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
AI agents cannot take legal or moral responsibility for decisions, cannot know what is not in their records or instructions, and cannot guarantee that what they produce is accurate. They also cannot safely judge situations that call for discretion, such as exceptions, disputes or decisions about people, so those need a person, with the agent preparing the evidence.
Most writing about AI agents lists what they can do. Planning well depends just as much on knowing where they stop. The limits below are not failings of one product or another; they follow from how agents work. A business that designs around them gets the benefit of agents without being surprised by them.
They cannot carry responsibility
When an agent sends a wrong reply or releases a wrong payment, the business is answerable, not the software. Directors' duties, consumer law, tax rules and data protection obligations all fall on the business and the people who run it. In the UK, for example, the ICO's guidance on AI and data protection makes clear that the organisation using AI remains accountable for how personal data is processed. Agents can prepare, recommend and act within limits, but a person must own the outcome.
This has a practical consequence. Whoever owns an outcome needs to be able to see how it came about. If an agent can act without leaving a trace, you carry the responsibility without the means to discharge it. That is why the record of each action and its source is not a technical nicety but the thing that lets an owner, an accountant or a regulator understand what was done in the business's name.
They cannot know what you have not told them
An agent works from its instructions and the records it can fetch. It does not know that a long-standing customer gets a softer approach, that a supplier agreed a discount on the phone, or that you never ship to a certain region. If it is not written down and available, it does not exist for the agent.
- Unwritten policies are applied inconsistently or not at all.
- Records kept on paper or in personal inboxes are invisible to it.
- Context from conversations, such as a promise made at a meeting, is lost unless someone logs it.
Our data readiness checklist shows how to close the most common gaps.
They cannot guarantee accuracy
Language models can produce fluent, confident output that is wrong. NIST's generative AI profile lists this, which it calls confabulation, among the main risks of the technology. Agents reduce the risk by working from your records and citing sources, but they can still misread a document, match the wrong transaction or carry an early error through later steps.
The practical response is not to avoid agents but to design for error: ceilings that limit the size of any single mistake, records that let you trace it, and sample checks that catch patterns early.
Example only. An agent matching 400 customer payments a month gets 98% right. That sounds good until you see that the 8 it got wrong include a $6,000 payment posted to the wrong customer, who then received a reminder for an invoice they had paid. A ceiling that sends any match above $2,000 to a person when references are unclear, and a weekly look at a sample of matches, would have caught it.
They cannot exercise discretion well
Some decisions depend on judgement that is hard to write down: whether to make an exception for a loyal customer, how to handle a complaint from someone in difficulty, whether to extend credit to a struggling client. Agents can apply rules, but discretion means knowing when to set a rule aside. That is a person's job.
| Decision type | Agent's role | Person's role |
|---|---|---|
| Exceptions to policy | Spot the request, gather history and the relevant clause | Decide whether to make the exception |
| Disputes | Assemble the evidence and timeline | Judge the case and respond |
| Pricing above a band | Draft the proposal at standard terms | Set the price |
| Decisions about people | Prepare records, such as hours or payments | Decide on hiring, discipline, pay or credit |
| Vulnerable customers | Flag signs and route to a person quickly | Handle the conversation |
Decisions about individuals also carry legal weight. Under UK GDPR, people have rights in relation to solely automated decisions that have legal or similarly significant effects on them, as the ICO explains, and other countries have their own rules. Keeping a person in such decisions is both good practice and, in many cases, a legal requirement.
Other practical limits
- Physical work
- Agents cannot inspect goods, count stock on a shelf or visit a site. They depend on someone recording what was found.
- Relationships
- They can draft a warm message, but trust with a key client is built by people.
- Strategy
- They can summarise performance and options; deciding what the business should become is the owner's call.
- Hostile input
- They can be misled by instructions hidden in emails or documents, a risk OWASP calls prompt injection. Narrow permissions limit the damage.
- Knowing when they are wrong
- They do not reliably sense their own errors. Checks have to be designed in.
Where Pryme fits
Pryme's agents are built around these limits. Each works inside a mandate with a ceiling, and above it a person decides. Every action is recorded with its source. The Customer Support agent answers from your own records and policy and escalates disputes; the Compliance agent records refusals where its policy set is silent rather than improvising. Read our guide to human in the loop decisions for the finance side.
Questions people ask
Can an AI agent be held legally responsible?
No. Responsibility stays with the business and the people who run it, which is why ceilings and approvals matter.
Will AI agents get better at judgement?
Models are improving at reasoning, but discretion depends on context, values and accountability that a business should keep with people.
Can an agent tell when it is unsure?
Some systems flag low confidence or missing sources, but this is not reliable on its own. Design checks rather than relying on the agent to know.
Should agents ever handle complaints?
They can acknowledge, gather facts and route complaints quickly. Judging and resolving a complaint, especially from a vulnerable customer, should involve a person.
Sources
- ICO: Guidance on AI and data protection · accessed Oct 2026
- ICO: Rights related to automated decision-making including profiling · accessed Oct 2026
- NIST AI 600-1: Generative AI Profile · accessed Oct 2026
- OWASP: LLM01:2025 Prompt Injection · 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.