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
When you change a policy that an AI agent applies, the agent's mandate changes too, and its earlier testing no longer covers the new rule. A safe change is versioned, approved by a named person, tested against real cases and only then switched on, with the date recorded so past answers can be traced to the old version. Where an agent is certified, a material change should mean certification is earned again before it acts under the new rule.
Why a policy change is a change to the agent
An AI compliance agent that answers only from your policy set behaves the way your policies tell it to. Change a refund window from 14 to 30 days, raise a payment ceiling, add a country or remove a prohibition, and you have changed what the agent does. That is the point of policy-driven agents: you control behaviour through rules you can read, not settings buried in software.
It also means every policy change carries the risks of a software release. A clause that reads well can still produce surprising answers when applied to hundreds of cases. NIST's AI Risk Management Framework treats ongoing measurement and management of AI behaviour as continuous, not one-off, and guidance from the UK Information Commissioner's Office on AI stresses documenting changes to systems that affect people.
Changes that matter most
| Change | Why it is material | Extra care |
|---|---|---|
| Raising a ceiling | More happens without a person | Check recent cases near the new line |
| Adding a country or currency | New rules, new lists, new risks | Add obligations and screening lists first |
| Removing a prohibition | Something refused becomes allowed | Record why and who approved |
| New data source | Answers may draw on new records | Check quality and access rights |
| Wording clarification | Can still change outcomes | Test against cases that used the old wording |
Not every change is material. Fixing a typo or reformatting a clause need not trigger a full review. But anything that changes what happens, who decides or how much moves without a person should be treated as a change to the mandate.
Who owns the policy set
Every policy set needs an owner, usually the business owner in a very small firm or an operations or compliance lead in a larger one. The owner decides when a change is material, approves new versions and makes sure each change is tested. Without a single owner, policies drift: one person raises a ceiling to clear a backlog, another adds a clause after a complaint, and nobody sees how the pieces now fit together. A short change log, with the date, the clause, the reason and the approver, is usually enough to prevent this.
A safe change in five steps
- Draft the new version alongside the old, with a reason for the change.
- Approve it. A named person signs off and the approval is recorded.
- Test it. Run recent real cases through the new clause and compare outcomes with the old one. Look for cases that change from refuse to allow, or from allow to refuse.
- Switch on a set date, not mid-way through a batch of work.
- Watch the first weeks. Check refusals, holds and decisions under the new rule and adjust if outcomes surprise you.
Keep the old version
If a customer, regulator or auditor later asks why an answer was given in March, you need the policy as it stood in March. Keep every version with its effective dates. An agent that cites the clause and version with each answer makes this simple; one that cites nothing makes it close to impossible.
Tell the people affected
A rule change often affects staff and customers as well as the agent. If a refund window lengthens, customers may need to see the new terms; if a payment ceiling rises, approvers need to know which payments will no longer reach them. Communicate the change before it takes effect, so that nobody is surprised by what the agent does on the first morning under the new rule.
A worked example
Example only. A UK training provider raises the limit below which its agent can approve course refunds from £200 to £500. Before switching, it runs the last three months of refund requests through the new rule. Forty-one requests that went to the operations manager would now be handled within policy. Thirty-eight look fine. Three involved learners who had already attended most of the course, which the old ceiling had caught by accident. The provider adds a clause requiring a person to decide any refund where attendance exceeds half the course, then switches both clauses on together on the first of the month. The manager's refund queue drops by about two-thirds, and the three awkward cases still reach a person.
For drafting clauses an agent can apply, see how to write a policy set an AI compliance agent can use. For where the line should sit on money, see finance decisions a person should make.
Where Pryme fits
Pryme Business agents are built, trained and certified by Pryme Intelligence. They are certified before they act and must re-earn certification when their mandate or data changes. Each agent works inside a mandate with a ceiling, and the Compliance agent answers only from the policy set you load, with the clause attached to the answer. Every action is recorded with its source.
Questions people ask
Does every policy edit need re-testing?
No. Typos and formatting do not. Changes to what happens, who decides or how much moves without a person do.
How long should I keep old policy versions?
Long enough to explain past decisions, in line with your records policy and any legal retention periods that apply. This is general information, not legal advice.
Can the agent apply a new rule immediately?
It can, but safer practice is to test first and switch on a set date, so you know which rule applied to which case.
What if the new rule produces bad outcomes?
Roll back to the previous version, record why, and fix the clause before trying again.
Sources
- NIST: AI Risk Management Framework · accessed Oct 2026
- ICO: Guidance on AI and data protection · accessed Oct 2026
- COSO: Internal control · 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.