AI agents for finance and accounting: what to hand over, and what to keep

AI agents for finance do well at preparing work and badly at approving it. They capture and match invoices, check expense claims against policy, chase payments, propose reconciliation matches and draft month-end commentary. People keep the approvals, the payments and the judgment. Set up that way, an agent takes a large share of the keying and chasing off a finance team without weakening a single control.

This guide is for finance leads, controllers and the owners of accounting practices. It covers the jobs an agent can take on, the controls that keep it safe, when a finance platform is enough and when a custom agent is worth it.

#What a finance agent can take on

#Accounts payable

• Invoice capture. The agent reads invoices arriving by email or upload, extracts the supplier, dates, amounts, tax and line items, and suggests the account coding based on how that supplier was coded before.
• Matching. It matches the invoice to the purchase order and the goods receipt, and flags differences in quantity or price.
• Exceptions. Duplicates, missing purchase orders, totals that do not add up and unfamiliar suppliers go to a person with the reason attached.
• Routing for approval. The clean ones go to the right approver, with the source document one click away.

#Expenses

• Policy checks. Each claim is checked against your policy: limits, missing receipts, weekend spending, duplicates, split transactions.
• Questions back to the employee, so the finance team is not the one sending them.

#Accounts receivable

• Payment reminders. Polite, correctly timed reminders that know the invoice, the history and the customer's usual behavior. The tone gets firmer on your schedule, not the agent's initiative.
• Cash application. Matching incoming payments and remittance advice to open invoices, and proposing the allocation.
• Dispute triage. Reading a customer's reply, working out whether it is a query, a dispute or a promise to pay, and routing it.

#Close and reporting

• Reconciliation preparation. The agent proposes matches, and explains what is left unmatched and why.
• The month-end checklist. Tracking who owes what, chasing it, and showing the controller what is blocking the close.
• Variance commentary. A first draft of the explanations for this month against budget and prior period, written from the ledger detail, for a person to correct and finish.
• Questions about the numbers. "Why did gross margin fall in March?" The agent runs the queries and answers with the figures and their sources.

#Supporting work

• Audit requests. Collecting the supporting documents for a sample.
• Contracts and leases. Extracting the dates, amounts and terms finance needs.

#What people keep

• Approvals and payments. Always.
• Judgment. Accruals and estimates, revenue recognition, provisions, accounting treatment.
• Tax positions, and anything sent to a tax authority, regulator or auditor.
• Changes to supplier bank details. These are a common route for fraud. An agent should flag a change request and stop, never process it.
• Conversations with customers in real difficulty.

#The controls that make it safe

Finance already has the right instincts. Apply them to the agent as you would to a new team member.

• The agent prepares, a person approves. This is segregation of duties, and it is the rule everything else hangs on.
• No conflicting permissions. The agent cannot both create a supplier and release a payment. It cannot approve its own entries.
• Read access by default. Write access only where the job needs it, and only to draft or pending states.
• The systems do the arithmetic. Totals, tax and allocations are calculated by your ledger or by code. A language model writes words and makes suggestions. It is not a calculator.
• Every figure links to its source. The invoice, the bank line, the ledger entry.
• Confidence thresholds. Confident extractions go to the approval queue. Uncertain ones go to a person. A sample of the confident ones is checked every month.
• A complete audit trail. What the agent read, what it proposed, who approved it and when. Your auditor will ask.
• Data handling. Financial data goes only to AI providers under business terms that rule out training on your data, and in the regions you require. If you are subject to SOX or a similar regime, involve whoever owns your controls before you start.

#A finance platform or a custom agent?

Finance is well served by products, and for standard jobs they are usually the right first step.

Typical reasons finance teams end up with something custom: freight, construction or insurance paperwork that standard tools read poorly, several entities with unusual rules, or a reporting pack that nobody's product produces.

#A sensible first project

Pick one, prove it, then move to the next.

1. Invoices from your ten largest suppliers. They are frequent and consistent, so results come quickly. Measure how many reach the approval queue without anyone keying anything.
2. The payment reminder cycle. Low risk, easy to check, and it shortens the time you wait to be paid.
3. Variance commentary. The agent drafts, the controller edits. It saves hours at the busiest point of the month and touches no transactions at all.

#What to measure

#What it costs

• Accounting system AI features are usually included in your plan or a higher tier.
• Dedicated finance platforms are typically priced per user, per invoice or by payment volume.
• A custom finance agent typically costs $8,000 to $40,000 to build, depending on how many systems and document types are involved. A single automation, such as invoice intake for one entity, sits below that. Running costs are usually small next to the hours saved. See how much does an AI agent cost.

For reference, our own automation projects start from $4,800 and take about two weeks, and custom AI agents start from $9,800 and take three to five weeks. Each is quoted as one fixed price before work begins, with a 30-day fix window.

#Where to start

Count last month's invoices, expense claims and reminder emails, and estimate the minutes each one took. The largest total is your first project. If your accounting system or a finance platform can do it, start there. If your documents or systems are not standard, book a free 20-minute call and we will tell you what a build would involve. We have also written a step-by-step guide to automating data entry, which covers document capture in more detail.

This guide is part of our series on AI agents for business.

Common questions.

• What can AI agents do in finance?

  They can read invoices and receipts and enter them with the right coding, match invoices to purchase orders, check expense claims against policy, send and follow up payment reminders, propose bank reconciliation matches, track the month-end checklist, draft variance commentary, and answer questions about the numbers. In each case the agent prepares the work and a person reviews and approves it.

• Is it safe to let an AI agent work on the books?

  It is, if the normal finance controls apply to the agent too. It prepares and a person approves. It cannot both create a supplier and pay one. It has read access by default and write access only where needed. Calculations are done by your systems, not by the language model. And everything it does is logged, with each figure linked to its source document.

• Can an AI agent replace a bookkeeper or accountant?

  It can remove much of the keying, matching and chasing, which is often a large part of a bookkeeper's week. It cannot take responsibility for the accounts, make judgment calls on estimates and accounting treatment, or deal with your auditor and tax authority. In practice one finance person supported by an agent can handle more entities or more volume.

• How accurate is AI invoice processing?

  On clear, typed invoices from regular suppliers, modern extraction is very accurate, and it is weaker on handwriting, poor scans and unusual layouts. No tool is perfect, which is why a good setup uses confidence thresholds: confident results go straight to the approval queue, uncertain ones go to a person, and a sample of the confident ones is checked every month.

• What is the best AI agent for finance?

  If the job is a standard one, such as payables, expenses or collections, a dedicated finance platform or the AI features in your accounting system are usually the best start, because the controls and integrations are built in. A custom agent suits documents and rules particular to your industry, older ERP systems, or commentary on your own management accounts.