The ERP, the spreadsheet and the project plan, agreed.

1Three plans, one factory.

The ERP holds the work orders: what to make, how many and by when. It rarely holds everything that decides when a run can actually happen, such as which tooling is free, which changeovers fit together and who is on shift. So the production planner sequences the real work in a spreadsheet. New products have their own plan, in a project tool, where validation runs reserve lines months ahead. In case studies of planning in manufacturing, spreadsheets still dominate over ERP and planning software as the planners’ main support.¹

Each plan is kept by different people, updated at different times, and right about something. Where they disagree, nobody sees it until a line is booked twice or a run starts without its material. Spreadsheets add errors of their own: in an audit of 50 working business spreadsheets, 43 had at least one error that produced a wrong result.² And the people who could catch it are busy relaying: in a study of seven schedulers over nineteen shifts, they spent more time passing on information than making scheduling decisions.³

2Read every plan as it is.

Neil reads each plan in its own format. From the ERP: work orders, routings, lines and their capacity. From the planner’s workbook: its tabs, colours, notes and hand-built totals. From the project plan, as exported from MS Project: tasks, resources and dates. Every run Neil reads keeps a link to the cell, row or task it came from.

Neil also reads what the plans assume. Material arrival comes from open orders and the supplier inbox, read the same morning by supplier order follow-up, and demand from the customer forecasts read by customer forecast intake.

3Match every run to one work order, line and week.

Each run is matched to the same work order, line and week across the plans: by work order number where there is one, and by part, quantity and week where there isn’t. A run that exists only in the project plan, such as a validation run, is kept as itself. A run that matches nothing anywhere is shown, not dropped.

4Check in code for what can’t be true.

Then code checks the plans against one another and against the factory. A line booked twice in the same week. A tool set on two lines at once. A run whose material arrives after it starts. A line booked past its hours. A run for demand that no longer exists. Every hand-built total is recomputed from its rows, so a formula that stops short is found.

A language model is not asked to build or check a schedule on its own. Asked to produce project plans directly, frontier models met every constraint 8.3% of the time on average in a 2026 preprint,⁴ and in an earlier benchmark GPT-4 completed 0.6% of realistic travel plans with many constraints.⁵ The plans are your planners’; Neil checks them in code and proposes how to resolve each conflict.

5Weeks 47 to 50 at Acme Medical.

Acme Medical is a fictional ISO 13485 maker of tubing and catheter components, with four extrusion lines and a braiding line. Its ERP holds the work orders, its production planner sequences them in Excel, and its new-product planner keeps validation runs in MS Project. On Thursday 5 November 2026, Neil’s 06:00 run found five conflicts.

The first two are one problem seen twice. Acme Vascular’s ramp needs WO-31877, the validation runs need Line 3 the same week, and the material for the run has moved to December in an email nobody answered.

6People decide, and update their own plans.

Neil proposes a reconciled schedule, with each conflict, the evidence on each side and what each resolution moves. The production planner, the new-product planner and the production manager decide. Then the owner of each plan updates it: the ERP, the spreadsheet and the project plan. Neil never edits any of them. Each conflict is kept with its evidence, the decision and who made it.

7Validate it on past weeks.

Take a quarter that has ended, with the ERP’s work orders, the planner’s spreadsheet and the project plan as they stood each Monday. Neil checks each week as if it were live. Compare what it found with what happened: the double bookings found on the floor, the runs moved at short notice, the overtime and the runs that waited for material.

Count what Neil found in time, what it missed and what it raised that didn’t matter, and have the planners judge the proposals as they would a colleague’s.

8Start with one quarter.

Bring one quarter of work order exports, the planner’s spreadsheets as saved each week, and the project plans from the same weeks. Historical files are enough; no live access to the ERP is needed to prove it.