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 should hand over when it cannot find a source for the answer, when money involved is above its ceiling, when the customer asks for a person, and when the case involves a complaint, vulnerability, a legal threat or a safety issue. The handover should pass the full history, the relevant records and a short summary, so the customer never has to start again.
Escalation is the part of AI customer support that customers remember. A routine answer delivered in seconds is pleasant; being stuck with a bot that will not let you reach anyone is infuriating, and it is one of the problems the US Consumer Financial Protection Bureau highlighted in its 2023 report on chatbots, where customers described “doom loops” of repeated, unhelpful replies. Good escalation is designed in from the start.
The triggers
Triggers fall into three groups: what the agent knows, what the case involves, and what the customer wants.
| Trigger | Example | Why a person |
|---|---|---|
| No source | “Do you price-match?” and no policy says | Guessing creates promises |
| Below evidence threshold | Two clauses could apply | Ambiguity needs judgement |
| Above ceiling | Refund of £900 when ceiling is £200 | Money decisions need a named approver |
| Exception requested | Return after the window, “as a loyal customer” | Goodwill is a business choice |
| Complaint | “This is the third time…” | Complaints need ownership and a fix |
| Vulnerability | Bereavement, illness, financial hardship | Care and discretion |
| Legal or regulatory | Mention of a lawyer, ombudsman or regulator | Needs a person and maybe an adviser |
| Safety | Injury, faulty electrical item, food reaction | Possible reporting duties |
| Customer asks | “Can I speak to someone?” | Respect the request |
The last trigger is non-negotiable. If a customer asks for a person, they should get one, even if the agent thinks it can answer. Under UK GDPR, where a decision based solely on automated processing has legal or similarly significant effects, people have rights including to obtain human intervention; most support replies will not reach that bar, but refusing someone a refund or a service might, so take advice if your agent makes decisions of that kind.
A fourth group is less obvious: patterns. A customer who has contacted you three times in a week about the same order should reach a person on the fourth contact even if each message on its own looks routine. So should a case that has been open longer than your promised response time, and a message that arrives on one channel about a case already being handled on another. These signals sit across conversations rather than inside one, which is why the agent needs the full case history rather than the latest message alone.
What a good handover contains
- The conversation so far, in full, across channels.
- The customer and their records: orders, payments, previous cases.
- A two-line summary: what the customer wants and why the agent stopped.
- The relevant policy clause, or a note that none exists.
- The clock: how long the customer has been waiting and any deadline.
- A suggested reply, clearly marked as a draft, if the agent has one.
The person receiving it should be able to act within a minute without reading back through the thread. The customer should be told plainly that a person is taking over and roughly when they will hear back. Do not pretend the agent is a person, and do not leave the customer guessing; see telling customers they are talking to AI.
Handover is also a moment where trust is won or lost. A short, specific message works better than an apology script: “I have passed this to Ada, who handles refunds over £200. She will reply by 3pm today. You do not need to send anything else.” It names a person or team, gives a time and removes the customer’s uncertainty about what to do next.
Setting the thresholds
Start cautious and relax with evidence. In the first weeks, set a low money ceiling, a high evidence threshold and a long list of excluded topics. Then read what was escalated. If many escalations were simple and the draft was right, lower the threshold for that type of question. If any sent answer was wrong, raise it.
Worked example (illustrative)
A small travel agency starts with every refund going to the owner. After a month, 70 of 90 refund cases were full refunds for cancellations made more than 30 days before departure, which the written policy covers exactly. The owner allows the agent to prepare those refunds up to a set amount for a single approval tap, and keeps everything else, including partial refunds and anything involving an airline, with a person. The numbers are an example; set your own from your own cases.
Who receives the escalation
Every trigger needs a named owner, not a shared inbox nobody watches. In a one-person business that is the owner, with a clear daily time to clear the queue. In a small team, split by type: money to whoever approves refunds, complaints to the most senior person available, sales enquiries to sales. An AI chief of staff can do this routing and bring the owner only the cases that need a decision. Our explainer on decisions a person should make sets out the same principle for finance.
- Set a response time for each escalation type and show the customer that time.
- Track escalations that wait too long, not just average waiting time.
- Review why cases were escalated each week, because repeated escalations point to a missing policy.
Where Pryme fits
The Pryme Customer Support agent escalates any answer below its evidence floor and any request above its ceiling, passing the case to a person intact with its history and evidence pack. It cannot grant exceptions or move money. The Chief of Staff reads across all the agents and brings you only what needs your decision, so a disputed payment and a related complaint arrive together. Every handover is recorded on the audit trail.
Questions people ask
How do I make sure customers can always reach a person?
Make “speak to a person” a trigger that always works, on every channel, and tell the customer when to expect a reply. Test it yourself regularly.
What should the AI tell a customer when it hands over?
That a named team or person is taking the case, roughly when they will reply, and that they will not need to repeat themselves.
How many cases should go to a person?
It depends on your records and customers. Start with more escalations than you need and reduce them only where review shows the agent’s drafts were right.
Does UK GDPR require human review of AI decisions?
Where a solely automated decision has legal or similarly significant effects, people have rights including to obtain human intervention. Check with an adviser whether your use reaches that bar.
Sources
- Chatbots in consumer finance (Consumer Financial Protection Bureau) · accessed Oct 2026
- Rights related to automated decision making including profiling (ICO) · accessed Oct 2026
- ISO 10002: Guidelines for complaints handling in organisations (ISO) · 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.