One number everyone trusts

You have been in the meeting. Finance has a workbook. Operations has another. Both are “right” for a definition of right that nobody wrote down. The hour becomes a reconciliation, not a decision. Someone promises a new dashboard. Six weeks later the same argument is still in the room — now with nicer charts.

This is not a visualization problem. It is also not an AI problem, even though a lot of teams will try to treat it as one. If two extracts disagree, an agent grounded on both of them will just argue faster.

What “one number” actually means

It does not mean a single Power BI workspace, or a lakehouse logo on a slide. It means a named metric, at a named grain, with a named owner, computed from tables people already treat as production. If those four are not true, you do not have a number. You have a claim.

Write it in one sentence before you build anything:

Booked revenue is invoice amount at close date, excluding tax, owned by Finance, sourced from the certified billing table — not from the sales CRM export.

If you cannot say that out loud without a fight, the fight is the work. The report is not.

Where the argument really lives

Most “which spreadsheet is right?” fights collapse into four places:

• Grain. Booking date versus close date versus invoice date. Daily versus monthly. Customer versus contract. Two “revenue” columns that are not the same fact.
• Silent filters. Active customers only. Exclude internals. Drop the one region that “doesn’t count.” Those rules live in someone’s head or in a hidden Excel column.
• The last-mile transform. A join, a mapping table, a manual overlay that never made it back to the warehouse. The workbook became the system of record because it was faster than waiting for a pipeline.
• Two golds. A “certified” table that nobody retired when the next one shipped. Both have owners on paper. Neither is the one people actually use.

Dashboards inherit all of that. So do AI demos. If the model is reading a PDF dump or a one-off extract, it is reading the argument, not resolving it.

Why another demo makes it worse

A demo is designed to look decided. Clean prompt, clean answer, a room that nods. The parts that fail on Monday are the ones the demo hides: freshness, grain, access, and who is allowed to say the number is wrong.

You are still in demo land when any of these are true:

• The answer cannot cite a table and a grain.
• A prompt tweak is how you “fix” a wrong number.
• There is no frozen set of questions it must still pass next week.
• The people who own the metric have not signed off on the output.
• Nobody can run it besides the person who built the notebook.

Production AI, on this problem, is boring on purpose. It reads certified tables, fails closed when the eval drops, and has a named human gate. If the warehouse is not trusted yet, the honest move is to say so and fix the tables — not to wrap a chatbot around the mess.

A sequence that actually settles it

Pick one contested KPI. Not a domain. One number.

1. Write the grain in one sentence, including what is excluded. If Finance and Operations cannot agree, stop here. That disagreement is the deliverable.
2. Name an owner. Not a steering committee. One person who can say “this is the number” and be wrong in public.
3. Point at one source table — or admit you do not have one and need a pipeline before a report. Side extracts get marked unofficial or they get killed.
4. Publish the metric. Same grain, same filter, same owner, in the warehouse and in the semantic layer. Workbooks may read it. They may not redefine it.
5. Then put a report on it. Then, if there is a real job for it, put an agent on it. Eval the agent against the same goldens the humans already use.

That sequence is slower than a demo week. It is faster than a year of meetings that never quite close.

What we will not do

We will not reskin a dashboard on top of two disagreeing extracts and call it alignment. We will not stand up a chatbot on a PDF pile and call it production AI. We will not pretend a notebook is a pipeline.

AKAF takes the unglamorous path: stabilize the tables, name the grain, ship something your team can run. Data engineering is usually the first move. Analytics & BI is the published number. AI that ships only after the number is real.

If this is the constraint

Send the metric, the two spreadsheets, and the decision they are blocking. We will tell you whether the work is a definition, a pipeline, a report — or whether you should not build it yet.

Start a conversation →