Capacity Planning Template: Can You Take That Order? Check With AI

Compare the new order's hours on each machine, setups included, with the free hours left after existing orders, week by week. If the tightest machine has enough free time before the date the customer wants, you can accept; if not, quote the later date or a split delivery. AI does the sums, the planner checks them.

The error that sinks most "can we take it?" answers is checking the total hours in the workshop instead of the hours on one machine. A shop with 200 free hours in a week can still be unable to take a job that needs 75 hours on the one mill that is already full.

Follow me on Instagram@sagnikteaches

The four tables the AI needs before it can answer

An assistant can only reason about your shop from what you paste in. Four small tables are enough, and each one fits on a screen.

Connect on LinkedInSagnik Bhattacharya

1. Effective hours per machine per week. Not the hours on the door. Take shifts times days, then multiply by an availability factor for maintenance, breakdowns, changeovers between jobs and the operator being elsewhere. Start with 80-85% for machines you trust and 70% for old ones, then adjust from your own records.
2. Booked load per machine per week. Every confirmed order's run hours plus setup hours, placed in the week the machine will actually do the work, not the week the order is due.
3. The routing of the new item. Which machines it visits, in what order, with setup time per batch and run time per piece on each.
4. The promise-date rule. A few lines in plain words: how full a machine may get, how long the steps after the bottleneck take, and what happens when material is late.

If you do not yet have a routing for the item, the quoting side comes first; the weekly production schedule tutorial shows how to build the job list those hours sit in.

Subscribe on YouTube@sagnikteaches

A bracket order walked through, hour by hour

This is an illustration with invented figures. A small sheet-metal and machining firm has four work centres. A customer asks for 600 brackets in four weeks. The planner builds the table of effective weekly capacity first.

Next the booked load for the coming four weeks, taken from the order book (run time plus setups):

The free hours are the effective capacity times four, minus the four weeks of load. Mills: 64 x 4 = 256, minus 196 booked, leaves 60. The brackets' routing is then multiplied out: run minutes per piece times 600, plus one setup per machine.

Read it by looking at the Gap column, not the total. The mills are 14.5 hours short in four weeks and the lathe has half an hour to spare, so the lathe is a second bottleneck that a single breakdown would break. The total free hours across the shop (250) dwarf the 149.5 hours needed, which is exactly why the shop-wide view misleads.

Promise-date rules you can paste into the prompt

The answer depends on a rule, and the AI will invent one if you do not give it yours. Write it down as numbered lines, so the model quotes them back.

PROMISE-DATE RULES
1. Never plan a machine above 85% of its free hours for a new order.
   (Usable free hours = 0.85 x free hours.)
2. The order is done when the LAST machine in the routing has finished,
   not when the bottleneck finishes. Add 1 week for operations after the
   bottleneck, packing included.
3. Material must be in the building before the week the first
   operation starts. If the supplier lead time is longer, the start
   moves, and so does the date.
4. Overtime counts only if the shift is already agreed with the
   people who work it. One Saturday on one mill = 8 h x 0.8 = 6.4 effective hours.
5. A split delivery is allowed if the customer accepts it in writing.

Rule 1 does the quiet work. Machines at 100% have no room for a rush job from a regular customer, a scrapped batch or a late-arriving tool. Whatever percentage you choose, choose it once and keep it.

The prompt that runs the check, with a sample answer

Paste the three tables as plain text or upload the sheet, then add the rules and the request. A version that works:

You are helping a production planner at a small machining firm.
Below are: (1) effective weekly hours per machine, (2) booked load for
weeks 1-4, (3) the routing of a new order of 600 brackets with run time per
piece and setup per machine, (4) our promise-date rules.

Task:
- For each machine, work out the free hours in each week and cumulatively.
- Apply rule 1 (85% usable) and say in which week each machine has enough
  usable hours for the order.
- Say which machine is the bottleneck and the earliest delivery date under
  rules 1-5 if the order starts in week 1.
- Show every calculation line so I can check it. Do not round until the end.
- List anything you had to assume. If a figure is missing, ask me.

[paste tables and rules]

An illustrative reply, trimmed: "Mills: cumulative free hours are 6, 18, 36, 60. Usable at 85%: 5.1, 15.3, 30.6, 51.0. The order needs 74.5 hours, so the mills do not reach it by week 4. Week 5 adds 34 free hours (load 30 against 64), cumulative 94, usable 79.9, which covers it, so mill work finishes in week 5. Lathe: usable at 85% over four weeks is 27.2 hours against 31.5 needed, so the lathe also finishes in week 5. With the one-week downstream rule, earliest delivery is the end of week 6. Assumptions: week 5 loads of 30 mill hours and 18 lathe hours are already booked, so the lathe's cumulative free time by week 5 is 46 hours, usable 39.1."

Now do what the planner does with every reply: pick the bottleneck line and recompute it by hand. Here 0.85 x 60 = 51, and 51 is below 74.5, so the claim is right. If the model had said the mills fit, the one-line check would have caught it.

Rush-order scenario: the customer insists on four weeks

A refusal loses the order; a counter-offer often keeps it. Ask the AI for options and make each one cost something visible.

• Option A, split delivery. Send 200 pieces by the end of week 4 and the other 400 by the end of week 6. Rule 2 means the first 200 must be off the mills and the lathe by the end of week 3, a week before delivery. They need 24 + 2.5 = 26.5 mill hours (usable by week 3: 30.6) and 10 + 1.5 = 11.5 lathe hours (usable by week 3: 15.3). Both fit. Try 300 and the sum fails: 38.5 mill hours against 30.6, which is why the split is 200 and 400 rather than half and half.
• Option B, Saturdays on one mill. Three Saturdays add 3 x 6.4 = 19.2 effective hours, so four-week mill free time rises from 60 to 79.2, usable 67.3. Still below 74.5, so the full order does not make week 4 under the 85% rule. It makes week 5 comfortably.
• Option C, subcontract the lathe operation. Removing 31.5 hours from the lathe takes the second bottleneck out, but the mills remain short, so this alone does not rescue week 4.
• Option D, decline the rush and quote week 6. No extra cost, and the existing customers keep their dates.

The planner should state aloud which rule each option bends. Option B bends nothing but needs people's agreement. A request that works only by breaking rule 1 is the point where a manager decides, with the table in front of them.

Where AI gets this arithmetic wrong

These are the failures worth testing for before you trust the tool with a quote.

• Minutes against hours. A routing that lists 7 minutes a piece turns into 7 hours if the units are not stated. Label every column with its unit.
• Setups dropped or counted per piece. The model may multiply a 2.5-hour setup by 600. Name setups "once per batch" in the header.
• Cumulative against weekly. Asked "is week 3 enough?", the model sometimes compares the order with week 3 alone. Ask for cumulative free hours and read both columns.
• Parallel machines treated as one. Two mills are 64 effective hours only if the part can run on either. A part tied to the newer mill has half that capacity.
• Silent assumptions. Requiring a list of assumptions in every answer exposes things like "I assumed no holidays" or "I assumed material arrives Monday".

A related mistake is allowing the model to fill gaps. Say "if a routing time is missing, stop and ask". Otherwise it will produce a plausible number for a time nobody has measured.

A cheap fix when the booked-load table is out of date

The table is only as good as the last time someone updated it. Picture a planner who pastes a load table in which a 38-hour mill order still sits in week 2, although the customer moved it to week 4 a few days earlier. The assistant sums the table correctly, reports 38 free mill hours in week 4 that no longer exist, and the new order is promised against them. The fix is not better prompting; it is a column labelled "last confirmed on" next to every order, and a rule that any row older than seven days is flagged in the AI's answer.

Ask for that flag explicitly: "List every order whose confirmation date is more than seven days old and say how much load it adds." Ten seconds of reading catches the stale rows.

Building the table once in a spreadsheet

You can run all of this in the sheet with no AI at all, which is a good way to test the AI. Put orders in rows with columns for machine, week, run hours and setup hours. Then free hours per machine per week become one formula: effective capacity minus SUMIFS(hours, machine, A2, week, B1). A cumulative row adds across weeks.

The AI's best role in this setup is not the sums but the questions around them: reading the routing from a customer's drawing notes, turning "can you do 600 by the 28th?" into the request above, and writing the reply with options. If the model's numbers and the sheet's numbers differ, trust the sheet and find out why. If you want help with the broader scheduling side, the spreadsheet scheduling test shows where the model's own planning starts to fail.

Two more orders through the same table

One example proves the method works once. These two show it giving different answers, which is the point of using it.

A small repeat order that fits easily. A regular customer asks for 80 spacer sleeves in two weeks. The routing is lathe only: 0.08 hours a piece plus a 1-hour setup, so 7.4 hours. Lathe free hours in weeks 1 and 2 are 2 and 6, a cumulative 8, and 85% of that is 6.8. Under the 85% rule the order does not fit in two weeks, even though it looks tiny: the raw 8 free hours cover 7.4 with 36 minutes to spare, which one late tool or a scrapped first-off would wipe out. The AI's answer is "week 3" (cumulative usable 15.3 hours), and the planner agrees. A person looking at the order size alone would have said yes and found out on the second Friday.

A large order that is fine because it avoids the bottleneck. Another customer wants 1,000 powder-coat-ready plates that need only the press brake (0.02 hours a piece plus 1.5 hours of setup) and finishing (0.03 hours a piece plus 0.5 hours). That is 21.5 brake hours and 30.5 finishing hours. Press brake free time over four weeks is 42 (usable 35.7) and finishing 116 (usable 98.6). The order fits comfortably in four weeks despite being the largest job on the book by piece count. Piece count and hours are different things; the table keeps them apart.

What a day of overtime really costs against what you win

A quick sum helps when the answer is "only with overtime". Say the order earns a gross margin of $4,800 and overtime on one mill adds 6.4 effective hours a Saturday at a loaded labour cost of $46 an hour for the operator and $14 an hour for electricity, tooling wear and supervision. Three Saturdays cost 3 x 8 x ($46 + $14) = $1,440 if the operator works the full 8 hours, so 30% of the margin. If the same sum on a different job shows overtime eating 70% of the margin, you quote the longer lead time instead. Put the sum in the prompt as a rule: "recommend overtime only if its cost is under 35% of the order margin". All of these figures are invented for the illustration; use your own labour and machine cost per hour.

Before and after: the quote reply

Before the table, a typical answer from a busy owner is: "Should be fine, let me check with the shop and get back to you." Three days later the customer has gone elsewhere or has been given a date that nobody checked. After, the same message reads: "We can deliver 200 by the 24th and the other 400 by the 8th. If you can confirm by Friday, material goes on order that day." The second reply is faster to write because the AI drafts it from the decision, and it is safer because the date has been tested against the mills and the lathe.

Replying to the customer with a date you can defend

Once the planner agrees with the table, the AI can draft the message. Give it the decision, not the whole analysis:

Draft a short reply to a customer who asked for 600 brackets in 4 weeks.
Decision: we can deliver 200 by the end of week 4 and the remaining 400
by the end of week 6. Do not mention our other customers or our
capacity numbers. Offer the split in writing and ask them to confirm
by [date]. Tone: plain, helpful, no apologies beyond one sentence.

A usable reply would run to about eighty words: the two dates, the quantity for each, a request to confirm, and one line saying what would let you pull the second delivery forward (for example, if the customer supplies material). Read it for any promise not in your decision, because models like to add "we'll do our best to deliver sooner".

Keeping the table honest after the order is accepted

The check is only useful if the next one is better. After every job, enter the real run and setup hours next to the estimate. A bracket that took 0.15 hours a piece against an estimate of 0.12 means the mills were short by another 18 hours, and you want to find that in week 2, not week 5. A short daily note from the floor, summarised with a daily production report, gives you those actuals without a new system.

Then reduce the guesswork by routine. Keep a column showing estimate against actual per part family, and adjust the availability factor by machine twice a year. A mill that really delivers 72% needs 72% in the table, however good the nameplate looks. If a design change alters a routing, update the table before the next quote; the engineering change order tutorial covers keeping that in step.

Checking that the capacity check works

Test it on three orders you already know the outcome of: one that shipped on time, one that slipped, and one you turned down. Run the same prompt with the load table as it stood on the day. If the check would have flagged the late job, and would not have blocked the on-time one, you can use it. If it says yes to the order that slipped, find the machine that ran out of hours and see whether the table was wrong, the availability factor was too kind, or something outside the routing, like a rework loop, ate the time.

Measure three numbers for a quarter: the share of promised dates met, the number of orders declined on capacity grounds, and the average gap between estimated and actual hours on the bottleneck. A shop that meets 90% of promised dates and declines a few jobs is doing better than one that says yes to everything.

Keep the planner in the loop on three points the model cannot see: a key operator on leave, a tool that is on order, and a customer who always changes the quantity after the quote. Add those as lines in the assumptions list, and the next answer accounts for them. If you want to map this with someone against your own routings, that's what my AI implementation consultation covers.

Questions planners ask about the capacity check

Do I need an MRP system for this?

No. The check needs four lists: machine hours available per week, hours already booked, the routing for the new item (run time and setup per machine), and your promise-date rule. A spreadsheet holds all four. An MRP or ERP system helps once the lists go stale faster than you can update them by hand.

How often should the capacity table be refreshed?

Refresh the booked-load columns whenever an order is confirmed, moved or cancelled, and at least once a week on a fixed day. Refresh the available-hours columns monthly, or straight after a breakdown, a new machine, a holiday week or a change in shifts.

Which AI tool should I paste the table into?

Any current chat assistant that can read a pasted table or an uploaded spreadsheet will do the arithmetic. Use a business plan or turn off training on your chats, because order books and customer names are commercial information. The planner still checks the bottleneck figures by hand.

What if two orders compete for the same machine?

Run the check for both together. Add their hours to the same bottleneck and see which delivery dates survive. If both cannot hold, the decision is commercial (margin, customer, penalty clauses), so give the AI's table to the person who owns that call.

Can AI estimate run times for a part I have never made?

It can suggest a range from a drawing description, but treat that as a starting guess only. Use times from similar parts you have actually timed, or run one trial piece, before you promise a date based on it.

Further reads

• RMA Process for a Small Manufacturer: Returns, Repairs, Credits — Plan capacity for returns and rework too.
• Calibration Log Template: AI Tracks Due Dates for Gauges and Scales — Gauge calibration takes machines out of the week.
• Is Cin7 ForesightAI Forecasting Worth It for a Small Maker? — See whether forecasting software earns its price.
• Where Should a Small Manufacturer Start With AI? — Pick the first job worth giving to AI.
• AI Work Instructions From Photos and Short Videos of the Line — Keep setup times down with clear instructions.
• 8D Report With AI: Answer a Customer Complaint Step by Step — Handle the order that did slip.
• Print Production Scheduling Software: Where AI Actually Helps — Seven print software vendors read for AI claims, a capacity model for presses and finishing, a due-date sum, list prices and a one-week board trial.
• AI Systems for Manufacturing: MRP Add-On or Chatbot on Your Own Data? — Compare native MRP AI replenishment tools against desktop chatbots running over CSV exports, with BOM thresholds and shop-floor tests.
• Capture Tribal Knowledge With AI Before Your Best Machinist Retires — Interview prompts, voice-note transcription and a sign-off loop that turn a retiring machinist's know-how into usable work instructions.
• Factory Energy Audit With AI: Find Peak Loads and Idle Waste — A step-by-step factory energy audit from a half-hourly meter export: find the peak, the idle baseload, and what to hand an auditor.
• Furniture Lead Times: Set Realistic Made-to-Order Dates With AI — How a small furniture workshop works out an honest delivery range from its capacity and queue, and tells customers at the right moments.
• Machine Shop Quote Template: AI Turns RFQ Emails Into Draft Quotes — Extract part, quantity, material and finish from RFQ emails, price from your rate card, and let an estimator approve every number.
• Manufacturing Cost Estimating: How AI Helps Price a New Product — One new steel product costed end to end with AI help: materials, labour minutes, overhead, scrap, tooling, margin and what moves the price.
• Sawmill Software: Where AI Helps With Orders, Yield and Quotes — Three admin jobs for a small mill: order emails to order lines, yield per log from a tally sheet, and quote drafts that a person prices and sends.
• How to Use AI in a Machine Shop: Five Jobs Beyond Quoting — RFQ triage, work instructions and inspection paperwork costed in hours and dollars, plus scheduling notes and customer updates for a 12-person shop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the worked figures are illustrative; no vendor data was used. Check any AI tool's data-use settings on its own plan page.