Trust and governance
Accuracy, lineage, access control, and a graceful "I do not know". The precondition that makes every other capability believable.
It needs a forecast we have not validated.
Here is what I can show instead.
A system that can say "I do not know" is the one you can trust the rest of the time.
For years, governance in BI meant a folder of certified reports and a hope nobody went off-piste. The implicit promise was “the official dashboard is correct, so trust it.” When a question fell outside it, you were on your own, and the answer you cobbled together carried no warranty at all.
Generative BI changes the stakes. A system that answers any question in plain language is far more useful and far more dangerous, because it can produce a confident, fluent, wrong number for a question nobody certified. The shift is from “the AI said so” to answers you can verify, attribute, and control. Trust stops being a label on a report and becomes a property of every individual answer.
Why now
Every generative BI effort hits the same wall, and it isn’t the model; it’s trust. A pilot dazzles in a demo, then someone in finance catches a number that’s subtly wrong, and adoption quietly dies. The technology works; the confidence doesn’t.
The fix isn’t a smarter model. It’s the discipline around it:
- Grounding answers in governed definitions
- Lineage so every number can be traced
- Access control so people see only what they should
- Continuous evaluation against known-good results
- Graceful refusal when there’s no trustworthy basis to answer
None of these are new ideas in data management. What’s new is that they’re now the difference between a system people rely on and one they abandon. A plausible-but-wrong answer is worse than no answer, because someone acts on it before anyone checks.
What it looks like
A sales director asks, “what’s our win rate on enterprise deals in the new vertical we just entered?”
The vertical was added last week and isn’t in the certified model yet. A naive system would find something close, join a few tables, and return a clean percentage that means nothing. A governed system does the opposite: it recognizes that “enterprise deals” and the new vertical aren’t yet defined in the semantic layer, so it declines to invent a figure. It says it has no certified definition for that segment yet, shows what it does know, and offers to route the request to the data team or start a governed contribution so the definition gets added properly.
When the same director asks something inside the model, every answer arrives with its lineage (the definition used, the query run, the sources touched) and respects row-level access, so a regional lead sees their region, not the whole book. Where the number came from, and whether they’re allowed to see it, is on the answer.
Where it’s heading
Centralized governance doesn’t scale to a system answering thousands of questions across dozens of domains; one team can’t certify fast enough without becoming the bottleneck that kills the speed generative BI promised.
The trajectory is federated governance. Each domain owns the trust of its own data: definitions, tests, access rules, its sense of what “good” looks like, while the organization keeps one shared standard for lineage and accountability, so a number from one domain is legible to another. Local ownership keeps it fast and accurate; the common standard keeps the company on one truth rather than a dozen private ones.
How we think about it
A system earns trust by knowing the limits of what it knows. A system that answers everything isn’t trustworthy; it’s reckless. The valuable behavior is the boundary: answering well inside what it can stand behind, and saying so plainly outside it. Grounding, lineage, access control, and evaluation are how you draw that boundary; graceful refusal is how you honor it. Get that right and trust stops being a slogan and becomes something the system demonstrates with every response, exactly what makes chat with your data and everything built on it believable. People won’t talk to a system they can’t trust, and they can’t trust one that won’t admit what it doesn’t know.
Trust and governance, in short.
Why is trust the foundation rather than just a feature?
Because every other capability depends on it. Conversational answers, AI-drafted reports, and automated insights are only useful if people believe them. The moment a system produces one confident but wrong number, trust collapses and the effort stalls. It's not a layer you add later; it's the precondition for everything else.
What does "graceful refusal" actually mean?
The system declines or routes a question it can't answer well, rather than improvising. Outside the certified model, the right behavior is "I don't have a trustworthy basis for that" plus a path to a human, not a plausible figure pulled from a guess. Knowing its limits is what makes the answers you do get believable.
Does central governance slow domains down?
It does when one team approves everything. The trajectory is federated: each domain owns its definitions, tests, and access rules while the organization keeps one shared standard for lineage and accountability. Local ownership keeps it fast; the common standard keeps it coherent.
Keep exploring
The semantic layer
The shared definition of metrics and entities, and the single biggest lever for AI accuracy. Answers come from governed definitions, not raw-schema guesswork.
A company-wide memory
A shared memory of business logic, requirements, modeling conventions, and expectations that every agent reads and writes. Institutional knowledge as living context.
Where could this take your BI?
If this is the direction you want to head, we should talk.