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
In the first 30 days, let an AI chief of staff read before it routes: week one, read-only briefs you check against reality; week two, routing of routine items; week three, ceilings and decisions; week four, a review that adjusts the rules. Keep its authority small, check its sources, and widen its scope only when the record shows it is right.
Most disappointing AI projects fail in the first month, not because the software is bad but because it was given too much, too soon, with no way of checking it. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. A slow, deliberate first month addresses all three. It also follows advice Anthropic gives to people building agents: find the simplest solution possible and add complexity only when it is needed.
Before you start
- Write the mandate. What it may read, where it may route, what it must raise, what it must never do. See how to write a mandate.
- Count your baseline. For one normal week, note how many items reach you a day, how many need a decision, and how long you spend in inboxes.
- Name the owners. For each type of work, which agent or person owns it.
- Name a deputy for decisions when you are away.
The baseline is easy to skip and hard to recreate later. Without it, you will not know whether the month helped.
Also decide what the chief of staff will not read in the first month. Payroll, personal messages and anything confidential to a client can wait until you have seen how it handles ordinary records. Adding a source later is easy; taking back something it has already summarised to the wrong person is not.
Week by week
| Week | Focus | What you do | Done when |
|---|---|---|---|
| 1 | Reading | Read the brief and check five items a day against the source | The brief matches reality for three days running |
| 2 | Routing | Let it route routine items; re-route any it gets wrong, noting why | Fewer than one in ten routes need correcting |
| 3 | Ceilings | Set starting ceilings for each agent; decide every item raised | Every decision arrives with a reason and evidence |
| 4 | Review | Look at what you approved unchanged, overruled and what was refused | You have changed at least one rule based on the record |
Week 1: reading only
The chief of staff reads finance, sales, support and compliance and sends a morning brief, but routes nothing. Your job is to check it. Pick five lines a day and open the source. Is the overdue invoice really overdue? Did the customer really ask for a refund? Errors this week are cheap and tell you where its reading is weak, often a system it cannot see properly.
Week 2: routing
Allow it to hand routine items to their owners: invoices to finance, enquiries to sales, complaints to support, deadlines to compliance. Keep anything unusual with you. Each time you re-route an item by hand, note the reason; those notes become better routing rules.
Week 3: ceilings
Set a starting ceiling for each agent that can commit money: a release line for finance, a discount band for sales, a refund limit for support. Start low. Every item above a ceiling comes to you as a decision. Expect a busy week, which is the point: you are learning which decisions you make the same way every time.
Week 4: review
Look back. If you approved twenty refunds under £75 without change, that limit can probably rise. If you overruled the same kind of discount three times, the band is wrong or the sales agent is reading the customer history badly. If items were refused because no mandate covered them, decide who owns that work.
Throughout the month, keep the agents' own powers where they were on day one unless the record supports a change. The temptation in week three is to raise every ceiling because the decisions are piling up. Raise only the ones where your approvals show a clear pattern, and write down why.
What to measure at day 30
- Items reaching you each day, against your baseline.
- Decisions as a share of those items. It should rise: more of what reaches you should need you.
- Routing corrections per week.
- Promises kept on time, from the promises register.
- Decisions waiting more than 24 hours.
More on reading these numbers in how to tell if it is working.
A worked example
Example only. A Calgary property maintenance firm with nine staff records a baseline of about 70 items a day reaching the owner, of which 8 need a decision. By day 30, about 15 items reach the owner a day, of which 7 are decisions. Routing corrections fall from 14 in week two to 3 in week four. The owner raises the support refund limit from C$100 to C$200 after approving every refund in between unchanged, and names an office manager as deputy for supplier payments. One refusal pattern stands out: emergency call-out requests on weekends, which no agent owned. The owner assigns them to the on-call engineer and adds the rule to the mandate.
Where Pryme fits
With Pryme Business there is nothing to install: the Chief of Staff and the other four agents come with the account. Each works inside a mandate with a ceiling you set, every action is recorded with its source, and agents are certified on Pryme Intelligence before they act. See how it works.
Questions people ask
Can I skip the read-only week?
You can, but it is the cheapest way to find where the agent misreads your business. Errors found later cost more.
What if the brief is wrong a lot in week one?
Check which source the errors come from. It is usually one system the agent cannot read fully, which is fixable.
Should staff be involved?
Yes. Tell them what the chief of staff will route to them and ask them to flag misroutes.
What comes after day 30?
A monthly review of ceilings and a weekly look at refusals. Widen its scope one recorded change at a time.
Sources
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027 (June 2025) · accessed Oct 2026
- Anthropic: Building effective agents · 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.