Claude for finance: the numbers it can't be trusted to ship

The promise

It is a genuinely appealing pitch. Point Claude at NetSuite or QuickBooks, Salesforce or HubSpot, ask for this month's numbers, and watch it draft revenue, margin, and variance commentary in minutes. It is the same pitch whether you are running the monthly close, sending the investor update, building budget versus actuals, or putting together a board deck. For a finance team under a monthly deadline, that is a tempting amount of time saved.

Then someone checks the numbers.

The disqualifier

The board report was built on figures the model invented, and when challenged, the model admitted it. That is not a one-time bug. It is what non-deterministic output means: the same question can return a different number, with no definitions, no exceptions, and no audit trail.

Most work can absorb a little variance. Finance cannot.

What "non-deterministic" looks like at close

Picture a controller asking the same question twice. Monday: "What was Q3 revenue?" Claude returns $4.21M. Tuesday, same prompt, same underlying data, it returns $4.27M. Neither run shows its work, and nothing flags that the two disagree.

It is not lying. A language model predicts an answer rather than computing one, so without a fixed query underneath, the path from question to number can shift between runs. This is a documented limitation of how language models generate output, not a setting you can switch off. A $60,000 swing in a number leadership acts on is the difference between a clean meeting and an awkward one, and you would only catch it by re-deriving the number yourself.

• No repeatability: ask twice, and the totals can differ. A number you report cannot move because you re-ran the prompt.
• No audit trail: you cannot click a number to see how it was calculated or where it came from.
• The re-audit tax: to trust anything, the team re-checks everything, which can cost two full days a week and erases the time the AI saved.

When you cannot take a number for granted, the speed was never real.

How to ship numbers you trust

Invert the order. Instead of asking the model for the number, have the AI plan and write the SQL, let the SQL produce the number deterministically, and have a connector refresh it on schedule. The model never touches the data path, so the figures are repeatable and every one clicks back to its source.

You still get the speed of AI for the thinking and the drafting. You just stop betting a leadership meeting on a number nobody can trace.

How the fix works

1. AI drafts the SQL. You describe the metric in plain language, the model writes the query, and you review it once.
2. The definition is locked. Revenue, margin, and churn live in that query, not re-guessed every time someone asks.
3. The connector refreshes it. The numbers pull from NetSuite, QuickBooks, Salesforce, or HubSpot on a schedule, so your reporting is current without a rebuild.
4. Every figure explains itself. Click a number to see the SQL and the source rows behind it. That is the audit trail a finance number needs.

What a report-ready AI number looks like

Before a number goes into your monthly reporting, it should be:

• Repeatable: the same question returns the same number, every time.
• Defined: the calculation lives in a query anyone can read, not inside a prompt.
• Auditable: you can trace it back to the source rows in one click.
• Refreshable: it updates on a schedule, so it is current at meeting time.

From a static report to a live web dashboard

A monthly report is a snapshot the moment you export it. A Coefficient AI dashboard is the same finance view, published to a shareable URL that stays current on its own. Leadership opens a link, sees this quarter's numbers, and never needs a NetSuite or QuickBooks seat to do it.

Because the figures sit on grounded SQL, every metric on the dashboard is clickable: open any number to see the query and the source rows behind it. Edits are clicks on a card, not another prompt, and the whole view refreshes on the schedule you set. It is the reporting finance actually wants: live, shared, and auditable, instead of an export that is already stale by the meeting.

Where Claude still helps in finance

None of this means closing Claude. It is genuinely useful for the parts of finance that are language, not arithmetic. Point it at a grounded dashboard and it will draft the commentary, summarize what moved and why, turn a variance into a plain-English narrative, and structure the deck so it reads cleanly.

The split is simple: let the AI write the words, and let grounded SQL produce the numbers those words describe. You keep the speed on everything that does not have to be exact, and you stop betting the meeting on the one thing that does.

See how teams run finance reporting on Coefficient, or read more on the full Claude Boomerang pattern and why Claude in your spreadsheet cannot keep reporting live.