AI for Financial Reporting in Manufacturing
It is the fourth working day of the close. Three plants have submitted, one has submitted twice, and the last has sent a trial balance that does not agree to the stock ledger behind it. Group finance is not analysing anything yet. It is still working out whose numbers to believe, and gross margin will not settle until it does.
That first stretch is where a manufacturing close spends its time. Most of the balance sheet is inventory. Most of the movement in profit is explained by how cost was absorbed rather than by anything anyone did. And the numbers come from plant controllers who each know their own site and none of the others. Our position at Adopt AI Partners is unchanged: agents assemble, reconcile and draft with evidence attached; a qualified person approves and posts.
The close that runs in a different shape at every plant
Every site closes in its own dialect. One books accruals before the stock count, another after. One clears production orders weekly, another only at period end. The group timetable says the steps are identical; the working papers say otherwise. An agent runs the group checklist against each plant's actual ledger rather than against what the plant says it did: count booked before cut off, order variances cleared, stock ledger agreeing to the general ledger, the goods received accrual not stretched to hold a late invoice. Where a site departs from the pattern, it names the step that ran late and what that moved.
Standard cost and the roll that ages quietly
Standard costs are set once a year on an assumed volume, material price and labour rate. By mid year several are wrong. Nothing breaks; the variance simply grows, and a product line's margin reads better or worse than it is. An agent compares standard to recent actual cost and reports where the gap exceeds your tolerance, separating material price drift from a routing time that no longer matches how the part is made, or a scrap allowance set for a process since improved. Ahead of the roll it prepares the revaluation effect by item and plant.
Purchase price variance as a reported result
The term covers two jobs. In accounts payable it is an exception on one invoice to clear before payment. In reporting it is a line in the pack, read by people who take it as a verdict on the buying team. That reported number covers unrelated causes: price genuinely paid away from standard, a standard set months ago on a different assumption, currency, a freight or duty element the standard never anticipated. An agent decomposes it by supplier, commodity and plant, separates a buying outcome from an artefact of standard setting, and drafts the commentary with the entries cited.
Absorption, and the variance that volume creates
Labour and overhead are absorbed at rates built on a planned volume. Run below plan and cost sits unabsorbed in the profit and loss; run above and the plant over absorbs. Either way a variance appears that has nothing to do with efficiency, and someone has to explain why a site looks worse in a month when nothing on the shop floor changed. An agent separates the volume effect from rate and efficiency, reconciles absorbed cost to the hours and activity recorded, and names the source: a line that ran short, a shutdown, a rate that no longer reflects the cost base. Whether that is a signal to act on stays a conversation between plant and group.
Work in progress, scrap and the order left open
Work in progress is where valuation errors hide. An order stays open long after the goods shipped, holding cost that belongs in cost of sales; another closes early, leaving a variance that surfaces next period. Scrap, rework and yield losses hit the ledger with no story attached, so the explanation is reconstructed weeks later by someone who was not there. An agent reviews open orders against production and despatch history and flags those whose status does not match reality: shipped and still open, open across periods with no movement, closed with material still issued. On losses it links the posting to the order, the item and the shift record while the detail still exists in somebody's memory.
Counts, adjustments and the story that has to go with them
A count adjustment is easy to post and hard to explain. The number takes a moment; why it arose, and whether the perpetual records can still be relied on, takes far longer, and that is what the auditor asks. An agent prepares each physical and cycle count for review rather than the total alone, grouping differences by location, item class and value, testing whether a loss in one place matches a gain in another, usually a movement booked to the wrong location, and looking for transactions around the count date that explain timing rather than real loss. A person still posts, but the adjustment arrives with its explanation written.
Slow moving stock and the provision a policy cannot judge
Excess and obsolete provisioning is mechanical on paper and not at all in practice. The policy provides against stock that has not moved for a defined period. Such a part may be a service obligation held deliberately, a component for a programme restarting next quarter, or material genuinely dead that nobody wants to write off in a difficult month. An agent applies the policy exactly as written, then assembles what the policy cannot see: demand history, open orders, the service commitment behind the part, provisions released before and why. It presents the calculated figure beside the case for departing from it, marked as a case and not a conclusion.
Plant to plant transfers and the margin sitting in stock
When one plant makes for another the transfer carries margin, and until the goods leave the group that margin is not earned. The elimination depends on both sides posting in the same period, and one side almost always posts late. An agent reconciles transfers pair by pair while the window is open, matching despatches to receipts and separating what is in transit from what is missing. It calculates the margin held in closing stock from the transfer price and the cost behind it, and presents the elimination with the movements attached. Where the sides disagree it names the entries and the plant responsible, not a net figure somebody unpicks by hand.
The machine that runs before anyone capitalises it
A capital project on a production site has no clean completion date. The line is commissioned, runs trial batches and produces sellable output while it is still being adjusted. The spend sits in assets under construction, depreciation has not started, and the costs booked around the project include items belonging in repairs. The correction lands months later and moves both the balance sheet and the result. An agent tracks spend against scope and against the operational record, commissioning sign off and first saleable output, flags projects producing output that has not been capitalised, and assembles each case with invoices and scope documents.
The plant pack and the review it has to survive
The commentary is written last, by a controller who has been closing for days, so it restates the movement instead of explaining it. "Material usage unfavourable due to higher material usage" is a sentence that appears in real packs. An agent builds the pack from the reported ledger so tables and narrative cannot diverge, and drafts each explanation from the transactions that caused the movement: what changed, what caused it, what is timing rather than permanent. The controller adds the part no system holds, which is what operations intend to do about it.
Testing against closes you have already audited
A manufacturing close has a near perfect test set: periods already closed, reviewed and audited. Take one plant and one period you have signed, run the agent against the data as it stood, and compare its reconciliations, decomposition and commentary with what your team produced. Sort the disagreements three ways: where the agent was wrong, where it was right and your team was wrong, and where both are defensible and the difference is judgement. The first tells you what to constrain; the third tells you which decisions stay with a person for good. Run it across more than one plant, because the failure worth finding is an agent that performs where the data is tidy and degrades quietly where it is not.
Where AI fails here
Standard costing is full of conventions that exist only in your organisation's memory. Which overheads sit in the rate, how a joint process splits its cost, why one plant's routing carries an allowance another does not. An agent reading the ledger infers the convention from the data, and it will infer confidently and wrongly, producing analysis that is arithmetically sound and conceptually incorrect. That failure survives review longest, because it looks right.
Inventory data quality sets a hard ceiling. If receipts are booked late, if the bill of materials does not reflect how the part is really made, if scrap is recorded against the shift that noticed it rather than the shift that caused it, no analysis of those numbers is worth much. An agent surfaces the inconsistency faster, but it cannot repair the underlying record and should not be sold as though it can.
Provisioning and capitalisation are judgements rather than calculations, and a well written recommendation reads like an answer, so teams start accepting it. The protection is procedural: the agent presents a case and never a conclusion, and the person approving records their reasoning in their own words. Closes are also compressed, and compression is where controls slip. An agent producing hundreds of well evidenced flags on the fourth working day is not helping if nobody can read them, so ranking by likely effect on the reported result matters as much as detection.
The first thing to run at your next close
Start with the stock ledger to general ledger reconciliation across every plant. Almost every other number here inherits from it, and it can be tested in an afternoon against a period you have already reported. Add the open production order review next, as the cheapest way to find valuation sitting in the wrong period. Absorption and variance decomposition come after that.
None of this posts anything. The agent assembles, reconciles, drafts commentary with the ledger evidence attached and flags what does not tie. It sets no provision, signs nothing and approves nothing. If you want to talk about where that line sits in your own close, Adopt AI Partners can be reached at hello@adoptaipartners.com.