Most monthly operating meetings contain a moment everyone recognizes. Finance presents a revenue figure, operations presents a different figure for the same month, and the next twenty minutes go to working out whose spreadsheet is wrong. The decision on the agenda gets deferred, somebody is asked to reconcile the two numbers before the next meeting, and the reconciliation duly happens.
Then the same moment arrives four weeks later, attached to a different metric.
Both numbers are usually correct
The instinct in the room is that one of the two must be broken, and that a careful review will find the error. In our experience both figures are calculated correctly and both are defensible, because the two teams are answering slightly different questions using the same word.
Finance counts revenue when it can be recognized. Operations counts it when the work was delivered. A customer is active in the CRM if somebody logged a conversation this quarter, and active in billing if an invoice cleared. On-time delivery is measured against the original promise date in one system and against the revised date in another.
Every one of those definitions is reasonable for the purpose it was built to serve. The difficulty is that none of them was ever written down, so the divergence stays invisible until two versions meet in a meeting.
Three places definitions come apart
Timing. The two reports use different cutoffs. One closes the month on the last calendar day, another waits for late invoices to post, and a third pulls whenever the scheduled job runs. A four-day difference in cutoff produces a real difference in the number, and neither report is wrong about the period it describes.
Inclusion. The teams disagree about what counts. Internal transfers, cancelled orders, refunds, intercompany work, and trial accounts each get included by one team and excluded by another, usually for a defensible reason established years ago by someone who has since left.
Source. The same metric gets pulled from different systems. The CRM, the billing platform, and the operational database each hold a version of the customer record, and they were never designed to agree. We worked with a nationwide services business running 35 disconnected systems where this was the daily condition, and consolidating them lifted data accuracy by 95% against a single source of truth while cutting manual reporting labor by 70%.
The cost is larger than the meeting
The twenty minutes are the visible part. Three larger costs sit underneath.
Trust stops being metric-specific. Once leadership watches two reports contradict each other, confidence drains from every report, including the ones that were always right. People start asking for the underlying data so they can check it themselves, which is how a management team ends up doing analysis instead of managing.
Reconciliation becomes a standing job. Somebody assembles the bridge between the two numbers every month, and that work is invisible in every budget because it lives inside roles that were hired to do something else. One regional bank we worked with was spending more than twenty hours a week on operational reporting before that cycle was compressed to ten minutes.
Decisions arrive late. A pricing change or a staffing decision waits for the numbers to agree. The delay is rarely recorded as a cost, and it is frequently the largest one.
An observation about the week just past
In early September, OpenAI reported that an internal model running roughly ten thousand concurrent agents spent about eighty-eight hours on the Navier-Stokes equations and produced a proof that smooth fluid velocities can blow up under certain conditions. The question traces back to Jean Leray’s work in 1934. Mathematicians have since raised pointed questions about how the result was reached, and that argument is some distance from settled.
Take the claim at face value for a moment anyway. We can apparently make progress on a problem that has resisted the field for ninety years, and we cannot get the revenue figure on two internal reports to agree.
The obvious thought follows quickly. Paste both spreadsheets into GPT-6 Astra, ask it which number is right, and reclaim the twenty minutes.
PLEASE DO NOT DO THAT
There are two reasons, and the second one matters more.
The first is that the model would answer. It would produce a confident, well-formatted adjudication of which figure is correct, while having no way to know which definition your business intends, because that decision exists only as an unspoken habit inside two departments. You would be asking a stranger to settle a dispute about a word that only your company defines.
The second reason is the data. OpenAI has stated that it cannot rule out that de-identified data derived from customers’ use of its products helped improve its models. The mathematicians who pressed that point had been working through Codex, a paid product, and they did not get a direct answer about training data. If that is where the ambiguity sits for people operating at the frontier of their field, a finance team pasting a customer-level revenue extract into a consumer chat window is in a considerably worse position.
That extract is your customer list, your pricing, and your margin structure. We have written about where this leads in shadow AI is a map, and at greater length in our paper on the hidden costs of shadow AI, including the community bank that crossed a disclosure threshold in two days with no attacker involved.
The repair is a written definition with a named owner
What settles the argument is smaller and less satisfying than a frontier model. For each metric that matters, write one paragraph stating what counts, what is excluded, which system it comes from, and when the period closes. Give that paragraph a named owner who is allowed to decide, and put it where anyone reading the number can find it.
Most mid-market businesses have somewhere between eight and fifteen metrics that drive decisions. Defining them properly is a week of work, and it holds for years.
Why it stays broken
If the repair is that cheap, the interesting question is why businesses live with the problem for a decade. The answer is structural. The disagreement spans finance and operations, each team is behaving correctly inside its own function, and settling it requires the authority to tell one of them that their reasonable definition is now the wrong one. That decision sits above both, with an owner who has twenty things ahead of it that feel more urgent.
Where this leaves you
Two reports disagreeing is a definitional problem wearing a technical disguise. Both numbers are usually right, the divergence traces to timing, inclusion, or source, and the cost compounds through eroded trust, standing reconciliation work, and decisions that wait. The repair is a written definition with an owner, and the shortcut that looks fastest this week is the one most likely to put your customer data somewhere you cannot retrieve it from.
Expect the temptation to grow. Every model released from here will make pasting the spreadsheet look more reasonable, and the businesses that come through this well will be the ones that settled their definitions first.
That work needs someone above the functions rather than inside one, with the authority to adjudicate and the experience to know which definitions carry real decisions. If your last three operating meetings each lost twenty minutes to the same argument, start with a conversation. We would be glad to help you settle it.