Know what's authoritative before the AI reasons

A decade ago, I helped run the first enterprise big data group at Credit Suisse. Everyone thought the hard part was storing data. It wasn't.

The hardest part was getting a global firm to agree on what its data meant.

Today's AI has the same problem. It's just faster and more confident.

The trust gap

Ask an AI tool a question about your business and it will give you a fluent, well-organized answer. What it usually won't tell you is which source it treated as the truth, and whether a different source would have said something else.

That's the trust gap, and it's showing up in the numbers. In a Templafy survey (opens in a new tab) of 2,000 knowledge workers in the US and UK, published in October 2026, 42% said their AI tools lack sufficient access to approved company source content. Only 30% said approved templates or content are applied automatically.

Put another way: many people are using AI that can't reliably find, or doesn't know to prefer, the material their own company has approved.

Authoritative isn't the same as available

A model reads whatever it's given. It has no idea that the spreadsheet in the shared drive was superseded two years ago, or that the policy document on the intranet is the draft, not the final. Everything it can reach is, from its point of view, equally true.

People know better, or at least some people do. Someone in operations knows which client list is current. Someone in compliance knows which version of a policy was approved. Someone in finance knows that "revenue" means one thing in one report and something slightly different in another.

That knowledge has to be written down somewhere a system can use it. Someone has to say which record wins, and when, before the AI starts reasoning. Otherwise the model makes that call silently, and nobody can tell it did.

One version of the truth is a myth

The usual answer to this problem is to build a single source of truth: one golden version of every customer, product and number. It's a lovely idea. I've never seen it fully exist.

Credit Suisse, JPMorgan, BNY: every enterprise I've worked in had more than one version of the truth. That isn't failure. Different teams need different views of the same thing, for legitimate reasons. Risk sees a client one way. Sales sees them another. Both can be right.

The mistake is pretending otherwise. Trustworthy AI doesn't bury contradictions. It shows you where they are and where they came from, so a person can make the judgment call. I gave a talk on exactly this at the Knowledge Graph Conference in 2023, "In Defense of Inconsistency: On Managing Truth in a Knowledge Graph," with Michael Pool. The title still sums up my view.

Here's a hypothetical. Two systems disagree about a client's address. One was updated last week by a relationship manager. The other comes from an official registry and hasn't changed in a year. Which wins depends on what you're doing. For a holiday card, take the newer one. For a regulatory filing, you may need the registry, and a person should look at why they differ. A model can't make that call on its own, because the answer depends on purpose, not just on data.

Why knowledge graphs help

This is where knowledge graphs earn their place. Back in 2020 I wrote about the three features that make them useful: they link things directly, they use schemas that encode meaning formally, and they make integrating data from different places much easier.

Those features matter even more with AI in the picture. A knowledge graph can record that two systems hold different values for the same thing, where each came from, which one is authoritative for which purpose, and who decided. It keeps meaning and lineage in a place people can inspect, question and change.

That turns a hidden assumption into a visible decision. When an AI tool draws on the graph, its answer can point back to the source it used and the rule that made that source authoritative. When someone disagrees, there's something concrete to disagree with.

None of this has to be built all at once. Start with the handful of terms and sources that matter most to the decisions you're making with AI. Agree on what they mean, record where they conflict, and name an owner for each.

Transparent AI starts before the AI

Most conversations about transparent AI focus on the model: can it explain its answer? That matters. But an explanation is only as good as what sits underneath it. If nobody decided which sources count, the model's explanation will be a confident account of an arbitrary choice.

The work comes first. Decide what's authoritative. Write down what words mean. Show where sources disagree. Then let the AI reason, with people making the calls that need judgment.

AI doesn't simply have a hallucination problem. It has a truth-management problem. Fix that first.