Five kinds of finance work stay with a person however good the agent becomes: releasing money, changing who gets paid, making accounting judgements, signing anything the business is bound by, and deciding exceptions to policy. The agent may prepare all of them. A named person decides, because that person is the one auditors and the board will hold to account.
Why draw a permanent line at all?
Most limits on an agent should rise as it proves itself. A few should not, because the reason for them is not the agent's accuracy. It is accountability and separation of duties.
A director signs the accounts because the law expects a person to answer for them. A treasurer releases a payment because dual authorisation only works if two people look. An agent that never makes a mistake still cannot carry those responsibilities, so no amount of testing moves the task.
Which tasks, and why?
| Task | Why it stays with a person | What the agent may prepare |
|---|---|---|
| Releasing payments | The final control against fraud and error; a second person's eyes | The payment run with support |
| Changing supplier or employee bank details | The commonest route for payment fraud | The request and the call-back checklist |
| Accounting estimates and judgements | Provisions, impairments, revenue timing need judgement and are challenged by auditors | Data, options and the calculation |
| Signing the accounts, tax returns and attestations | Legal responsibility rests with named officers | The draft, the trail and the checklist |
| Exceptions to policy | Waiving a control is a decision about risk | The case for and against |
| Reopening closed periods | Changes figures others have relied on | The proposed correction |
| Writing off balances | Gives up a claim or hides an error | The ageing and history |
| Granting credit to customers | A commercial decision with consequences for the relationship | The credit review |
This list mirrors the one in which finance tasks an AI agent can do on its own. Read together, they mark both edges of the agent's role.
Isn't this the same line as in compliance?
It is close. Compliance draws the same line around decisions with legal effect, such as filing a suspicious activity report or approving a high-risk customer, for the same reason: accountability attaches to a named person. Our answer on which compliance decisions must stay with a person sets out that list. In finance the line runs mostly around money leaving the business and figures the business publishes.
A worked example
Take a manufacturer facing a disputed customer balance of 80,000 that is now a year old.
The agent pulls the ageing, the correspondence and the payment history, and notes that the customer has disputed three invoices for quality and paid the rest. It drafts two options: a specific provision for the disputed invoices, or a full write-off. It calculates the profit effect of each and attaches the trail.
The financial controller decides on a partial provision, records the reasoning, and asks the sales director to talk to the customer. The agent did most of the work. The decision, and the judgement the auditors will test, belong to the controller.
Which controls keep these tasks with people?
- Withhold the capability, not just the instruction. The agent's identity should not hold rights to release payments, edit the supplier master or reopen periods. If the right does not exist, the agent cannot use it by mistake.
- Separate the preparer and the approver. The person who approves should not be the one who set up the agent's draft. Our segregation of duties matrix for finance agents maps the pairs.
- Make the approval a real decision. The approver sees the evidence, not just a button. A held item should show what the agent relied on and what it was unsure of.
- Record the human decision. Who approved, when, and on what evidence, filed alongside the agent's draft.
The finance agent depth matrix shows these tasks capped at depth 2 permanently. The topic guide AI agents for finance: books you can audit explains how that fits the audit trail.
Where Pryme fits
In Pryme, a customer narrows what an agent may do with its own limits, so these tasks can be held for a named approver on every occasion. Some actions go further: the platform withholds them for every agent at every depth, such as reversing a settled payment on our own rail. Every held, refused and withheld verdict is filed.
Questions people ask
Could an agent release very small payments?
It could be technically allowed, but most finance teams decide against it. Small payments are where fraud tests a control first, and the time saved is slight.
What if the person just approves everything the agent drafts?
Then the control has failed even though it exists on paper. Measure how often approvers change or reject drafts, and train them to look for what the agent was unsure of.
Does this list change for very small businesses?
The tasks do not change, but the people might. A small business may have one director approving everything; the point is that a person, not the agent, makes the decision.