AI in Finance: What to Automate, What to Gate, and What to Leave Alone

AI in finance is doing real work in three places and creating liability in a fourth. Reconciliation, document extraction, and first-draft reporting are shipping in production teams right now. Forecasting and anything that posts to the general ledger still need a human gate. Finance is already the second-most AI-heavy sector in the US economy. Here is exactly where the line sits, and what the arithmetic looks like.

TL;DR

• The US Census Bureau's Business Trends and Outlook Survey put AI use in finance and insurance at 33.9 percent as of May 3, 2026, against 19.8 percent for US businesses overall. Only the information sector was higher, at 39.7 percent.
• A Federal Reserve FEDS Note published April 3, 2026 found about 63 percent of financial-sector workers using generative AI at work while only about 30 percent of financial firms had adopted it. Workers are ahead of their employers.
• Deloitte's 2Q 2026 CFO Signals survey of 200 CFOs at billion-dollar-plus organizations: 51 percent use AI for operational productivity, 44 percent for planning and budgeting, 41 percent for analysis. Only 43 percent felt confident in their AI governance.
• FinTradeBench, an arXiv benchmark from March 2026 covering 14 models and 1,400 questions, found retrieval helps textual fundamentals substantially and trading-signal reasoning barely. Numbers over time are still the weak spot.
• A month of reconciliation exceptions costs single-digit dollars at Anthropic's published Haiku 4.5 rates. Review time is the only cost worth modeling.

Finance is further along than most sectors and further behind than it looks

Two numbers explain the whole situation.

The Census Bureau, which surveys the actual US business population rather than a conference mailing list, found finance and insurance at 33.9 percent AI use as of May 3, 2026. That is roughly 14 points above the national rate of 19.8 percent. Finance is not a laggard sector.

Then the Federal Reserve's April 2026 FEDS Note, authored by Jeffrey S. Allen, broke it down by person instead of by firm. About 63 percent of financial-sector workers reported using generative AI at work, against roughly 30 percent of financial firms reporting adoption. That is a 30-point gap between what people are doing and what their companies have sanctioned.

That gap is the actual story of AI in finance in 2026. Your analysts are already pasting variance tables into a chat window. The question is not whether AI enters your finance function. It is whether it enters through a controlled path with an audit trail or through a browser tab nobody logged.

The three-gate rule

Every finance task sorts into one of three buckets, and the sorting rule is reversibility, not difficulty. A model finding a duplicate vendor payment is harder than posting a journal entry. Posting is the one that needs the gate.

Two practical notes on using this. First, a task can move up a band once you have volume data proving the error rate, but it moves one band at a time and only after you can show the numbers. Second, nothing moves down into "ship unattended" just because the drafts have looked good for a quarter. Approval design is its own build, and I wrote the mechanics of it in human in the loop AI agent.

Reconciliation is the correct first build

Reconciliation is the best entry point in finance for four reasons: it runs every month, the output is checkable in seconds, the error is caught before anything posts, and the current process already has a human reviewing exceptions.

The build is not "AI does the reconciliation." Deterministic rules should still do the bulk matching, because exact matches are a database join and you should never pay a language model to do a join. The model works on the residual.

Run the arithmetic on a mid-sized account. Four thousand lines a month, deterministic rules clearing 88 percent, leaves 480 exceptions. Each exception costs roughly 2,500 input tokens (the candidate matches, the memo text, the policy snippet) and 300 output tokens (the proposal and its reason). That is 1.2 million input tokens and 144,000 output tokens.

At Anthropic's published rates for Claude Haiku 4.5, $1 per million input and $5 per million output, the month costs $1.92. On Sonnet 5 at $2 and $10, it costs $3.84. The Batch API takes 50 percent off both for work that can run overnight, which reconciliation can.

So the model spend is a rounding error against a single hour of controller time. What actually costs you is the build: connecting the ledger, writing the exception schema, and setting up the approval queue. Call it a few engineering days. The payback is not the token savings. It is the three to four hours a month a controller stops spending on line-by-line matching, plus a close that starts a day earlier.

One rule that keeps this honest: every proposal must carry the specific evidence it used. Not "this looks like the Acme invoice," but the invoice number, the amount delta, and the date offset. Ungrounded proposals are how a reconciliation agent quietly teaches your team to rubber-stamp. The techniques for forcing that grounding are in how to reduce AI hallucinations.

Close prep, split into drafting and posting

The month-end close is the highest-value target in finance and the one most likely to be oversold. Split it in two and it becomes tractable.

The drafting half is fair game. An agent can assemble the reconciliation package, pull supporting documents into one place, match the routine items, propose accruals with the calculation shown line by line, and write a first-pass flux commentary that cites the accounts it moved on. Deloitte's 2Q 2026 CFO Signals survey found 51 percent of those 200 CFOs already using AI for operational productivity tasks and 44 percent for financial planning and budgeting, which is mostly this drafting work.

The posting half is not. Every entry that hits the general ledger goes through a named human, and the approval is logged with the proposal it approved. That is not caution for its own sake. It is what makes the process auditable later, and it is the difference between an agent that speeds up your close and one that creates a control finding.

The practical shape: the agent produces a close checklist with each item marked ready, blocked, or needs judgment. A controller works the second and third categories. The first category still gets sampled, because an agent that is right 97 percent of the time and never checked will eventually be wrong on something material.

Forecasting is where it breaks

This is the section most AI-in-finance articles get wrong, so here is the evidence.

FinTradeBench, posted to arXiv in March 2026 and revised in June, built 1,400 questions grounded in NASDAQ-100 companies across a ten-year historical window and ran 14 language models through them under zero-shot and retrieval-augmented settings. The headline finding: retrieval substantially improves reasoning over textual fundamentals, and provides limited benefit for trading-signal reasoning.

Read that carefully, because it maps exactly onto the finance function. Give a model a 10-K and ask what changed in the revenue recognition policy and it does well. Give it a series of numbers over time and ask what happens next and the retrieval that helped on text stops helping.

Forecasting is the second job. So build it that way:

• Keep the projection arithmetic in a spreadsheet, a SQL model, or a statistical model you control and can re-run identically.
• Use the language model to gather and normalize the driver inputs, which is a text and extraction job.
• Use it to check the narrative against the numbers, which is a comparison job. "The commentary says headcount cost is flat, the model shows it up 11 percent" is a catch worth having.
• Use it to draft the commentary once the numbers are locked.

What you do not do is ask it for the number. If you want the reasoning behind picking a model tier for each of those jobs, how to choose an AI model covers the tradeoffs.

Governance is a build task, not a policy document

Deloitte's 2Q 2026 survey put a number on the discomfort: 43 percent of the 200 CFOs felt confident in their organization's AI governance, while 53.5 percent felt only somewhat confident. Their top internal concern was cost uncertainty or lack of transparency at 46 percent, ahead of lacking confidence in using AI for key operations at 35 percent. Governance ownership was scattered, with 33 percent naming the chief information security officer and 19 percent the CFO.

Scattered ownership plus 63 percent of financial workers already using these tools is the risk. The fix is not a policy PDF. It is four things you build:

1. A logged path. Every agent action writes what it read, what it proposed, and who approved it. If you cannot reconstruct a decision six months later, it is not an auditable process. The instrumentation pattern is in AI agent observability.
2. A sanctioned tool. People use the browser tab because nothing better exists. Give them the internal version and the shadow usage drops on its own.
3. A confidence threshold with a real fallback. Below the line, it routes to a human. Not a retry, not a lower-quality guess.
4. A sampled review. Pick 5 percent of approved items every month and re-check them cold. This is the only thing that catches drift.

None of that is exotic. All of it is boring infrastructure that turns a demo into something your auditors will accept.

The bottom line

AI in finance in 2026 is not a question of capability. Finance is already the second-most AI-heavy sector in the economy at 33.9 percent, and 63 percent of the people in it are using these tools whether or not their firm has a program. The question is where you draw the gate.

Draw it by reversibility. Extraction, classification, and flagging can run unattended because a wrong answer costs one click. Reconciliation matches, accrual proposals, and commentary get drafted by an agent and signed by a person. Ledger postings, payments, filings, and published forecasts stay human, permanently, no matter how good the drafts get. Then instrument all three so you can prove which is which.

Start with reconciliation. It runs monthly, it is checkable in seconds, the exceptions are already a human queue, and the model bill is under $4 a month. If it works there, you will have both the pattern and the audit trail to justify the next one. For the broader function-by-function picture, AI in business has the adoption data and the first three builds outside of finance.

Take the reconciliation build first, and take the prompts with you. CLAUDE FINANCE AGENTS is the free pack of finance agent setups I actually run, including the exception-matching prompt and the approval queue structure. Then join the newsletter for the builds I ship each week, with the arithmetic included.