AI can take the rebuilding out of month-end, but not the judgement
Updated: Oct 28
In many food manufacturers and wholesalers, month-end follows a familiar ritual. Someone spends three or four days pulling exports from the accounts system, the stock system and a handful of spreadsheets. They clean them. Then they rebuild last month's reports. Over a year that is around forty to fifty days of a skilled person's time (illustrative). Perhaps half of it is mechanical. That half is where AI tools are now useful. The other half is judgement about what the numbers mean. There, a confident wrong answer costs far more than a slow close.
Adoption is wide, but the gains are still thin
Use has spread quickly. The ONS reports that the share of UK businesses with ten or more employees using at least one AI technology rose from about 12% in September 2023 to about 35% in June 2026. Only around 10% of those firms describe their use as extensive. The ONS also notes that manufacturers and wholesalers most often come to AI through free-to-use software.
Finance teams tell the same story. Gartner's 2025 survey of 183 finance leaders found 59% of finance functions using AI, against 58% in 2024 and 37% in 2023. Yet 91% reported low or moderate impact at first. McKinsey's State of AI survey, published in November 2025, found 88% of organisations using AI in at least one function. Only 39% reported any effect on profit at the level of the whole business. Nearly a third reported consequences from AI getting things wrong. Enthusiasm, in other words, has run some way ahead of results.
The mechanical half of month-end suits AI
The useful tasks have three things in common. They follow clear rules. They recur every month. And a person can check the result. In a food business that usually means the following. Pulling and reshaping exports into a consistent layout. Matching records that ought to agree, such as goods received against purchase invoices, or stock movements against production records, and flagging those that do not. Picking out lines that have moved more than usual, so someone looks at them first. Drafting the commentary for the management pack from the numbers, for a person to edit.
This is much what finance teams already use AI for. In Gartner's 2025 survey, the most common uses were knowledge management (49%), accounts payable automation (37%) and error and anomaly detection (34%). In McKinsey's 2024 CFO survey, 71% of finance functions already using AI said it had raised productivity.
Judgement and controls stay with people
AI should not decide what a number means. It should not make the calls on stock valuation, provisions for short-dated or slow stock, write-offs, accruals or rebate estimates. It should not replace the reconciliations and sign-offs that tell you a figure is right. Nor should it run unsupervised on data nobody understands. These are the figures the auditors test, the bank reads and the directors sign. A tool cannot explain its reasoning to any of them as the person who made the call can. Two practical points follow. Keep a person's name against every figure that leaves the finance office. And do not paste supplier prices, customer margins or payroll into a free tool without knowing where that data goes.
Use it to take away the rebuilding, not the thinking.
Four questions to put to a report
List every report produced at month-end, and the hours each one takes. Then put the longest through four questions. A report that passes all four is a sensible first project.
Hours: does it take at least half a day every month?
Rules: can every step from raw export to finished report be written down, with no judgement calls along the way?
Checking: can a person confirm the output is right in a fraction of the time it takes to build?
Inputs: do the source exports arrive in the same format each month?
A gross margin by customer report, built from the same sales and cost exports each month, often passes. A stock valuation with judgement on slow lines does not. Start with one report. Run the old and new methods side by side for two closes, and only then switch over.
The sums are modest, but real. Say that one report takes a day and a half each month, and two-thirds of it is mechanical. Taking those steps away saves around twelve days a year (illustrative). That is time the finance team can spend on the questions that matter at month-end. Why did margin move on a customer? Why is a line overstocked? Why did the chiller costs go up? Keep a record of each run, the inputs used and what the person checking changed. Then the process can be audited, and explained to the bank or the auditors when they ask.
Clean master data pays before any AI does
AI tools work best on consistent data. Gartner's respondents named poor data quality and availability among their largest obstacles. If product codes, supplier names or cost centres are entered three different ways, that needs fixing first. In my experience, fixing it shortens month-end on its own, before any new tool arrives. Most of the hours go on making exports agree, not on reading them.
Where to begin
List every month-end report with the hours it takes and who builds it.
Put the three longest through the four-question screen and pick one.
Clean the product, supplier and cost-centre codes that report depends on.
Write down which figures need a named person's sign-off, whatever tool produces them.
Agree what data may and may not go into an AI tool before anyone starts using one.
Sources
Figures are taken from the sources listed and were correct at the time of writing. Where rules are described, check the latest official guidance and take qualified advice before acting.
Comments