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 be set up to notice signs that a customer may be vulnerable, such as bereavement, illness, financial difficulty or trouble understanding, and to adapt or hand over to a person rather than continue a standard flow. It should never be the only route for these customers, and the business should make reasonable adjustments, such as other channels and accessible formats, available.
Any customer can become vulnerable: after a bereavement, during illness, after losing a job, or simply when dealing with something they find hard to understand. Support processes designed for the average customer can fail them badly, and automated processes can fail them faster. An AI agent can help, by noticing signals a busy team might miss and routing those customers to a person. It can also harm, by repeating a policy to someone who needed patience.
What vulnerability looks like
The UK Financial Conduct Authority's guidance on the fair treatment of vulnerable customers (FG21/1) is written for financial firms, but its framework is useful for any business. It groups the drivers of vulnerability into four areas:
| Driver | Examples a support agent might see |
|---|---|
| Health | Mentions of illness, hospital, mental health, disability |
| Life events | Bereavement, divorce, job loss, caring responsibilities |
| Resilience | Struggling to pay, debt, irregular income |
| Capability | Difficulty reading or understanding, low digital confidence, language barriers |
Signals are often indirect. "Sorry for the delay, things have been hard since my husband passed" is a bereavement disclosure inside a routine question. An agent should be configured to recognise phrasing like this, not just keywords.
What the agent should do differently
- Change the pace. Stop sending reminders, chasers or upsell messages while the case is open.
- Stop repeating the policy. Once a customer has said they cannot pay, quoting the amount again does not help.
- Offer a person early. Say plainly that a member of the team can help, and make that easy.
- Simplify. Shorter sentences, one step at a time, no jargon.
- Record what the customer has shared, with care, so they do not have to explain again. Health information is sensitive data in most data protection regimes, so record only what is needed, with access restricted.
- Never use the disclosure against the customer, for example by tightening credit terms because they mentioned an illness, without a proper and lawful process run by a person.
Where an agent helps, and where it falls short
An agent has some advantages here. It reads every message, at any hour, and does not get tired or irritated by the fifth message of the day from the same person. It can be configured to treat a disclosure the same way every time, which people under pressure do not always manage. It can also apply a pause on reminders instantly, before the next automated chaser goes out.
Its weaknesses are real too. It can miss signals that are implied, sarcastic or written in a second language. It can sound warm without understanding. And it cannot exercise the discretion that good treatment of vulnerable customers often needs, such as waiving a fee or agreeing a plan that fits someone's situation. Design for this: set the threshold for handing over low, review a sample of conversations that were not flagged, and make sure the people receiving the cases have the authority to help.
Access and reasonable adjustments
Equality and accessibility law in many countries expects businesses to make adjustments so disabled customers can use their services. In Great Britain, the Equality Act 2010 places a duty on service providers to make reasonable adjustments. In the US, the Department of Justice has published guidance on web accessibility under the Americans with Disabilities Act. The details differ, but the principle for AI support is consistent:
- A chat agent should not be the only way to reach you. Keep phone, email or in-person routes available.
- The chat interface should work with screen readers and keyboard navigation.
- Customers should be able to ask for information in other formats.
- A person should be reachable for anyone who cannot use the agent.
These are general points. Check your obligations with an adviser.
A worked example
Illustrative only. A home energy installer collects monthly payments on finance agreements and uses an AI agent for support. A customer writes: "My payment will be late again, I have been in and out of hospital." A standard flow would quote the overdue balance and the late fee. Instead, the agent:
- Acknowledges
- "Thank you for telling us. I am sorry you have been unwell."
- Pauses
- Automatic payment reminders on the account for 14 days
- Hands over
- Passes the case to the accounts lead, flagged as a possible vulnerability, with the message history
- Tells the customer
- That a named person will call or write within one working day to discuss options
- Records
- A short note of the health disclosure, visible only to the accounts team
The accounts lead, not the agent, decides whether to waive the late fee or agree a plan. The agent's value was in noticing, pausing and handing over cleanly. For when to keep cases with people more broadly, see when not to automate customer support.
Where Pryme fits
In a Pryme Business account, the Customer Support agent answers from your policies and hands cases to a named person when an answer falls below its evidence floor or a case is outside its scope, with the history attached. It cannot grant exceptions or move money; fee waivers and payment plans are decisions for your team, recorded by the Finance agent under the approvals you set. The Compliance agent can keep track of your policies, and every action is on the audit trail. For finance firms specifically, see AI customer support for financial businesses.
Questions people ask
How can an AI agent recognise a vulnerable customer?
By being configured to notice signals such as mentions of illness, bereavement, job loss or difficulty understanding, including indirect phrasing, and to hand over to a person.
Should vulnerable customers be kept away from AI agents?
Not necessarily, but the agent should not be their only route, and it should hand over early when there are signs of vulnerability.
Is the FCA vulnerability guidance only for financial firms?
FG21/1 applies to FCA-regulated firms, but its framework of health, life events, resilience and capability is a useful guide for any business.
What adjustments should a business make?
Keep other channels open, make the chat interface accessible, offer other formats and make a person reachable. Check local equality and accessibility law with an adviser.
Sources
- FCA: FG21/1 Guidance for firms on the fair treatment of vulnerable customers · accessed Oct 2026
- Equality Act 2010 (legislation.gov.uk) · accessed Oct 2026
- ADA.gov: Guidance on web accessibility and the ADA · accessed Oct 2026
- ICO: Guidance on AI and data protection · 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.