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 agent can build a sales forecast by weighting each open deal according to its stage and the evidence behind it, such as a signed proposal or a confirmed budget, and by showing the source for every figure. It improves consistency, but a person should still commit the final number, because the agent cannot know everything a customer has said off the record.
A forecast answers a simple question: how much will we sell, and when? For a small business the answer drives real decisions, such as whether to hire, how much stock to buy, or whether there will be enough cash to pay suppliers next month. Yet most forecasts are built from memory and optimism. Gartner reported in 2020 that fewer than half of sales leaders and sellers had high confidence in their organisation's forecast accuracy, and those were large companies with dedicated teams.
How a forecast is usually built
The common method is a weighted pipeline. Each stage gets a probability, each deal's value is multiplied by it, and the results are added up. The table below shows an illustrative set of weights.
| Stage | Meaning (example) | Weight |
|---|---|---|
| Enquiry | Contact made, need not confirmed | 5% |
| Qualified | Budget, decision-maker and timing confirmed | 20% |
| Proposal sent | Written proposal with price | 40% |
| Verbal yes | Customer has said yes, paperwork pending | 75% |
| Signed | Contract or order received | 95% |
The weights are where forecasts go wrong. They are often set once and never checked, and deals are placed in stages on hope. If your proposals actually win one time in five, a 40% weight doubles the forecast for that stage.
What an AI agent changes
- Evidence for every stage. The agent places a deal in a stage only if there is a message, document or note that supports it, and it links to that source.
- Weights from your history. It can calculate how often deals at each stage actually closed over the last year and suggest weights based on that, not on a textbook.
- Timing from behaviour. If deals of a certain size typically take 45 days from proposal to signature, a deal sent yesterday is unlikely to close this week.
- Stalled deals separated. Deals with no recent movement can be shown separately or weighted down, so they do not prop up the total.
- Currency handled properly. Deals in several currencies are shown in their own currency and converted at a stated rate for the total.
The result is not a prediction of the future so much as a checkable summary of the present. You can click on any number and see why it is there. That is what makes it possible to argue with the forecast, which is healthy.
Where AI forecasting still fails
An agent only knows what is recorded. It does not know that the customer's finance director told your account manager over coffee that budgets are frozen. It cannot predict a competitor's price cut or a currency shock. It can also be confidently wrong if your historical data is thin or messy: twelve deals last year do not make a reliable win rate.
For these reasons, a sensible split is that the agent prepares the forecast and a person commits it. The person adjusts for what they know that the records do not, and the adjustment is recorded with a reason. Over time, comparing the agent's figure, the committed figure and the actual result tells you who is better calibrated and where. The NIST AI Risk Management Framework makes the same point in general terms: test what an AI system produces against reality, and keep a person accountable for decisions based on it.
Small numbers, big swings
Small businesses face a particular problem: one or two large deals can make or break a quarter. A weighted forecast spreads those deals across probabilities, which is mathematically tidy but practically misleading. A £200,000 deal at 40% contributes £80,000 to the forecast, yet the real outcome will be either nothing or £200,000. It is often more useful to show large deals separately, with a best case and a worst case, and to plan cash against the worst case.
An agent can present the forecast this way automatically: a base made of many smaller deals, where weighting works reasonably well, and a short list of large deals that each need a person's judgement. It can also show how the forecast has changed week by week, which reveals whether optimism is creeping in.
A worked example
Illustrative. A Nairobi-based software firm selling to schools in three countries asks its agent for the quarter's forecast. The weighted pipeline is US$310,000. The agent notes that, on the firm's own history, proposals to schools close at 22%, not the 40% weight in use, and that eight deals worth US$96,000 have not moved in 30 days. Applying historical weights and excluding stalled deals gives US$171,000. The founder knows two of the stalled deals are waiting on a term-start budget release she has confirmed by phone, and adds them back at 50%: US$189,000, with her note attached. The quarter closes at US$182,000.
Her committed figure came within 4% of the result, while the original weighted pipeline would have been 70% too high. That difference would have shaped a hiring decision she was due to make that quarter.
Where Pryme fits
The Pryme Sales agent forecasts from deals stage by stage, weighted across the currencies you trade in, and every number cites the conversation it came from. It will not commit a forecast number that no certified desk has accepted, however much the quarter needs it. The Finance agent can use the committed forecast for the cash view, and the account handles several currencies. Before forecasting, see pipeline hygiene, and for cash planning, AI agents for business finance.
Questions people ask
How accurate is AI sales forecasting?
It depends on the quality and volume of your data. An agent improves consistency and shows its evidence, but it cannot see what is not recorded.
What stage weights should I use?
Your own historical close rates by stage, where you have enough data. Textbook weights are a starting point, not a rule.
Should the AI or a person commit the forecast?
A person. The agent prepares it with sources; the person adjusts for what they know and records why.
How do I forecast in several currencies?
Keep each deal in its own currency and convert to a reporting currency at a stated rate, so exchange movements do not hide in the total.
Sources
- Gartner: less than 50% of sales leaders and sellers have high confidence in forecasting accuracy (2020) · accessed Oct 2026
- NIST AI Risk Management Framework · 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.