How much does an AI finance agent cost?

Expect four cost lines, not one price: the platform that runs and governs the agent, the model usage per task, the work to connect it to your ledger and feeds, and the time your people spend reviewing what it drafts. The last two usually decide the total, and the depth you grant each capability drives both.

What are you actually paying for?

Licence fees get the attention, but they are rarely the largest line over the life of a finance agent.

For the general model behind these lines, see how much it costs to run an AI agent. Finance differs from support in one way that matters: a support agent's cost tracks conversations, while a finance agent's cost tracks documents and the depth of what it does with them. Compare how much an AI customer support agent costs.

Why does depth change the cost?

Each step up in depth shifts cost from people to the agent, and from recurring to one-off.

At depth 1 the agent reads and answers questions. Cost is mostly platform and model usage, and the saving is analyst time spent finding numbers.

At depth 2 it drafts journals or payment proposals. Now you add review time: every draft needs a person. If drafts are good, review is fast. If they are poor, review costs more than doing the work.

At depth 3 it executes routine items within your limits. Review time falls sharply, but testing and certification rise, because you need evidence before granting it and again at each renewal. The finance agent depth matrix shows which capabilities usually reach which depth.

A worked example

Take a business with 3,000 supplier invoices a month and two payables clerks who spend most of their time on coding. For illustration only:

The agent codes supplier invoices to accounts and cost centres. In the first month it drafts every coding (depth 2), and the clerks review all of them. Review is quicker than keying, but not by much, because they are learning to trust it. After certification, matched invoices under 2,000 are coded at depth 3, and the clerks review only exceptions and a sample.

The budget question is then simple. Does the cost of the platform, model usage and recertification come in below the clerk time released, and is that time going somewhere useful, such as supplier queries and cut-off review? If the honest answer is that review time has not fallen, the agent is not yet worth its cost at that depth.

What costs do teams forget?

• Exceptions. An agent that holds too many items for review moves work rather than removing it. Track the hold rate.
• Recertification. Evidence expires. Budget for a replay at each renewal, and after any change to the chart of accounts.
• Segregation of duties. Adding an agent can force a redesign of who approves what. Our segregation of duties matrix for finance agents helps size that work.
• Audit time. Your external auditor will want to understand the agent. Budget for a walkthrough in the first year.

What should stay with a person regardless of cost?

Releasing payments, approving supplier bank changes, posting above materiality and signing off the accounts. Saving money by moving those to an agent is a false economy: one error costs more than years of review time.

Which controls protect the business case?

• Measure before and after. Record time per invoice and error rates before the agent starts, so savings are real, not estimated.
• Grant depth on evidence. Depth earned through testing produces lower review time than depth configured on hope.
• Watch model usage per task. A sharp rise usually means the agent is reading more than it needs to.

The topic guide AI agents for finance: books you can audit sets out the controls the budget has to pay for.