By Zoe Hagfeldt
Demand Generation Lead at Metric Insights
Standing up an assistant that answers questions over company reporting is no longer the hard part. The hard part is what the person on the other end of the answer can actually tell about it.
The Number You Cannot Check
Picture an ordinary Thursday. Someone is finishing a deck for a Monday review and needs last quarter's figure for one region. They ask the assistant. Back comes 4.2 million, a confident sentence of context, and possibly a link to a dashboard.
Now ask the only question that matters. What can that person actually tell about the number they are about to put in front of their leadership?
They can tell that the assistant reached something. That is the whole of it.
The number could be any of three things, and from where they sit the three are indistinguishable.
It could be the certified number, from the report the data team reviewed and approved for exactly this purpose. It could be a stale copy, from a real report built by a real analyst, superseded eight months ago and never retired because nobody could prove who still depended on it. Or it could be a plausible reconstruction, assembled from a model containing a similar measure under a similar name. The arithmetic is sound. The measure is not the one they meant.
The assistant's tone is identical in all three cases. Fluency is not evidence, and the person reading has no way to separate the three without going and asking somebody, which is the thing the assistant was supposed to save them from doing.
Something Quiet Has Been Removed
For years the trust mechanism in most large organizations was a person.
An analyst knew which of four revenue reports was the current one. She knew the third was built for a different business definition, and she caught it before it reached a board pack. That knowledge was never written down anywhere, which was tolerable while she sat between the question and the answer.
An assistant is faster than the analyst and does not know what the analyst knew. Removing the human from the loop also removes the check nobody realized was a check.
This is one of the most common reasons these programs stall after a good demo. According to S&P Global, the average organization scraps nearly half of its AI proof-of-concept projects before they reach production. The demo ran against one curated dataset where the provenance of every number was obvious. Production has thousands of uncurated ones where it is not.
A Source Link Is Not an Assurance
The usual answer to all of this is a citation. The assistant tells you which dashboard it used.
A citation names an artifact. It does not say whether that artifact was the approved one, who owns it, when it was last reviewed, or whether a newer version of the same report exists two platforms over. In a company running reporting across several BI tools at once, a link to a dashboard is a location, not an assurance. It answers where this lives, which was never the question. The question was whether to use it.
Four Questions, Asked at the Moment of Reading
Here is a practical test, and it is deliberately the kind a business user can apply without help. Before using a number, they should be able to answer four questions from the answer itself:
- Which report did this come from?
- Is that report an approved one, and who approved it?
- Who owns it, and when was it last reviewed?
- Was this read from live data, or from a copy?
If those four cannot be answered at the moment of reading, the answer is not auditable, whatever else it is. It may well be correct. The point is that nobody can tell, and an answer nobody can check gets used anyway, because the alternative is waiting two days.
What Certification Does to Accuracy
An assistant inherits the governance of whatever sits beneath it. That sounds like a slogan. It is mechanical. Two assistants running the same model behave differently depending on what they are allowed to answer from.
Grounded in whatever it can reach, the scope of an answer is everything accessible. The decision about whether a given report is fit to answer from gets made implicitly, at answer time, by relevance matching. No person made that decision, and no person can see that it was made.
Grounded in certified content, the scope is what has already been reviewed and approved. The decision was made in advance, by named people, and recorded on the report itself.
That is the real difference, and it is worth stating plainly. Certification moves the approval decision from after the answer to before the question. Accuracy stops being something you hope the model achieved and starts being something a team decided, in advance, in the ordinary course of governing their reporting.
What Certification Actually Has to Mean
None of that holds if certified is only a label. For the argument above to be true in practice, certification has to mean five specific things.
- It is a property of the report, not a row in a spreadsheet beside it. A certification list maintained separately from the content it describes is wrong within a quarter, and nobody notices until it matters.
- It has levels rather than a single yes. Bronze, silver and gold let a partially reviewed report be distinguishable from an unreviewed one and from a fully governed one. A binary flag collapses that distinction and gets applied generously.
- It names an owner. Every certified report has somebody answerable for it, reachable by name.
- It carries lineage. What the report was built from, so a question about the number can be followed somewhere rather than stopping at a dashboard.
- It is visible to the person reading the answer. Certification that only administrators can see leaves the business user exactly where they started.
Add usage data to those five and decommissioning becomes a decision rather than a gamble, because you can see who actually opens a report before switching it off. That is how a reporting layer stops growing purely out of distrust.
Where Metric Insights Fits
Metric Insights is a governed BI Portal. It connects the BI tools a company already runs into one place where somebody searches once and sees which report is the certified one, whichever tool it lives in, with its owner and what it was built from. It reads metadata from the source tools and sits above the BI and query layer, so nothing in the pipelines changes, no data moves, and no BI tool is replaced. Certification is a property of the report, administered centrally rather than once per platform.
Concierge, the Metric Insights AI layer, answers natural-language questions against that content. The business semantics it answers from are taken from the models behind reports that have already been certified, which is why there is no multi-quarter project to define and maintain semantics in the warehouse before anyone can ask a question. When Concierge queries a source BI tool it does so with a per-user token, so that tool's own permissions and row-level security decide what comes back.
What that produces is not a cleverer answer. It is an answer a business user can situate: this came from a report at this certification level, owned by this person, built from this model.
The Order Most Organizations Get Wrong
The order of work here is not the order most organizations attempt. Most start with the assistant and discover the content problem afterward, usually at the point where somebody senior asks where a number came from and the honest answer is that nobody is sure.
The governed reporting layer is worth having on its own merits, with or without AI on top of it. People do not use reporting they cannot find, cannot interpret, or do not trust, and that was true for the twenty-five years before anyone asked an assistant anything.
What AI changes is the cost of getting it wrong, because the answer now arrives with no human in front of it and sounds exactly as confident when it is wrong as when it is right.
If this is the problem in front of you, two things worth reading next, both ungated: our point of view on
why BI adoption has stayed flat for twenty-five years, and the
BI and AI Governance Masterclass, six on-demand sessions.
