Agents plus governed data
Analytics that answers the follow-up question.
An AI agent can do an analyst's work: read a question, find the data, run the query, check the result, and try again. What decides whether it is right is not the model. It is whether the data underneath has one definition, one copy, and one place access is enforced.
Either half alone gets you a demo.
Agents without governed data are fluent and wrong, at machine speed, with nobody checking. Governed data without agents is a well-run warehouse that still answers questions in a week.
Put them together and the long tail of questions, the ones that were never worth an analyst's time, becomes answerable. That is the whole proposition, and most of the work is on the data side.
Start here
What the term means, how it differs from the dashboard era, and why the earlier attempt at this stalled.
- What is agentic analytics An AI agent doing the work of an analyst against governed data, with the authority to run queries and the obligation to show where the answer came from.
- From dashboards to questions Dashboards answer questions somebody anticipated. Agents answer the ones nobody did, which changes what the data layer has to support.
- Text to SQL and why it stalls Generating SQL from a question was solved years ago. Generating the right SQL against a real schema was not, and the gap is definitional rather than technical.
Three layers, three failure modes
An agent is not a model
It is a model inside a loop, reaching outside itself through interfaces. Knowing which of the three is failing is most of debugging one.
- The agentic stack Three separable layers, model, harness, and open interfaces, each with its own failure mode. Most agent problems are misattributed across them.
- The model What the reasoning layer actually contributes, what it cannot know, and why treating model choice as the whole architecture leads people astray.
- The harness The program that runs the loop. It decides context, tools, retries, memory, and when to stop, which is where most agent behaviour actually comes from.
- Open interfaces The four contracts an analytics agent speaks through, why each should be a specification rather than a product, and what a closed one costs later.
The other half
Why an open lakehouse
Agents multiply readers, multiply questions, and multiply the temptation to make a private copy. An architecture where the data sits in open formats, with one catalog deciding access, gets better under that pressure. One where reading requires a particular vendor's engine gets more expensive.
Five Apache projects settle five separate questions. Each one is a specification with more than one implementation, which is the property that makes the stack assemblable at all.
Running it for real
The part that decides whether a pilot becomes a capability: identity, evaluation, cost, and the order to do the work in.
- Governance for agents Identity, scoped credentials, and audit, arranged so that an agent cannot reach what its user cannot reach, and so you can prove it afterwards.
- One copy, many readers The architectural property that decides whether agentic analytics scales past a pilot, and the specific way copy sprawl produces disagreeing answers.
- Evaluation and trust How to know whether an analytics agent is right, why accuracy on a benchmark is not the question, and what to measure instead.
- Cost and latency Agentic analytics changes the shape of the workload from few large queries to many small ones, which moves where the money goes and what needs optimising.
- Portability and lock-in What it would actually cost to replace each component of an agentic analytics stack, and why the semantic layer is now the most likely trap.
- Rolling this out enterprise wide What separates an agent pilot that impresses a steering committee from a capability a whole organisation uses, and the order to do the work in.
Longer reads
Published pieces that carry parts of this argument in more depth.
- Definition What Is Agentic Analytics? The definition, and what separates an agent answering a question from a chatbot guessing at one.
- Semantics The Semantic Layer: The Definitive Guide Why shared definitions became load bearing the moment machines started asking the questions.
- Governance Apache Polaris: The Catalog Standard for Lakehouses and AI One catalog that both engines and agents resolve against, with access decided in one place.