Can AI Schedule Production for a Small Maker? A Spreadsheet Walkthrough

Yes, for a small workshop with a few machines, and with checking. Given a sheet of orders, machine times, setup times and due dates, AI can produce a workable schedule in minutes, sometimes better than a planner's first pass. It will not reliably find the best one, so compare it with your own plan before using it.

The test below is one you can reproduce on paper before trying it on your own sheet. Seven jobs go through one coating line, where changing colour costs half an hour. A due-date-first plan, the plan a changeover-only instruction leads to and the plan a properly constrained instruction leads to are compared on lateness and setup hours. Every figure is worked out by hand, not taken from a chatbot run.

Follow me on Instagram@sagnikteaches

The test: seven jobs, one coating line, three plans

Take a six-person workshop that makes steel shelving and coats it in-house. Every job passes the coating line, which is the bottleneck. When the line moves from one colour to another, an operator spends 30 minutes cleaning and changing over; same colour after same colour needs no setup. Times below are working hours counted from 08:00 Monday, so 8 means the end of Monday and 16 the end of Tuesday. All figures are invented for illustration.

Connect on LinkedInSagnik Bhattacharya

Total run time is 16.0 hours whatever the order, so the only things a plan can change are the number of colour changes and which jobs finish late.

Subscribe on YouTube@sagnikteaches

Plan A: earliest due date first. This is what most planners do by instinct: do what is due soonest. The order is J1, J7, J2, J3, J5, J4, J6. It needs four colour changes (Black to White, White to Black, Black to White, White to Grey), so 2.0 hours of setup, and finishes at hour 18.0. Nothing is late. It is a respectable plan.

Plan B: the naive AI plan. Ask a chat tool simply to "minimise changeovers" and the plan that best answers that instruction groups by colour: Black, then White, then Grey. J1 and J3 run first (finishing at 5.5), a half-hour change, then J2, J5 and J7 (the rush job ends at hour 10.5), another change, then J4 and J6. Two changeovers, finishing at 17.0. It saves an hour against Plan A, but J7 is due at hour 8 and finishes at 10.5, so it is 2.5 hours late. An objective that only counts setups rewards exactly this plan, and a tool following it can describe it as better.

Plan C: AI with the constraint spelled out. Tell it that no job may finish after its due time and that late jobs count as failures. The plan that meets both rules is J1, J3, then (after a change) J7, J2, J5, then (after another change) J4, J6. J7 now runs 6.0 to 7.0, inside its deadline. Two changeovers, 1.0 hour of setup, everything on time, finish at hour 17.0.

Two lessons come out of one tiny sheet. First, the gain from a good plan here is one hour on a 17-hour run, about 6 per cent, which is real but not dramatic. Do not expect AI to halve your lead times from sequencing alone; most of the gain in a real workshop comes from catching clashes and re-planning quickly. Second, the biggest risk is not a bad plan but a plausible one that quietly breaks a promise (Plan B), which is why the checking section below matters more than the prompt.

Building the sheet the AI will read

Four small tables are enough for most workshops. Put each on its own tab, or paste them as separate blocks in the chat, with plain headings.

Orders

One row per job: job code (not a customer name), product, quantity, which machines it needs and in what order, run time per machine, due date and time, and priority. Run time should be the total for the quantity, not per piece, so the AI does not have to multiply. For a job that passes several machines, write each step on its own row with a "step" number, and a column saying "can start only after step N finishes".

Machines and calendar

Each machine, the shifts it runs (for instance Monday to Friday, 08:00 to 16:30 with a 30-minute break), planned maintenance windows and any operator limits ("the press brake needs the one certified operator, who is off Wednesday"). If you do not give the AI the calendar, it will assume the line runs around the clock, and a Friday afternoon job will be scheduled for Saturday night.

Setup matrix

The table that makes or breaks the plan. List every changeover and its time. For the coating line above it is simple: any colour change costs 30 minutes. In other shops the matrix is uneven, and the AI needs it spelled out:

An uneven matrix like this, where going from dark to light costs more than light to dark, is the situation where a human planner's rule of thumb gets weak and a sheet-based solver earns its keep. Ask your operators for these times rather than guessing; they are often different from what the office thinks.

Material and tooling availability

A job cannot be scheduled before its material arrives. Add an "earliest start" column, filled from your purchasing records or supplier confirmations. This is the column most often left blank and then blamed on the AI.

If your data is not yet tidy, cleaning a messy sheet with AI is a sensible first step: the colours must be spelled identically, dates must all be real dates, and run times must be numbers rather than text such as "approx 2h".

A prompt that asks for a schedule and its proof

The prompt below states the objective in order of importance, so the AI cannot trade a missed deadline for fewer setups. It also asks for a table you can check.

You are helping me schedule a small workshop. Use only the data below; do not invent jobs, machines or times.

Objective, in this order:
1. No job finishes after its due time. If that is impossible, say which jobs are late and by how much.
2. Minimise total setup time using the setup matrix.
3. Then minimise the finish time of the last job.

Rules:
- One job at a time per machine. Setups happen between jobs of different colour and take the time in the matrix.
- Machines work only inside the shifts in the calendar. A job may not start before its earliest start.
- Times are working hours from 08:00 Monday.

Output a table: job, machine, start, end, setup before it (minutes), due, slack (due minus end). Then give totals: number of setups, setup hours, finish time, number of late jobs. Then list the assumptions you made.

[paste orders, machines and calendar, setup matrix, earliest starts]

Ask for slack on every row. A slack of zero or less is a flag, and a slack that is large on every job suggests the plan has room to spare and the due dates are not the real constraint.

What a typical reply looks like, and what to correct

Here is an invented reply to the seven-job sheet, written the way a chat tool might report it:

Schedule (optimal):
J1 0.0-3.0 | J3 3.0-5.5 | setup 0.5 | J2 6.0-8.0 | J5 8.0-9.5 | J7 9.5-10.5 | setup 0.5 | J4 11.0-15.0 | J6 15.0-17.0
Total setups: 2 (1.0 h). Finish: 17.0. All jobs delivered on time.

This is Plan B described as Plan C. J7 ends at 10.5 against a due time of 8, so "all on time" is false, and "optimal" is an unsupported claim. The fix is mechanical: never read the summary, read the slack column. If any slack is negative, tell the AI "J7 finishes 2.5 hours after it is due; fix this and rerun," and check the slack column again on the new version. If it keeps getting it wrong, ask it to write the check as a spreadsheet formula instead, which brings us to the next section.

Checking the schedule with three formula columns

Do not trust a schedule you cannot check in under five minutes. Paste the AI's table into a sheet and add three columns next to it:

1. Late by: =MAX(0, end - due). The sum should be zero for a plan with no late jobs.
2. Overlap check: for jobs on the same machine sorted by start time, =start_this - end_previous - setup_this should be zero or more. A negative number means two jobs are on the machine at once, or a setup was skipped.
3. Run time check: =end - start should equal the run time you gave. A difference means the AI shortened or stretched a job.

Take Plan B's J7 as the test. Late by is MAX(0, 10.5 - 8) = 2.5, so the first column catches the problem on its own. The AI can write these formulas for you if you describe the columns, and the Excel route is covered in the guide to using ChatGPT with Excel. Between them, the three columns catch the errors that matter most in a chat-built schedule: late jobs, two jobs on one machine at once and jobs quietly shortened.

How your manual plan fares in a fair comparison

Do not run the AI on a hypothetical week. Take last week's actual orders, build the plan you would have made on Monday morning using only what you knew then, and run the AI on the same data. Score both on four numbers:

• Jobs late and total hours late.
• Total setup hours.
• Finish time of the last job.
• Minutes the planner spent making it.

Say your planner needs 40 minutes to plan a week of 25 jobs and the AI route takes 10 minutes plus 5 minutes of checking. That is 25 minutes saved a week, about 22 hours a year, which is worth having but not a reason to restructure the business. If the AI plan also cuts setups by an hour a week at an internal cost of, say, $45 an hour for operator and line time, that is another $45 a week, around $2,300 a year. These are made-up numbers; use your own rates. If the comparison is a draw, keep your planner's plan and use the AI as a second pair of eyes that spots clashes.

Where chat-built schedules go wrong in real shops

Five faults to test for once the job list gets longer than a demo:

• Scale. Once the list runs to several dozen jobs on several machines, arithmetic done inside a chat becomes hard to trust and slow to check. Ask it instead to write a short Python script (using an open-source solver library such as Google OR-Tools, whose documentation includes a worked job-shop scheduling example in Python) or a spreadsheet model, and run that. The script does the sums; the AI only writes it. Have someone who can read code look at it before it drives a real plan.
• Silent dropped jobs. A list of 28 jobs comes back as 27. Add a count to the prompt: "Confirm the number of jobs in your table equals the number in my sheet."
• Calendar blindness. Jobs spill across lunch breaks, weekends and the maintenance window. Cure: put the calendar in the sheet and add a rule that no job may span a break unless it is marked "can be split".
• Frozen work. A job already half-run on the machine gets moved. Add a "frozen" column (yes for jobs that have started) and tell the AI frozen jobs cannot move.
• Stale data. The plan is built from a Monday export, but a customer changed a due date on Tuesday. Re-run when the sheet changes, and put a date on every export.

A rush order arrives at 10:30 on Tuesday

This is where AI scheduling earns its place. Say a customer needs 40 steel racks coated in white by Wednesday afternoon, and you are already committed through Thursday. With a paper plan the planner spends half an hour moving things around and hoping. With the sheet, add the rush job (run time 3.0 hours, due hour 24), mark everything already running as frozen, and rerun. Ask for two outputs: the new schedule, and a list of every other job whose finish time moved, with the new slack. In a couple of minutes you have an answer to give the customer, and a list of the other customers you now need to warn.

A machine goes down

The press brake fails on Wednesday morning and the engineer says it will be back on Friday. Add a downtime row (machine, start, end) and rerun. Ask specifically: "Which jobs now miss their due time, and what is the smallest set of changes that fixes it?" The answer might be "move J9 to the second shift" or "outsource J12's bending". The AI is useful here because the question is a search over many small moves, which people tend to do badly under pressure.

Overtime, shifts and people

Small makers often schedule people as well as machines: the one welder who is certified for stainless, the van driver who also loads. Add a "skills" table (operator, tasks they can do, hours available) and a "needs skill" column on the jobs. Keep it to one or two skills at first; the more skill constraints you add, the likelier the plan contains assignments no human would accept, such as the same operator on two machines in the same hour. The overlap check handles machines; add a second overlap check on the operator column to catch this.

For rota-style problems, where the question is who works which shift rather than which job runs next, AI shift scheduling for small teams is a closer match than this tutorial.

The point where software stops being optional

A sheet plus a chat assistant is a good tool for a small shop with one or two bottlenecks and a handful of changes a day. Buy a proper scheduling or manufacturing-planning tool when several of these are true:

• You plan more than three machines or work centres, and operators move between them.
• Orders change more than once a day, and re-keying them into a sheet is the slow part.
• Material arrival times drive start dates, and a stock system or purchase orders already hold that data.
• Customers or sales staff need to see live promise dates without asking the planner.
• You need a record of what was planned versus what happened, for audits or for quoting.

Software of this kind usually sits in one of two places: a planning module inside an MRP or ERP system, or a stand-alone finite-capacity scheduler. The comparison of MRP add-ons and a chatbot on your own data lays out how to choose, and it is worth reading before you ask for quotes. When you do, give each vendor the same seven-job style test from your own shop and ask them to schedule it live. A tool that cannot beat your sheet on your real jobs is not worth a monthly fee. Prices vary widely and change; ask for a written quote and a trial period rather than relying on a figure from an article.

A one-week trial you can start on Monday

1. Pick the bottleneck machine. Do not try to schedule the whole shop.
2. Build the four tabs (orders, calendar, setup matrix, earliest starts) for next week's jobs. Budget two hours the first time.
3. Write your own plan first and save it, with its late jobs and setup hours worked out.
4. Run the prompt above, then the three check columns on its output.
5. Compare both plans. Pick the better one, and tell the AI what it missed ("the sander needs a clean-down on Fridays").
6. At the end of the week, record what actually happened against the plan. The difference between planned and actual run times is the most valuable data you will gather, because it makes next week's sheet more honest.

If after a month the AI plan has not beaten yours on late jobs or setup hours, that is a useful result too: you know the planning is not your bottleneck, and the time is better spent on quality or on the supply side. AI quality control for small manufacturers is a sensible next place to look. If you would like help mapping this to your own shop, my AI implementation consultation covers that, and a 1:1 call can be booked here.

Questions makers ask before trusting an AI schedule

How many jobs can ChatGPT schedule in one go?

There is no published limit, and arithmetic done inside a chat gets harder to trust as the list grows. Keep chat-built schedules to a list you can check with the formula columns, such as one week on one bottleneck. Past that, ask it to write a small Python or spreadsheet-formula solution and run that instead, so the sums are done by code rather than by the language model. Always check the output against a formula column.

Does the AI find the best possible schedule?

Not reliably. It finds a plausible one and often says it is optimal when it is not. Treat the result as a candidate to compare with your own plan on lateness and setup hours. If it beats yours on both, use it; if it ties, keep yours.

What data should I keep out of the chat?

Customer names and prices are not needed to build a schedule. Replace them with job codes. Keep drawings, formulas and any confidential customer specification out of a chat tool unless your plan's data terms and your customer agreements allow it.

Can it handle machine breakdowns?

Only if you tell it. Add a downtime table with machine, start and end, and rerun. The AI does not know a press has failed this morning unless the sheet says so. Re-planning in minutes after a breakdown is where this approach pays back fastest.

When is spreadsheet scheduling no longer enough?

When several machines share operators, materials arrive late and change the plan daily, or you re-plan more than once a day. At that point a dedicated scheduling or ERP module that reads live orders will save more than it costs. Ask for a trial on your own jobs before you commit.

Further reads

• AI Job Scheduling and Dispatch for Trades and Field Service Teams — See dispatch-style scheduling for crews rather than machines.
• Where Should a Small Manufacturer Start With AI? — See where scheduling sits among a maker's other AI options.
• How to Use AI to Set Reorder Points and Prevent Stockouts — Set reorder points so material is there when a job starts.
• Print Production Scheduling Software: Where AI Actually Helps — Another sector's take on scheduling software versus AI help.
• Bakery Production Planning: Oven Slots and a Prep Timeline With AI — A smaller version of the same sequencing problem.
• AI Inventory Forecasting for Small Businesses: How It Works — Forecast material needs so the schedule has stock to run on.
• Landed Cost Calculator in Excel: Test Tariff Changes With AI — Cost out imported components that feed your orders.
• Automation in Clothing Manufacturing: What a Small Workshop Can Automate — Order intake, tech pack checks, fabric and cutting-plan sums, production tracking and customer updates for a cut-and-sew workshop, with tool costs and limits.
• AI for Industrial Manufacturing: Read Downtime Logs to Find Repeat Stops — A practical CSV walkthrough for finding repeated line stops, checking the arithmetic and choosing a maintenance action.
• Capacity Planning Template: Can You Take That Order? Check With AI — Compare a new order's machine hours, setups and existing load with real capacity, then set a promise date you can keep.
• Can ChatGPT Write G-Code? What a Machine Shop Must Check First — ChatGPT can draft simple G-code. Here is a faulty draft taken apart, a ten-program test log, what simulators miss and the sign-off before the spindle turns.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the worked schedules are hand-calculated illustrations, not measured results; Google OR-Tools documentation (overview and job-shop scheduling example), checked 5 November 2026.