AI agents for business finance: what they can and can't do
What AI agents can do in bookkeeping, close, collections and compliance today, where a person must stay in charge, and how to judge a product.
The short answer
An AI chief of staff routes work by classifying each incoming item, matching it to the agent or person whose mandate covers that type of work, and checking the item sits within that owner's ceiling. Anything no owner covers is refused and shown to a person rather than guessed at. Good routing is visible: you can see where each item went and why.
Routing sounds like the dullest part of the job. It is where most of the value sits. Every item that lands with the right owner first time is one the founder did not have to read, forward and chase.
Routing is a known pattern
Anthropic's guide to building effective agents describes routing as a workflow that classifies an input and directs it to a specialised follow-up task, and an orchestrator-workers pattern in which a central model breaks work down and delegates it. An AI chief of staff is a business version of both: it classifies, delegates and keeps the results together. The same guide recommends finding the simplest solution possible and adding complexity only when needed, which is a sound rule for routing rules too.
In practice that means routing should start as a short, written set of rules that you can read: invoices to finance, enquiries to sales, complaints to support, deadlines to compliance. The AI part is in reading messy inputs, such as a single email that contains a complaint, a question about an invoice and a request for a new quote, and splitting it into items each owner can act on.
The four questions behind each hand-off
- What is this? An invoice, a complaint, a quote request, a deadline, a payment request.
- Who owns this type of work? The agent or person whose mandate names it.
- Is it within their ceiling? If a refund is above what support may settle, it goes to a person instead.
- Is anything else connected? A complaint from a customer with an overdue invoice should reach both support and finance, with a note linking them.
| Incoming item | Usual owner | Goes to a person when |
|---|---|---|
| Supplier invoice | Finance | It is above the release ceiling or the bank details changed |
| New enquiry or quote request | Sales | The price asked for is outside the discount band |
| Customer complaint or refund | Customer Support | The refund is above its limit or it is a dispute |
| Filing deadline or screening match | Compliance | A match is found or the policy set is silent |
| Anything no mandate covers | Nobody, yet | Always: it is refused and shown, never guessed |
Refusing a hand-off is part of routing
The last row matters most. When an item does not fit any mandate, a weak system pushes it to whichever agent seems closest. A good one refuses the hand-off and puts it in front of a person. The OWASP GenAI Security Project warns that agents given functions, permissions or autonomy beyond what their task needs create excessive agency, and that in multi-agent systems a compromised or confused peer agent can trigger damaging actions. A chief of staff that only hands work to an owner whose mandate covers it is a practical control against that.
Payment-detail changes are the clearest case. Action Fraud, the UK's national fraud reporting centre, describes mandate fraud as criminals pretending to be an organisation you pay and asking you to change the bank details you send money to. A routing rule that sends every bank-detail change to a person, whatever the amount, costs little and stops a lot.
A worked example
Example only. A Lagos logistics firm receives an email at 8am: a long-standing client says a delivery arrived damaged, wants a ₦350,000 credit and mentions that last month's invoice is still unpaid because of the problem.
- The chief of staff classifies this as two linked items: a complaint with a refund request, and a payment dispute.
- The complaint goes to Customer Support with the delivery record attached. ₦350,000 is above the ₦100,000 support may settle, so the credit becomes a decision for the owner.
- The unpaid invoice goes to Finance, flagged as disputed so it is not chased as an ordinary late payment.
- The promise "we will respond on the credit by Friday" is added to the promises it holds.
Each hand-off is recorded with the original email as its source, so if the client later disputes what was agreed, the trail shows who received what and when. Without routing, the owner gets one email and has to do all of that sorting personally. With it, they get one decision with the context already joined up.
Routing will still go wrong sometimes. An email may be misclassified, or a request may sit in a channel the agent cannot read, such as a personal phone. The fix is not to give the router more power but to make its errors cheap: because it executes nothing, a misrouted item is a delay, not a loss, and correcting it should teach the rules something.
How to check routing is working
- Look at a week of routed items and count how many you had to re-route by hand.
- Check that refused items reached a person the same day.
- Ask whether linked items were linked, or arrived as separate unrelated tasks.
- Confirm every hand-off is recorded with the source, as described in audit trails for AI agents.
Where Pryme fits
The Pryme Chief of Staff reads finance, sales, customer and compliance as one list, decides which certified agent answers, and refuses a hand-off its certificate does not cover. Agents are built and certified on Pryme Intelligence, which makes a governed agentic operating system, and must re-earn certification when their mandate or data changes. Every routed item is recorded with its source.
Questions people ask
What if an item needs two agents?
It should go to both, linked, with one named as the lead. Splitting it silently is how disputes get chased as ordinary debts.
Can I override a routing decision?
You should be able to, and the override should be recorded so the rules can be improved.
Does routing need AI at all?
Simple rules handle a lot. AI helps with messy inputs such as emails that contain several requests at once.
What happens to items it cannot classify?
They should be shown to a person, not dropped or guessed. A growing pile of unclassified items is a signal to add a mandate.
Sources
- Anthropic: Building effective agents · accessed Oct 2026
- OWASP GenAI Security Project: LLM06 Excessive Agency · accessed Oct 2026
- Action Fraud (UK): Mandate fraud · accessed Oct 2026
This article is general information, not financial, legal or tax advice. Products described are provided under the terms on this site and are subject to regulation and eligibility.