unstructured data, because NLP processing of unstructured data has largely been the realm of research labs and research applications. Everyday businesses struggle to process unstructured data. So for the last 5 to 10 years, a common strategy has been to convert it into structured data, so you can process it with traditional tools: analytical tools, relational databases, reporting tools, dashboarding tools and so on.

Sunil: Interesting.

From Screen Scraping to Normalized Databases

Sam: This is a very interesting use case, Sunil. In earlier years, if you were building a network management platform or a piece of infrastructure, you needed to talk to real devices to get data. We couldn't get every vendor to write APIs, so some companies did screen scraping. They scraped the screens, pulled out unstructured data and pushed it into their databases. I'm talking about the 2011 to 2013 timeframe, when people wrote drivers, not just device drivers but CLI drivers and other types, trying every possible way to get information and collectively push it into a normalized database. Why would you do that? Gathering the data in unstructured form is something you build on the fly. But once the data is in your database, you can reason against it.

SQL, Logs and Analytics

Sajjad: Yes, because SQL is such a common language. Literally every analyst knows how to use SQL, and that helps unlock insights. Even traditional server logs and instrument logs are unstructured, in the sense that each is a long string of text. Embedded in that text are useful bits of information, like latency, lag or errors. You want to aggregate those logs, extract that information and push it into a table in a structured database. Then you can see how often you've had latency problems, how many times the application errored out, and which application or microservice was the biggest contributor to errors or latency. Those are the kinds of insights you get by converting unstructured logs into structured data in a database.

Structured and Unstructured Data Must Interoperate

Sam: That's a huge thing. Even though we say unstructured data represents a huge portion of what's out there, most of the processing world still depends on databases. There's no escaping it. I'd say 70 to 80% of systems running today use databases for their day-to-day operations. If that's the case, we need a strategy where unstructured and structured data interoperate.

Sajjad: And that's been the pursuit: how do we make unstructured data processing tools readily available to analysts? That's why LLMs and chat services have enormous potential. You can ask any question in any language, of content in any other language, and still get the insights you're looking for. That's been huge.

Custom Benchmarks vs. Runtime Verification

Sunil: Another thing that stood out to me as I catch up on this report is evaluation. It says evaluation goes beyond general benchmarks to custom benchmarks specific to an enterprise's data or task. That's interesting. It means companies are using benchmarks internally to compare different solutions and see which gives them the highest-quality, most trustworthy outputs in their context.

Sajjad: Exactly. LLM evals, or model evals, are about this: given your knowledge of your applications and use cases, you design some tests and put a variety of LLMs through them to see which performs better for your kind of data and applications. But that's about it. You can qualify an LLM, but the eval isn't going to ride shotgun in that Jeep when you take the LLM for a test drive in your production applications. That's where you need real-time, in-process verification, which is what ClaimGuard is all about. ClaimGuard does the same kind of verification, but at runtime, in a live environment.

Sunil: Interesting.

Databases for the Agent Era: OLTP Limits and Lakebase

Sam: Exactly. And Sunil, you raised a very interesting point earlier about databases for the AI era. The report talks about it. OLTP databases are designed to efficiently handle high volumes of real-time transactional data, and traditional OLTP databases have powered most operations. But they were built around human workflows: predictable transactions and infrequent schema changes. You never expect somebody to change the schema on the fly. You go through ALTER commands, somebody inspects them, and provisioning is manual.

Sam: Think about what's happening at enterprises adopting agentic AI today. Those assumptions no longer hold. Agents generate continuous, high-frequency read and write patterns, and they create and tear down environments programmatically. They run complex, read-intensive queries as part of multi-step reasoning loops, and you see similar behavior in our product. That places demands on databases that traditional OLTP systems were never architected to support. Agentic workflows require infrastructure that supports massive concurrency at millisecond scale, with branching, isolation and real-time responses at scale, beyond human workloads.

Sunil: Is that the Lakebase they talk about in the report?

Sam: Lakebase, that's exactly it.

Sajjad: Yeah.

Sam: This has given rise to a new category of operational databases built for elasticity, programmability and scale. Lakebase exemplifies this class of operational database, enabling AI agents to create, branch and orchestrate Postgres transactional databases in milliseconds. That's a huge thing.

OLTP, OLAP, HTAP and Delta Lake

Sajjad: Traditionally, you had OLAP systems, which are analytical, and transactional systems. During the day, people transact, buying your products and services, on an OLTP database. Toward the end of the day, you extract summary information and push it into an OLAP database so analysts can process it. But there's a lag between your customers transacting on the OLTP database and your analysts processing that data. The question is how to reduce that lag. So people started talking about hybrid transactional and analytical processing, HTAP. With the Apache Iceberg format, and with Databricks and Delta Lake, which everyone is coming around to, you can build applications that are both transactional and analytical.

Stanford HAI's AI Index Report

Sunil: Maybe it's time to move to the second report. Sajjad, you had the Stanford report. Tell us a little about it.

Sajjad: Stanford HAI, the Institute for Human-Centered Artificial Intelligence, publishes an annual report called the AI Index Report. This one was published many months ago, so it's a little old, but I think it's still quite relevant if you just look at the summary of their findings.

AI Capability Isn't Plateauing

Sajjad: Key finding number one is that AI capability is not plateauing. In their view, it's accelerating. They were looking back at 2025, but it may still apply to 2026. Capability is accelerating and reaching even more people, and they have statistics and metrics to show it.

The US-China Model Gap Has Closed

Sajjad: An important insight is that the US-China AI model performance gap has effectively closed. DeepSeek was a light-bulb moment. Since DeepSeek, we've seen a slew of Chinese open-weight models. You can download the weights and run the models in your own data centers for sovereign AI, and they're almost as good as the frontier models.

Data Centers and Chip Concentration

Sajjad: The third is that the United States hosts the most AI data centers, with most of their chips fabricated by one Taiwanese foundry. We know who that is, and I think they're trying to mitigate that risk a little. I wouldn't call that an insight. It's a fact most of us knew.

The Jagged Frontier: Math Olympiad Gold, but Can't Tell Time

Sajjad: The fourth one is interesting: AI models can win a gold medal at the International Math Olympiad but cannot reliably tell time. They call this the jagged frontier, and that's effectively what we've been dealing with. On one hand, these models are tackling Nobel Prize-worthy problems in mathematics, and increasingly in theoretical computer science and biology. On the other hand, you give them a basic business workflow, and they hallucinate and fail. So what accounts for this jagged intelligence, and how do you mitigate the risks?

Neuro-Symbolic AI and the Lean Verifier

Sajjad: In our interpretation, this is where neuro-symbolic AI helps. Neuro-symbolic architectures are helping even in the Math Olympiad results and the Navier-Stokes problem we saw. I think those critically depended on the Lean verifier, which is a symbolic verifier. That's an interesting insight for us.

Responsible AI Isn't Keeping Pace

Sajjad: Another finding: responsible AI is not keeping pace with AI capability. Safety benchmarks are lagging and incidents are rising sharply. We've been saying this for a while, so there's nothing new in it for us. And these are only reported incidents; we don't know about the unreported ones.

US Investment Leads, but Talent Attraction Is Declining

Sajjad: The US leads in AI investment, but its ability to attract global talent is declining. Immigration policy and the H-1B visa situation definitely aren't helping attract global talent, and I think that will be a serious concern for this country. But the US still leads AI investment. We're seeing literally trillions of dollars invested, and now people are waiting for the returns.

Adoption at Historic Speed

Sajjad: AI adoption is spreading at historic speed, and consumers are getting substantial value from tools they often access for free. I think that's consequential. We've grown so dependent on tools like ChatGPT, Claude, Perplexity and Google Gemini, and it's having downstream effects. Yesterday, Google reported its numbers. Search volumes are holding steady, but revenue isn't growing at the same pace, partly because they haven't yet figured out how to manage ads effectively inside their AI search tools. That's a concern, but I think they'll solve it eventually. There's some turbulence for now, but it speaks to the speed and ferocity with which consumers have adopted these tools for all kinds of applications.

Learning AI at Every Stage of Life

Sajjad: Formal education is lagging behind AI, but people are learning AI skills at every stage of life. That's another interesting one. Talk to friends and family anywhere in the world, and you'll find they're using the same tools we use. In some sense, it's democratizing access to education and information. Anybody, anywhere in the world now has access to the same insights and information, so I think competition will get much more intense.

AI Sovereignty and Open-Weight Models

Sajjad: Ninth, AI sovereignty is becoming a defining feature of national policy, but capabilities remain uneven, even as open-source development helps redistribute who participates. Think of the European Union AI Act. Even individual US states have their own AI legislation. Most countries have had data locality requirements for quite some time, and I think they'll adapt those regulations for AI. Companies are feeling the need for sovereign AI more than ever, so open-source and open-weight models play a big role in helping them comply.

Public Trust and Runaway Agents

Sajjad: AI experts and the public have very different perspectives on the technology's future, and global trust in institutions to manage AI is fragmented. That's important too. Over the last few weeks, we've heard about runaway agents causing havoc, hacking into bank accounts and so on. I wouldn't say that's largely because of AI. It's because of poor security posture and a lot of day-one bugs in many of these tools and applications, which AI can help discover. For the malicious few who want to use AI to attack, the opportunity presents a much bigger surface today. I wouldn't blame that entirely on AI.

The Router Is the Control Point

Sam: The main thing is that you need a control point, because machines are becoming operators. The notion of a single brain doing everything is out, and a team of specialists is in. That leads to an interesting equation: the model is becoming a commodity, and the router is the control point. Fifty-nine percent of companies now run three or more models. So the durable asset isn't the model. It's the layer that decides which model handles which data under which rules. That control point, the router, becomes very important. AI is also moving into the loop: 96% of inference is now real time. We used to think it was batch.

Verifiable Trust: From Pilots to Production

Sam: Some of this is moving in the right direction for us, especially for DaaX, because trust is a scarce resource. In fact, the report talks about 95% of pilots never reaching production. Why? The organizations that ship are the ones that can prove their agents behave. The report ties governance and evaluation together, which means we need verifiable trust, which is at the center of our conversation.

Sunil: Right.

Sam: That's why I think we're heading in the right direction. We're talking about verifiable trust, which becomes central, because it's the only way to move from pilots to production. That's how I see these two reports: they show that the neuro-symbolic side of things, verifiability and the evals we run are quite important. It's not like you throw something out and say, "Oh, it works." There will come a point when the industry starts asking how many of these pilots made it to production, and that really requires verifiable trust.

Mistral Large 4, "Le Chonk"

Sunil: We only have a few minutes left, and I wanted to ask you about one more thing: Mistral. I've started hearing more about their Large 4, nicknamed "Le Chonk," which, as I understand it, means something like a big, chubby cat.

Sam: Yep.

Sunil: It's a 1-trillion-parameter open-weight model, I believe. I wanted to see if you'd heard about it, what you think of it, and whether you've had a chance to look into it. It's gotten some buzz recently. I'm putting you on the spot, since I haven't asked you about this before.

Sajjad: No, it's a good thing. It's interesting because it's an open-weight model, so you can download it from Hugging Face. It's a trillion-parameter model with 52 billion active parameters and a million-token context, and it supports text and image inputs. Properly used, you're now getting a frontier-class model in downloadable form, and I wonder what this will do to the frontier labs. I can imagine it performing at literally frontier levels, whether in cybersecurity, coding or any domain you can think of.

How It Compares to Luna, Sol and Astra

Sunil: Is it equivalent to something like Luna, or is it more like Astra?

Sajjad: Probably like Luna. We don't know how big Luna is, but one can imagine it's in that ballpark of performance, somewhere between Luna and Sol, or between Luna and Astra. We don't know. But I'd imagine it's in that ballpark, so you're now within striking distance of the best of the best models, with an open-weight model you can download and run on your own. Effectively, you can control your cost and comply with government sovereignty requirements in different countries. I don't know as much about Mistral's multilingual capabilities, but it sounds like a really promising model.

Sunil: It says it has three modes: low, medium and high reasoning.

Sajjad: Yeah.

Open-Weight Models and Growing Active Parameters

Sunil: Sorry, Sam, go ahead.

Sam: One thing: open-weight models are very important for the ecosystem. That's point number one. And they're becoming humongous. We talked about Kimi K2 before. But the number of active parameters in these mixture-of-experts models is exploding. I don't know which way to go. We used to prefer MoEs that keep the active count small even when the model is very large, but there's been a shift, with bigger models increasing their active counts. I'm wondering why that is. Maybe Sajjad has an explanation for why we need 52 billion active parameters when other MoEs fit into something smaller.

Sajjad: The other models we've been looking at were 100 billion parameters or less, so their active parameters look small by comparison.

Sam: Yep.

Sajjad: But here, 52 billion active parameters means you certainly need a bigger GPU cluster to host and run it, and that requires specialized skills to run, host and manage. But for those who can, I think it offers a very attractive alternative.

What the Reports Miss: Permission, Proof and Accountability

Sam: Certainly, but my question is at what cost, and what extra bells and whistles it provides. We've talked about huge context windows and things getting dropped in the middle.

Sam: All of this is good, but here's what all these reports don't catch, in my view. They're all about agents reading and writing, provisioning databases and so on. None of them talks about what actions an agent is allowed to take, how an action is approved, or how it's proven afterward. When agents become autonomous, the question becomes: what did the agent do? The data layer was the first thing agents took over. The control layer, meaning permission, proof and accountability at machine speed, is next, and nobody in these reports is talking about it.

Sam: Say an auditor comes in and wants to look at the actions agents have performed. They don't want to depend on what the organization says; they want an independent review. What tools can provide that kind of tracing? That brings in sovereignty and who can access what. Permissions, proof and accountability are huge, and these reports don't discuss them. Why does that matter? Because we should stop designing systems for people who use AI and start designing systems for agents that run operations. We're moving into that arena. In fact, that's the more interesting thing we should talk about in our next podcast.

Wrap-Up

Sunil: Next week?

Sam: Yes. Next week we can talk about data security and what it entails. But that's the gist of these reports.

Sunil: Or maybe the week after. I think we have a guest in two weeks, but we can talk about it next week, no problem. Okay, I think we've run out of time. Thank you, guys. We'll talk to you later.

Sajjad: Thank you.

Sunil: Thank you.