Finance and the dice-rolling AI

Ask an AI the same question twice and you can get two different answers. Banks, insurers and consulting firms use Claude for money work anyway. The trick is not fixing the dice: it is never letting the dice touch anything that must be exact.

The whole idea in one picture

Two seats. The wrong seat: the AI as the calculator and the ledger, crossed out, because a wobbly answer becomes a wrong number in the record. The right seat: the AI as the analyst whose work is checked, ticked, because a wobbly answer is only a draft.
1

First, the worry — and it is real

An AI builds its answers a bit like rolling weighted dice, one word at a time. Ask the same question on Monday and again on Tuesday and the sentences come out different, and once in a while so does a number. Nobody at Anthropic or at the banks claims to have fixed this at the source. Everything below is what they do instead.

One question, asked Monday and Tuesday, passes a pair of dice and produces two answers with the same gist but different words.
Under the hood

Large language models sample each token from a probability distribution, so outputs vary run to run (and even “temperature 0” is not a full cure in practice). The honest engineering position across every deployment the source thread examined is the same: architect around non-determinism rather than deny it — push everything that must be exact into deterministic components, and confine the model to work where variability was always tolerable.

2

Money work was never one-answer work

Give the same company file to two analysts and you get two different memos: different words, different emphasis, slightly different comparisons. Firms have always been fine with that, because only a narrow band must be exact: the sums, the source numbers, and the ledger. The AI is allowed into the wobbly part. It is kept out of the exact part.

Two analysts produce two different memos from the same file, which was always fine. Beside them, the part that must be exact: the sums, the source numbers, the ledger, kept out of the AI's hands.
Under the hood

The question “how do they solve non-determinism?” carries a slightly wrong premise: finance work was never fully deterministic. Judgment, synthesis and drafting always varied between humans; the firm tolerated it because the exact subset — arithmetic, source data, the system of record — was handled by deterministic computation. LLMs slot into the variable layer; the exact layer is handled the way it always was.

3

The AI never does the sums

When Claude works in a spreadsheet it does not invent totals from its dice. It writes the formula; the spreadsheet does the arithmetic. The dice can change how a sentence is phrased. They cannot change what =SUM adds up to.

The AI writes the formula =SUM(B2:B14), the spreadsheet computes 4,712.00, and a crossed-out die shows the dice never touch the arithmetic.
Under the hood

Claude in Excel reads formulas, edits them and processes data — the spreadsheet engine computes the results. It is the same shape as any tool-calling agent: the model writes the formula or the code, a deterministic runtime executes it, and anything numeric routes through code execution or a database rather than token prediction. Sampling temperature cannot change what =SUM(B2:B14) returns.

4

Facts arrive through pipes, not memory

The AI is not asked to remember what a company is worth or what a rating is. Governed pipes bring in the real, current numbers from the official sources, so the model works with verified facts instead of recollections. If the facts are exact going in, the dice have nothing exact to spoil.

Market data, credit ratings and a company registry flow through governed pipes into the AI, while its own memory is crossed out: never asked to recall a fact.
Under the hood

This is the connector ecosystem: governed, real-time access to data providers such as FactSet, S&P Global, Moody’s and LSEG. Dun & Bradstreet’s CDO makes the point explicitly — verified business identity data is what lets AI systems deliver the “deterministic, auditable outcomes” financial workflows require. The determinism lives in the data source, not in the model.

5

Every answer carries its receipts

Here is the quiet reframing that makes the whole thing work. You do not need the same answer every time. You need an answer you can check, every time. Every figure in the output points back to where it came from, so a reviewer can follow the trail before anyone acts. Two byte-identical memos that both invented the same number would be worse, not better: perfectly repeatable, and uncheckable.

A memo whose numbers carry receipt tags pointing to ledger page 3 and filing page 12. Identical memos without receipts get a cross; a memo with receipts gets a tick, because a reviewer can follow every number home.
Under the hood

Every output is source-attributed so teams can verify the work before acting on it, and every number can be traced back to its source. That swaps the requirement from reproducible to auditable — the same standard firms already apply to human work, where no two analysts were ever byte-identical either.

6

A person holds the stamp

The AI drafts; a person decides. The pitchbook, the credit memo, the filing — each is reviewed and signed by someone whose job is to find the draft’s mistakes before it goes anywhere. And some doors the AI simply has no key to: it cannot place a trade, write into the ledger, or approve a loan on its own.

The AI's draft flows to a person who reviews it, then to a stamp marked signed: only now does anything happen. Below, three locked doors: place a trade, write the ledger, approve the loan.
Under the hood

Every vertical repeats the same split: Claude assembles the analysis while you control the valuation call, structures the writeup while you own the credit decision, drafts the filing while you retain sign-off. The agent templates add verification subagents (comparables selection, methodology checks), and managed deployments add per-tool permissions, credential vaults and a full audit log of every tool call. Notably absent from all the marketing: autonomous trades, bookings, or credit decisions. Even the FIS story is about compressing AML investigations; the decision still goes to a human.

7

And at your kitchen table?

On a laptop you get the same tools: code that does the sums, pipes to your real files and accounts, your own rules written down. What you do not get is the firm’s safety net — the reviewer of your tax working is you, the person who asked because you didn’t know. So borrow the discipline: make it do sums in code, feed it real documents instead of memories, ask it to show the actual rule, and let an accountant hold the stamp on anything that matters. The AI’s job is to make that conversation cheaper, not to replace it.

At the firm, the AI's draft sits above a reviewer, checkers with audit logs, and insurance. At your kitchen table, the same AI sits above empty dashed slots: the reviewer is you, no second pair of eyes, a wrong return is yours alone. Below, four borrowed habits: sums run in code, real files not memory, show me the rule, an accountant signs.
Under the hood

What transfers: code execution for anything numeric, MCP connectors to real data (accounting software, bank exports, actual invoices), and file-based skills that encode your own conventions the way the finance plugins encode a firm’s.

What doesn’t: human-in-the-loop only works when the human can evaluate the output; there are no methodology-check subagents, eval suites or compliance readers on one pass with one reader; pasted-in numbers and recalled tax rules are unverified, and jurisdiction-specific thresholds are exactly what a model states confidently and wrongly; and the stakes concentrate — no firm amortises your mistake.

The translation: insist calculations run in inspectable code, connect real sources, ask for the rule citation and check it, and route consequential positions through a qualified professional with Claude’s work as the prepared material.