For accountants
From assistant to agent: what agentic AI really changes about forecasting
For accounting practices weighing up what to adopt, what to resist, and what to charge for.
The language changed in 2026. Last year the pitch was a copilot: something that sat beside you, answered questions and drafted commentary. This year it is agents, plural, described as a virtual team that does the work rather than helps you do it. Intuit has launched exactly that framing for QuickBooks. Xero's chief product and technology officer has said that 2026 is about moving from AI as a feature to AI as the core engine.
The sentiment data has moved with it. Deloitte's Q2 2026 UK CFO Survey found 73% of CFOs more optimistic about AI's business impact than they were twelve months earlier, up from 39% in Q3 2024. Worth noting that this is a survey of 58 large-corporate finance chiefs, so it tells you where the FTSE mood is heading rather than what your owner-managed clients are doing. Consero's Q1 2026 survey of 102 mid-market finance leaders, US and investor-backed, found 42% with AI broadly or fully embedded in finance, up from 22% the year before.
So the direction is not in doubt. The useful question for a practice is narrower: which parts of a forecasting engagement can you actually hand to an agent, and which parts still have your name on them?
Where agents genuinely earn their keep
Four jobs in a typical forecasting cycle are now safely automatable, because each one is checkable in seconds and cheap to get wrong.
Data preparation. Pulling actuals from Xero, QuickBooks or Sage, spotting unmapped accounts, flagging periods that have moved since last month, catching a supplier who has been coded three different ways. This is the least glamorous part of the job and by some distance the biggest time sink.
Variance commentary drafting. An agent that reads a month's movement and produces a first draft of "why" saves a client manager forty minutes and produces something a human can correct faster than they could write from scratch.
Assumption capture. Turning a client's messy voice note about a new hire in October and a price rise in January into structured, dated drivers that sit properly in the model.
Consistency checking. Does the balance sheet still balance after the forecast overlay. Does the cash flow reconcile to the movement in cash. Has anyone overwritten a historical actual. Machines are better at this than tired people on a Thursday afternoon.
Consero's respondents put the fastest payback in management reporting and variance analysis, at three to six months. That matches what the four jobs above have in common: high volume, low judgement, immediately verifiable.
Where forecasting is different
Bookkeeping errors sit still. Forecasting errors compound forward. Miscategorise one invoice and you have one wrong line. Get a growth driver wrong and every month to the horizon is wrong, the KPI built on it is wrong, and the covenant headroom your client is showing their bank is wrong.
That asymmetry is why three things should stay with a person.
Choosing the drivers. Whether revenue is best modelled on headcount, on pipeline or on seasonality is a judgement about the business, not a pattern in the ledger. An agent will happily fit something to twenty-four months of history that has no causal basis at all.
Sense-checking the shape. Clients recognise when a forecast looks wrong before they can explain why. That instinct is the most valuable review control in the process and it does not transfer.
Owning the number. As one panel at a recent finance AI discussion put it, if you put the numbers in the spreadsheet, they are your numbers, and AI is no different. Your professional indemnity insurer takes the same view.
There is a practical constraint underneath all of this. Consero's respondents named data readiness the single biggest blocker to getting a return from AI, ahead of everything else. If a client's chart of accounts is a decade of accumulated improvisation, an agent will not save you. It will produce the same wrong answer faster and with more confidence.
The dividing line, in practice
| Hand to an agent | Keep with a person |
|---|---|
| Pulling and mapping actuals from the ledger | Choosing the forecast drivers |
| First-draft variance commentary | Sense-checking the shape of the output |
| Structuring client assumptions into dated drivers | The client conversation about what to do next |
| Integrity checks across the three statements | Signing off and owning the number |
What this means for how you run the work
The firms getting value are not the ones buying the most tools. They are the ones that have redrawn the line between preparation and judgement, then priced accordingly.
That looks like three changes.
First, stop selling preparation time. If an agent compresses the data prep on a monthly forecasting engagement from four hours to forty minutes, a time-based fee turns your own efficiency into a pay cut. Price the output.
Second, build the review layer deliberately. Decide in advance what a human checks every single month, write it down, and make it a signed step. Regulators, insurers and clients will all eventually ask you what your control was, and "the system did it" is not an answer.
Third, insist on seeing the reasoning. The market has already moved on this. Reported research suggests 72% of financial institutions are only partially aware which of their vendors are using AI at all, and finance leaders are increasingly refusing outputs they cannot trace back to a source line. Xero calls its version "accountable intelligence", requiring the model to show its working at each step. Whatever the vendor calls it, the test is the same: can you click a forecast figure and see the calculation, the inputs and the assumption behind it. If you cannot, you cannot defend it, and you should not be putting your name on it.
The uncomfortable conclusion
Agentic AI does not make forecasting easier. It makes the mechanical part nearly free, which means the value of an engagement now sits almost entirely in judgement, in review, and in the client conversation about what to do differently. That is a good trade for practices that were already selling advice. It is a difficult one for anyone whose forecasting service was really a spreadsheet-building service.
The number still has your name on it. That has not changed, and it is the part worth charging for.
Sources
See your own forecast in about five minutes.
Connect Xero or QuickBooks and build a live, connected model with no spreadsheet to maintain.