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 support agent can handle most of an online shop's routine questions, such as where an order is, how to return something and what the delivery options are, as long as it answers from your own order records and written policies. Refund exceptions, damaged-goods disputes, chargebacks and angry repeat customers should go to a person, with the case history attached.
Most online shops find that a handful of questions make up the bulk of their inbox: where is my order, can I return this, when will it be back in stock, can I change my address. These are repetitive, time-sensitive and answerable from data the shop already holds. That makes them the natural place to start with an AI customer support agent. The harder part is deciding where the agent stops.
The questions an agent handles well
A support agent is useful when the answer already exists in a record or a written policy and only needs to be found and explained. For a typical shop that covers:
- Order status. Reading the order and shipment record and giving the customer the latest tracking event, not a guess.
- Returns within policy. Explaining the return window, generating the return instructions and confirming when the parcel is received.
- Delivery options and costs. Quoting the published rates and cut-off times for each country you ship to.
- Product facts. Sizes, materials, compatibility and care instructions, taken from the product listing.
- Address or contact changes before dispatch, where your process allows it.
- Stock questions. Whether an item is listed as available and whether a restock date has been published.
What these have in common is that the agent can show its source. If a customer asks why their refund is 14 days away, the agent should be able to point to the clause in the returns policy that says so.
Where it should hand over to a person
Some cases look routine but carry money, legal exposure or a relationship at risk. These are the ones to keep with a named person:
| Situation | Why a person decides |
|---|---|
| Refund or credit outside the written policy | It is a goodwill decision with a cost, not a lookup |
| Item arrived damaged or not as described | It may involve evidence, courier claims and statutory rights |
| A card chargeback has been raised | Deadlines and evidence rules come from the card scheme and acquirer |
| Customer threatens legal action or a regulator complaint | Wording matters and may be used later |
| Third contact on the same unresolved problem | The process has already failed once |
| Suspected fraud or account takeover | It needs checks the agent should not perform alone |
The handover matters as much as the decision to hand over. A good agent passes the case intact: the order, the messages so far, the policy clauses it relied on and what the customer has been told. The person picking it up should not have to ask the customer to repeat anything.
The consumer rules sit behind every answer
An agent repeating your returns policy is only as good as the policy. In the UK, most online sales are distance contracts, and the Consumer Contracts Regulations 2013 generally give buyers 14 days to cancel after delivery and require a refund within 14 days of getting the goods back. In the US, the FTC's Mail, Internet, or Telephone Order Merchandise Rule expects you to ship within the time you state, or within 30 days if you state none, and to tell customers about delays and offer to cancel. Canadian provinces and Nigeria's consumer protection law set their own rules; see our guides for the UK and Nigeria.
Two practical consequences follow. First, if your written policy offers less than the law, the agent will repeat the shortfall, so have the policy checked by an adviser before you point an agent at it. Second, the business is responsible for what its agent says. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable for a refund promise its website chatbot made, rejecting the argument that the chatbot was responsible for its own words.
A worked example
Illustrative only. A homeware shop ships about 600 orders a month and receives roughly 450 support messages. The owner sorts a month of messages and finds:
- Where is my order
- about 210 messages
- Returns and exchanges within policy
- about 95
- Delivery costs and times
- about 50
- Product questions
- about 45
- Damaged or wrong item
- about 30
- Refund requests outside policy, complaints, other
- about 20
The first four groups, about 400 messages, are answerable from records and policy. The last two, about 50, need a person. The owner sets the agent to answer the first group, to open a case with photos requested for damaged items and pass it to her, and to refuse to promise anything outside the returns window. Her own inbox drops to the cases that need judgement. The numbers here are an example, not a forecast; your own split will differ.
Getting the setup right
- Write down the policies the agent may use. Returns, delivery, warranties, price-matching. If it is not written, the agent should not say it.
- Connect it to real order data so status answers come from the shipment record, not from the customer's description.
- Set a ceiling. Decide what the agent may complete alone, such as a standard return label, and what it must pass up.
- Read a sample of replies each week for the first months, and fix the policy text when an answer is unclear.
- Track repeat contacts. A customer writing three times about one order is a sign the agent, or the process, is failing.
Where Pryme fits
A Pryme Business account comes with five AI agents. The Customer Support agent answers only from the policies you point it at, attaches the source to each reply and hands a case to a named person when an answer falls below its evidence floor or a request is above its ceiling. It cannot move money or grant exceptions itself: a refund within policy is booked with the Finance agent under the approvals you set, and every action is on the audit trail. The agents are built and certified on Pryme Intelligence, the governed agentic operating system they run on. For channels and setup, ask Pryme. For how the five agents divide the work, see Business Account agents explained.
Questions people ask
Can an AI agent process refunds for my shop?
It can prepare or complete refunds that fall clearly within your written policy, if you allow it. Refunds outside policy, disputes and chargebacks should be decided by a person.
What share of e-commerce support can be automated?
It depends on your products and policies. Sort a month of real messages by type; the share answerable from order records and written policy is your realistic ceiling.
Am I liable if the agent gives a wrong answer?
Generally yes. Courts and tribunals have treated a business's chatbot as speaking for the business. Keep answers tied to written policy and review them.
Should the agent handle chargebacks?
It can gather the order record and messages into an evidence pack, but a person should decide whether to contest and submit the response.
Sources
- Consumer Contracts (Information, Cancellation and Additional Charges) Regulations 2013 · accessed Oct 2026
- FTC: Mail, Internet, or Telephone Order Merchandise Rule · accessed Oct 2026
- Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal) · accessed Oct 2026
- FTC: Chatbots, deepfakes, and voice clones: AI deception for sale · 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.