Blog/Why your AI agents should know what to trust
Nov 16, 2026/8 min read

Why your AI agents should know what to trust

Agents need controlled freedom

Most leaders I know are drowning in a luxury we built ourselves. I hold access to several thousand dashboards and reports at my company, just by being a member of the groups I belong to. Revenue, productivity, lines of code, adoption curves, but also prototypes and experiments — all of it just one click away. When I put a question to an AI analytics agent, my first instinct is to give it the same view I have, to swing open the doors and let a model roam.

That instinct, I’ve learned, is incomplete.

Agentic analytics is scaling faster than any governance function built to keep pace with it. Two anxieties travel with it everywhere. One is volume: agents generating insights faster than any team can read, triage or act on them. The other is trust: the fear that an agent will produce an answer that is confident, plausible and wrong. Look closely and both problems share a root cause. Organizations have worked hard on what their agents can reach and very little on curation — identifying which assets are certified and trustworthy.

The MIT Technology Review Insights report “Raising the Quality of Insights in Agentic Analytics”, produced in partnership with Snowflake, found that strengthening AI and data governance is the top priority for improving agent insight quality, cited by 60% of the senior leaders surveyed. That should come as no surprise; if you can’t verify an answer, you certainly aren’t going to act on it. And an answer is only verifiable when the data behind it has been vetted. Curation, then, is the key to turning generated insight into insight a leader will actually stake a decision on.

The problem is that most organizations haven’t decided which of their data is ready to support authoritative answers and which isn’t. They grant access to the full estate and hope the model will sort out what matters. The fix takes discipline, and it has two halves. An agent should be able to reach the information it’s authorized to use, but it should also know which assets have been curated and certified as ready to inform a decision.

Reach is necessary, and nowhere near sufficient

Start with access. For an agent to be useful, it has to reach the data behind the dashboard.

A chart shows one or two dimensions the author chose to foreground. An agent that reasons from the visualization will tell you, at length, what you can already see. Connect it to the data behind the dashboard and something different happens: It can test correlations that were never previsualized, and the answers are computed from real data rather than inferred from pixels.

Connect the agent to the right tools, and this compounds. A metric definition matters more than a column name. Power BI semantic models are today’s best-practice example: They capture business metrics, relationships and definitions that organizations have already invested in and validated. Agents should reuse those existing semantics, not recreate them. When I ask about sales last month, I want the agent to reach for the company’s actual definition of a sale, the one a human vetted, not to invent an aggregation over whatever column happens to be named sales. Give an agent access to models your data scientists have already built, and it stops faking a forecast with a straight line and uses the model that was built to be right. Each of these moves makes the answer more grounded.

“Governance gets treated as a constraint on ambition. I think it’s what makes the ambition safe enough to run at all.” Bogdan Crivat, Corporate Vice President, Azure Data Analytics, Microsoft

Trust is the half nobody talks about

The part of the conversation that gets skipped is the part that decides whether any of the above is trustworthy.

Inside most companies, data is refined in stages, from raw ingestion to finished output. Engineers and data scientists need every stage, because the intermediate layers are where the work happens. But those layers aren’t clean or finished, and they aren’t ready to inform a decision.

A business user who asks an agent a question doesn’t want to be pointed at the half-processed layer without knowing it. They want to know which version someone has looked at, checked and will stand behind. So the second discipline is about trust: Certify the artifacts that are ready for authoritative AI answers, and make that certification status visible to agents. In my organization, that means a certification step: A human accountable for the data marks a semantic object as trusted, and the agent can use that signal to distinguish authoritative answers from exploratory ones.

Governance gets treated as a constraint on ambition. I think it’s what makes the ambition safe enough to run at all.

Broad access, informed trust

Put the two halves together and you get an operating model I believe serious deployments should adopt.

An agent needs the technological capability to reach into any of the islands where a person’s information lives — the warehouses and lakehouses, the reports, the conversations and the source systems, subject to appropriate permissions. And it needs trustworthy signals inside each island that say: Of everything here, these artifacts have been curated and certified to inform a business answer. Broad access, with clear signals of trust.

The distinction is between what an agent is authorized to access and what it should treat as authoritative. Deciding what earns certification is a governance call, and it belongs with the people accountable for the numbers.

Here’s why it matters more than any model choice. When an agent draws on a semantic layer that humans have vetted — the tables that map to real business metrics, the hierarchies, the definition of what a sale is — its answers are more likely to line up with the numbers your own regulatory reports reflect. That matters more than it sounds. It is the single strongest defense against the thing every board worries about: an agent handing an executive a polished number that doesn’t match the number the company already stands behind to its shareholders. Human vetting is what reconciles the two.

There’s one more condition for any of this to hold. Most organizations run more than one analytics engine across more than one cloud, and the definitions humans have vetted rarely live in just one of them. Power BI semantic models have demonstrated the value of governed business meaning at enterprise scale. Now we’re embracing Apache Ossie (incubating), an open, vendor-neutral effort, to help make those same principles of trusted semantics available across tools and engines.

When curated meaning is portable, certification becomes a multiplier: Agents on participating engines can draw from consistent governed definitions.

Precision is the cure for overload

In the MIT Technology Review Insights survey, 52% of executives say their analytics agents generate too many insights to be adequately evaluated. This is where trusted grounding becomes essential — but it’s only part of the answer.

Left to their own devices, agents can flood you with output; they’re designed to satisfy the person asking, and more output is not necessarily more useful. But notice what happens as grounding improves. We start satisfying that need with accuracy instead of bulk. An anchored agent gives you more trustworthy insights; business context and prioritization then help it surface fewer, more actionable ones. Overload begins to ease not simply because the data is certified, but because agents can distinguish what is correct from what matters.

“Precision, in other words, requires both governance and relevance.” Bogdan Crivat, Corporate Vice President, Azure Data Analytics, Microsoft

Decide what’s ready before you decide what’s next

Go back to those several thousand dashboards I mentioned earlier. The temptation is still there every time I open an agent: Give it everything I can see and let it work. What I’ve come to believe is that the leaders who get the most from agentic analytics will be the ones who go one step further. They’ll make a deliberate call about which data is finished enough to inform a decision, put a human mark on it and make that status available to their agents.

That work is unglamorous. It means naming owners for definitions, agreeing on what a sale is, and accepting that some layers of the estate are exploratory rather than authoritative. But it pays twice (and, if you’ve done this with your Power BI semantic models, you’ve already collected that second payment). The answers your agents return can match the numbers you already stand behind, and with the right prioritization, the volume becomes something people can actually act on.

So before the next expansion of agent access, ask a simpler question: What have we certified as ready, and do our agents know the difference? In my experience, the honest answer is usually less than you expect, but that’s the right place to start.

Apache and Apache Ossie are either registered trademarks or trademarks of the Apache Software foundation in the United States and/or other countries.

Share this post

Subscribe to our blog newsletter

Get the best, coolest and latest delivered to your inbox each week

Where Data Does More