In Today’s Issue:

🧑‍💻 What “AI-written” code means at Starburst, and what it is measured against

📈 His plan for 2.5x the revenue on 1.5x the headcount, and a harder door for juniors

🏎️ Ferrari or bus: when to build your own agent and when to buy

🗄️ Why he thinks the query engine has to change before the dashboard does

🔐 Why he says the user’s own access rights are not enough

🏦 The workflow he says changes inside regulated banks

🔮 The assumption about agents he expects to be completely wrong

✨ And more AI goodness…

Dear Readers,

For decades, the enterprise data stack was built for a patient customer: an analyst who writes a query, waits for it to run, builds a dashboard and walks over to finance to ask which revenue figure is meant. In Borgman’s telling, AI agents are a different customer. They can run thousands of queries in parallel, and when nobody has pinned down which figure counts, an agent can silently settle on the wrong interpretation and keep going.

Justin Borgman has spent most of his career on the plumbing underneath that stack, first at Hadapt, which Teradata bought in 2014, and since 2017 as co-founder and CEO of Starburst, whose Trino-based engine lets large companies query data where it already sits. His argument in this exclusive is that enterprise agents will be won or lost in the layer underneath the model: shared business definitions, permissions enforced at query time, and an audit trail back to the data.

He is also running the experiment on his own company. By his account, about a third of Starburst’s code now starts as a draft from tools like Claude Code; the target is 80%, and the plan is to grow revenue 2.5x on 1.5x the headcount. We asked what counts as AI-written, what that plan means for people trying to get into the industry, what an agent should show before an executive acts on its answer, and whether the user’s own access rights are enough once an agent does the querying. Where he got concrete, and where he stayed general, is part of the story.

All the best,

Kim Isenberg

Starburst sells the layer between a company’s data and whoever is asking questions of it. Its platform is built on Trino, the open-source SQL engine that grew out of Facebook’s Presto project, and lets banks, insurers and retailers query data where it already sits instead of first copying it into one central warehouse. In April 2026 the company launched AIDA, its AI Data Assistant, which Starburst says answers plain-language questions by reasoning in several steps over live data samples and metadata rather than firing off a single generated query. It became generally available in May. We sent co-founder and CEO Justin Borgman eight questions; he answered in writing.

Q1. You’ve said that AI tools already generate roughly a third of Starburst’s code and that this could eventually reach 80%. What exactly counts as “AI-written” code, and how are you measuring its effects on development speed, code quality, security, and review time?

"AI-written," for us, means the first draft comes from a tool like Claude Code and a human engineer reviews, corrects, and ships it. The engineer is still the author of record. About a third of our code originates that way today, and our engineering leadership has a goal of reaching 80% originated code over time.

We track output, not input. Jellyfish lets us watch velocity, feature delivery, and time to market, and the clearest result so far is idea-to-production time down by up to 60%.

Even so, the human stays in the loop. AI code still needs someone to catch what looks right and isn't.

Q2. You’ve argued that AI’s near-term effect will be slower hiring and greater output per employee rather than immediate mass layoffs. How does that change the structure of a software company, and what does it mean for junior employees trying to enter the industry?

Our goal over the next couple of years is to grow revenue 2.5x while growing headcount only 1.5x, and that gap is productivity. That changes the shape of a company. Teams stay smaller for longer, each person operates with more leverage, and the work shifts from producing volume to reviewing and directing it.

For someone entering the industry, that's a harder door, not a closed one. The roles that were mostly about producing volume are the first ones AI absorbs. What holds value is judgment, knowing when what a machine produced is wrong. The junior job is shifting from writing the first draft to telling whether the first draft is any good.

Q3. You use a “Ferrari versus the bus” analogy when discussing whether companies should build their own AI agents or buy existing products. What are the clearest criteria for making that decision, and what mistakes are companies currently making?

It comes down to three questions: How specialized is the use case? Do you have the engineering capacity to build and maintain it? And do you have the scale to justify the cost? When the use case is specialized enough that no product will do it well, and you have the people and the volume, you build. Otherwise you buy.

Q4. You’ve described AI agents as the new consumers of enterprise data. If agents begin replacing some dashboards and traditional BI workflows, what has to change first in a company’s data architecture?

The engine was built for analysts who could wait. Agents can't. Agents can run thousands of queries in parallel, and when one of those queries stalls then the whole chain of reasoning is held up, making the query engine the bottleneck in the system. So the engine changes first: built for agents, not analysts.

Context is also key. A human can clarify what revenue means, GAAP versus ARR. An agent can't. It picks a definition and runs. That business context has to be available at query time, so every agent is working from the same governed definitions. Get those right and dashboards don't die, they get demoted, from static reports to conversational.

Q5. Enterprise agents can produce convincing answers that are nevertheless wrong. What evidence, provenance, or validation should an agent provide before an executive can safely act on its answer?

Most AI tools take your question, generate SQL, run it, and hand back whatever comes out. If it joined on the wrong key, or picked gross billings when you meant recognized revenue, you don't know. The answer looks confident and it's wrong.

Before an executive acts on an answer, the system should make three things clear: what governed definitions and business context it used, what data and analytical steps produced the result, and where the underlying evidence came from so the answer can be verified and audited. A better model helps, but model intelligence alone isn't enough. Enterprise trust comes from grounding the agent in governed meaning and making its outputs traceable back to the data.

Your teams are moving fast, burning tokens, and shipping more than ever.

But more output doesn’t mean more impact.

Jira Product Discovery brings your ideas, customer insights, and business context together so product teams can weigh the evidence, make the tradeoffs, and decide what’s actually worth building. Then connect those decisions directly to delivery in Jira, so everyone knows what you’re building and why.

Jira Product Discovery. For better product decisions in the AI era.

Get Started for Free

Q6. Starburst says AIDA operates with the same access rights as the user making the request. Is that sufficient when an agent can search, combine, and analyze data much faster and at a much larger scale than a human? What additional safeguards are necessary?

Providing the same access as the user is the right foundation, but it is not sufficient on its own. An agent should never be able to access data the person behind it isn't authorized to see. But agents change the scale of the interaction. They can execute and iterate across many queries far faster than a person can, and potentially combine information in ways that weren’t anticipated when those permissions were originally designed. That’s why governance has to be enforced consistently at query time, not added after the fact. The same access controls, masking policies, and governance rules need to apply regardless of how many queries the agent makes or which data sources it combines.

You also need auditability. Organizations should be able to see what data the agent accessed, what queries it executed, and how an answer was produced so they can investigate and verify the result.

Q7. Citizens Financial Group and GEICO have been mentioned as examples of organizations applying this technology. Without sharing confidential information, can you describe one concrete workflow that changed and any measurable result it produced?

I won't name specific customers here, out of respect for their confidentiality. But I can tell you what we see again and again in regulated finance. 12 of the top 15 U.S. banks run on Starburst, and those are the institutions with the strictest rules anywhere about moving data.

Here's the workflow that changes, and it's the oldest frustration in the enterprise. Someone needs an answer, and the old way meant filing a request with the data team and waiting for a new dashboard to get built. Now that same person just asks the question in plain language and gets a governed, reasoned answer right when they need it, working from the same definitions everyone else uses. That gives time back. What used to take weeks now takes minutes.

Q8. Looking two or three years ahead, which widely held assumption about enterprise AI agents do you think will turn out to be completely wrong?

The assumption that will turn out completely wrong is that the smartest model wins. Every vendor is racing to pick a model, fine-tune it, and plug in a co-pilot, betting that better agents come from better models. But in the enterprise the model is only part of the equation. The hard problem is whether the agent actually understands the data it is working with.

Two or three years from now, I think the companies making real progress with agents will be the ones that invested in giving them consistent business context, governed access to data, and a reliable way to interpret what the data means. Models will keep improving and changing. The data foundation underneath them is what will determine whether agents are actually useful in the enterprise.

This interview was conducted in writing in September 2026 and has been lightly edited.

Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.

Put It on Autopilot

Borgman’s most useful point is about definitions. When “revenue” could mean GAAP revenue or ARR, gross billings or recognized revenue, and nothing in the system says which one counts, an agent can silently pick the wrong interpretation and return an answer that looks just as confident as a correct one. In his account, that is the failure to plan for: a model that quietly answers a slightly different question than the one asked. His remedy is shared, governed definitions at query time and answers that can be traced back to their data and SQL. That is also the layer Starburst sells, which is worth keeping in mind, though the problem he describes does not depend on who sells the fix. "If it joined on the wrong key, or picked gross billings when you meant recognized revenue, you don't know. The answer looks confident and it's wrong."

What convinces. The operating numbers are specific, and he puts his name to them: by his account, a third of Starburst’s code starts as an AI draft today, the target is 80%, and the plan is 2.5x revenue on 1.5x headcount. They are Starburst’s own figures, not an independent measurement. His definition of AI-written code is precise: the engineer stays the author of record, and he counts code that originates with a tool rather than claiming the machine wrote it. The agent argument also matches what AIDA’s own interface shows, from the generated SQL a user can inspect to the model that sits in a dropdown.

What stays open. We asked how AI-drafted code affects quality, security and review time. His answer offered one figure, idea-to-production time down by up to 60%, alongside the velocity and delivery data Starburst watches in the engineering-analytics platform Jellyfish; for quality, security and review time he gave no figure. The build-or-buy answer named three criteria but not the mistakes companies are making. The customer question produced no named deployment and no figure beyond "weeks to minutes." And the claim that 12 of the top 15 U.S. banks run on Starburst is his own. The company’s February announcement uses a different yardstick, four of the top five banks in the Americas and seven of the top ten in EMEA, so the two figures neither contradict nor confirm each other.

What to watch. Whether a bank or insurer puts its name and a measured result to an AIDA deployment, and whether the push toward 80% AI-originated code comes with published data on defects and review load, the half of our first question that is still open.

Our thanks to Justin Borgman and the Starburst team for the exclusive.

Before Starburst, Borgman co-founded Hadapt in 2010 while studying for his MBA at the Yale School of Management, commercializing the HadoopDB research of Yale professor Daniel Abadi together with Kamil Bajda-Pawlikowski, also a co-founder of Starburst. Teradata acquired Hadapt in 2014, and Borgman ran its Hadoop product portfolio as vice president and general manager before leaving to start Starburst. The Boston Globe named him to its 2025 Tech Power Players 50 list.

Sources:
🔗 Starburst, "Starburst Announces AI Data Assistant (AIDA): The First AI Assistant That Reasons Across All Enterprise Data for Fast, Accurate Decisions" (14 April 2026): https://www.starburst.io/press-releases/starburst-announces-its-ai-data-assistant-to-bring-ai-to-the-business-user/
🔗 Starburst, "Starburst Unveils Enterprise Intelligence Platform at AI & Datanova" (28 May 2026): https://www.businesswire.com/news/home/20260528612095/en/Starburst-Unveils-Enterprise-Intelligence-Platform-at-AI-Datanova-Giving-Enterprises-a-Faster-Path-to-Trusted-AI
🔗 Starburst, "Introducing Guardrails: Keeping AIDA on Task": https://www.starburst.io/blog/introducing-guardrails-keeping-aida-on-task/
🔗 Starburst, "Starburst Crosses $100M ARR as Enterprises Move from BI to AI" (18 February 2026): https://www.starburst.io/press-releases/starburst-crosses-100m-arr-as-enterprises-move-from-bi-to-ai/
🔗 SiliconANGLE, "Starburst adds AI front end in a bid to make conventional business intelligence dashboards obsolete" (14 April 2026): https://siliconangle.com/2026/04/14/starburst-adds-ai-front-end-bid-make-conventional-bi-dashboards-obsolete/
🔗 Starburst, "About us," with Justin Borgman’s biography: https://www.starburst.io/about/