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 agent can handle refunds that your written policy covers exactly, up to a ceiling you set, with the payment itself going through your normal approvals. Anything outside the window, above the ceiling, partial, disputed or a goodwill gesture should go to a named person. The agent should never move money without the approvals your business has set.
Refunds are where customer support meets the bank account. Customers want them quickly; owners want them correct. An AI agent can help with both, provided the line between “settled by policy” and “decided by a person” is drawn clearly before it starts.
Three kinds of refund
| Kind | Example | Who decides |
|---|---|---|
| Policy refund | Unused item returned on day 12 of a 30-day window | Policy decides; agent prepares or completes within its ceiling |
| Statutory remedy | Faulty item, consumer entitled by law | Usually clear, but a person should check unusual cases |
| Discretionary refund | Outside the window, “goodwill”, partial for poor service | A named person |
The first kind is ideal for an agent: the conditions are written, the facts are in your records, and the answer is the same whoever asks. The third kind is a business judgement about a relationship, and should never be made by software. The second depends on the law in your country; for example, the UK Consumer Rights Act 2015 sets out remedies for faulty goods, and in Nigeria the FCCPC enforces consumer rights. See refunds in Nigeria for that market.
The steps an agent can take
- Identify the customer and the order or booking.
- Find the clause that applies and check every condition: dates, item condition, payment method.
- Calculate the amount from the records, not from the customer’s message.
- If it qualifies and is under the ceiling, prepare the refund for approval or complete it if your settings allow.
- Tell the customer what happens next and when, citing the clause.
- Record the case, the clause, the amount and the approval on the audit trail.
Step 3 matters. Customers sometimes quote the wrong amount, include delivery fees that are not refundable, or ask in a different currency. The agent should work from the order, and say so in its reply.
Speed is where the agent adds most value. Many refund complaints are not about the decision but the wait: a customer who returned an item and heard nothing for a week. An agent that confirms receipt, states the clause and gives a date removes most follow-up messages, even when a person still approves the payment.
Ceilings and approvals
A ceiling is the amount above which the agent stops and a person decides. Set it per action and per currency, and keep it low at first.
- Below the ceiling: the agent prepares the refund; depending on your settings, it goes ahead or waits for a single approval.
- Above the ceiling: it goes to a named approver with the case attached.
- Above a higher threshold: a second named approver is required.
- Always: the payment needs the authorisation your account requires, such as a transaction PIN.
This separation of duties is a basic fraud control. Without it, anyone who can talk the agent into a refund can get money out. With it, a manipulated conversation produces at most a draft that a person declines. The same logic applies to staff: nobody should be able to both promise and pay a refund alone.
Worked example (illustrative)
An online clothing shop sets a refund ceiling of £150 for the agent. A customer returns a £60 jacket on day 10: the agent checks the 28-day clause, prepares the refund, the owner approves it in one tap, and the customer is told the money will arrive within the time the policy states. Another customer asks for £240 back on a coat worn once, saying it lost a button. The agent explains that this may be a faulty-item case, sends it to the owner with photos and the order, and tells the customer a person will decide within a working day. The amounts are illustrative.
Currency deserves a line in your policy too. If a customer paid in euros, the refund normally goes back in euros to the same method; converting it creates a different amount and a possible complaint. State this, and let the agent quote it.
What to keep with a person
- Any refund that is not exactly described in your policy.
- Partial refunds and compensation for inconvenience.
- Refunds to a different account or payment method from the one used to pay, which is a common fraud pattern.
- Refunds linked to a chargeback or payment dispute; see disputes and chargebacks.
- Repeated refund requests from the same customer.
Refunds also change your books. Each one should reduce revenue in the right period and be matched to the original payment, which is a job for whoever keeps the books. Our guide to AI bookkeeping explains the finance side.
Finally, look at refunds as information. A rising number of refunds for one product, courier or service usually means something upstream needs fixing. An agent that groups refund reasons each week gives you that view without extra work; see learning from contact reasons.
Where Pryme fits
The Pryme Customer Support agent settles refunds that fall within your written policy and books them with the Finance agent, telling the customer what happens next. It cannot move money on its own, grant exceptions or offer goodwill credits above its ceiling; those go to a named person. Payments follow the approvals you set, including a second named approver above thresholds and your transaction PIN, and every step is recorded.
Questions people ask
Can an AI chatbot issue refunds automatically?
It can prepare or complete refunds your policy covers, up to a ceiling you set, with payments going through your normal approvals. Exceptions and larger amounts should go to a person.
What refund ceiling should I set for an AI agent?
Start low, at a level where a mistake would be easy to absorb, and raise it only after reviewing a month or more of correct cases.
How do I stop customers tricking an AI into refunds?
Work from order records not messages, never refund to a new account, require approvals above a ceiling and review repeated requests.
Do refunds need to be recorded in the books?
Yes. Each refund should be matched to the original sale and recorded in the right period. Ask your accountant how to treat them.
Sources
- Consumer Rights Act 2015 (legislation.gov.uk) · accessed Oct 2026
- Accepting returns and giving refunds (GOV.UK) · accessed Oct 2026
- Federal Competition and Consumer Protection Commission (Nigeria) · accessed Oct 2026
- AI Risk Management Framework (NIST) · 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.