Manufacturing Cost Estimating: How AI Helps Price a New Product

To estimate the manufacturing cost of a new product with AI, list every material and bought-in part, time each process step in minutes, add labour at a loaded rate, an overhead rate per labour hour, a scrap allowance and tooling spread over expected volume, then ask the AI to check the build for missing steps and show a sensitivity table.

The AI does not know your costs. What it does well is the structure: it asks what you have forgotten, lays the sum out so each line can be challenged, and re-runs it when an input moves. The numbers still come from your supplier quotes, your stopwatch and last year's accounts, and a costing that skips any of the three is a guess however tidy the table looks.

Follow me on Instagram@sagnikteaches

The product: a powder-coated steel plant stand, first run of 1,000

An illustrative case. A small fabrication workshop with four welders wants to add a product of its own: a welded steel plant stand, powder-coated, sold wholesale to garden retailers. The owner has a prototype built from offcuts and a hunch that it should sell for about $60. No one has costed it properly, because the prototype took a Saturday and nobody timed it.

Connect on LinkedInSagnik Bhattacharya

The question is whether $60 covers the real cost and leaves a margin, and if not, what price the product needs, and which costs are worth attacking first.

Subscribe on YouTube@sagnikteaches

Step one: turn the prototype into a bill of materials with AI

Describe the product to an assistant in plain words, and ask it to draft the bill of materials and the process route, then correct it against the prototype on the bench. This is the use where AI earns time, because it reminds you of the items nobody thinks of: packaging, fasteners, labels, the pallet share.

I make steel products in a small workshop. Draft a bill of materials and
a process route for this product, and list anything I might have
forgotten.

Product: welded plant stand, 3 legs and a ring top, mild steel tube,
powder-coated, ships flat-packed with 4 rubber feet and a printed
instruction card. Wholesale to garden retailers, first run 1,000.

Output two tables:
1. Bill of materials: item, quantity per unit, unit of measure, what I need
   to price it (supplier quote, weight, length).
2. Process route in order: operation, machine or station, whether I must
   time it, whether it is done in house or outside.
Then list assumptions and questions I should answer. Do not give prices.

A sensible reply names the tube, the sheet for the ring, the feet, fasteners, the powder coating (outside), cartons and the card. It also asks questions: Is the tube cut to length by saw or laser? Is the ring rolled in house? How are parts hung for coating? Does the outside coater charge by hanging position or by area? Each question is a cost you had not yet priced. Keep the questions and answer them before moving on, because a question the assistant raised is usually a cost line you were about to miss.

What you would fix in the reply: the first draft treated the powder coating as in-house and left out the carton. Correct both before you cost anything. Check any list like this against the prototype in your hands, since the assistant has never seen it.

Step two: price the materials from real quotes, not memory

Take the corrected list to your suppliers for quotes at the first-run quantity. For anything with a price break, note the break. The illustrative figures:

Give these to the assistant as inputs, with the source of each, so the calculation record shows where each number came from. A cost estimate nobody can trace is a guess with a decimal point.

Step three: time the labour, then ask AI to find the gaps

Build five to ten units and time every operation. Use minutes per unit, from the second and third unit made, not the first. Then split time into the productive part and the share lost to handling and waiting. For the stand:

Paste the route and the times into the assistant and ask it a different question: "What operations in this route have no time against them, and what common handling steps do products like this need that I have not listed?" It should notice that nothing covers moving frames to and from the outside coater, the incoming inspection of the coated parts, or the loading of cartons onto a pallet. Whether these deserve a time depends on your workshop, and your answer may be "two minutes a unit shared across the batch". Put a number on each rather than leaving a blank.

The labour rate must be the loaded rate: wages plus the employer costs and paid time off that come with them, divided by productive hours, not hours paid. An illustrative $32 an hour loaded is used below. Your payroll or accountant can give the real figure; the assistant cannot.

Step four: overhead, scrap and tooling, each on its own line

Three costs are easy to leave out and hard to defend if you add them as a vague percentage.

Overhead. Take last year's overhead (rent, power, insurance, indirect staff, software) and divide by the direct labour hours actually worked. In the example that is $18 per direct labour hour. At 28 minutes a unit, overhead is 28 ÷ 60 × $18 = $8.40. If your shop is machine-heavy, use a machine-hour rate instead, as in a machine shop's written rate card.

Scrap and rework. A first run of a new product wastes more than a mature one. Use 6 percent of material plus labour as the starting allowance, and test it at double. Here, (18.05 + 14.93) × 6% = $1.98.

Tooling. A welding jig that costs $600 and is expected to serve a first-year volume of 1,000 adds $0.60 a unit. Show it separately so you can see what happens at 500 or 2,000.

The result: unit cost, price and the answer to the owner's $60

The owner's $60 would have given a margin of 26.7 percent, ($60 − $43.96) ÷ $60, not the 35 percent wanted. It still makes money, so the product is viable; the question became whether the wholesale market would pay $67.63 or whether the cost had to come down. That is the question a costing exists to answer.

A note on the margin sum, because it is where assistants and people both slip: a 35 percent margin on price means dividing cost by 0.65, not multiplying by 1.35. The second gives $59.35 and a real margin of 25.9 percent. Tell the assistant which you mean, and check the result by working back.

Step five: a sensitivity table showing what really moves the price

Now ask the assistant to re-run the estimate with each input changed one at a time, and report cost, price at the same 35 percent margin, and the margin you would earn if you held the original price. Run it yourself in a spreadsheet to check, because the point is to understand the answer.

The table is the useful output. Labour minutes move the margin most (a six-minute slip costs nearly eight points), steel price and scrap come next, and tooling volume hardly matters at this scale. So the effort belongs in the welding time: a better jig, a fixed weld sequence, a second welder on the longest operation. Chasing a cheaper tooling supplier would be a waste.

The last row shows the combined bad case: a margin of 20 percent at the same price. Whether you can live with that is a business decision, but you now know the range before the first tube is cut.

The sensitivity prompt and what to check in the answer

A sensitivity run is tedious by hand and quick for an assistant, provided you hand it the whole model and ask for working. A prompt that keeps it honest:

Here is my unit cost model (lines, quantities, rates). Re-calculate the
unit cost and the price needed for a 35% margin on price (cost divided
by 0.65) for each of these changes, one at a time, then all together:
steel +15%, labour minutes 28 to 34, scrap 6% to 12%, volume 1,000 to
500, overhead rate $18 to $24. Show the arithmetic for each row as a
line, state the units, and flag any row where a number looks odd.
Do not round until the final column.

MODEL:
[paste the table of lines and rates]

When the answer comes back, check three things. First, that the base-case row matches your own total to the cent, since if it does not the model was misread. Second, that the "all together" row is not simply the sum of the single changes (costs that multiply, such as scrap applied to labour, compound slightly). Third, that the assistant did not quietly change an assumption, for instance by applying scrap to materials only. A model that gets one row wrong rarely gets the others right.

A short example of the second check going wrong: an assistant asked for the "all together" case once applied the labour increase to the scrap line but left the overhead rate on the old 28 minutes, so overhead did not rise with the extra six minutes. The total looked plausible and came out about $1.80 low. The giveaway was that the overhead line was identical in every row, when it should have grown whenever labour minutes grew. Any line that should move with an input and does not is the first place to look.

How the price moves with volume across three runs

A new product rarely stays at its first-run size. Showing the owner what the unit cost does at 500, 1,000 and 3,000 units helps decide how many to commit to. In the illustration, volume affects only tooling here, but at larger runs you would also expect better material prices and faster labour, so the table below adds an assumed 4 percent saving on materials at 3,000 and 3 minutes less labour from practice. These are guesses to be replaced by supplier quotes and run data, and the table should say so.

The jump from 1,000 to 3,000 comes mostly from the assumed labour saving, not the tooling. That reinforces the earlier finding: the thing to measure on the first run is how long the welding takes once the crew has done a few hundred.

Four more uses of the same method on other products

A wooden toy maker adding a new range. The big cost is sanding and finishing time, which is not obvious from a prototype made by a skilled hand. Time a second person doing it, and test the labour line at 1.5 times.

A small-batch food producer. Yield, not scrap, is the lever: the weight after cooking against the raw weight. Put yield on its own row and test it at two points lower. Add packaging and label printing as materials, as they are real costs.

A plastics moulder quoting a new part. Cycle time and mould cost dominate. The assistant is most useful in building the table across quantities (1,000, 5,000, 20,000) and showing where the mould repays itself.

A contract assembler. The materials are the customer's, so the cost is labour and overhead only. Test the plan at a 90 percent line efficiency, since waiting for parts is the usual loss.

Mistakes the AI will not catch for you

An assistant works with what it is given. These errors survive a polished-looking estimate:

• Prototype time used as production time. A skilled hand building one unit is slower in some steps and faster in others than a batch routine. Time unit 3 to 10, not unit 1.
• Forgotten costs after the factory gate. Freight to the customer, pallets, returns and breakage on a wholesale product. Add a distribution line even if it is an estimate.
• Selling prices confused with costs. A supplier's list price rather than the price you pay at your volume, or the other way round.
• Currency or unit slips. Weight in kilograms against a price per pound, millimetres against inches. Ask the assistant to state every unit in its working and check the conversions yourself.
• The estimate that never gets updated. After the first run you hold real data. Replace each guessed line with the actual one and note the difference. Over a few products you will see your own habits: perhaps labour is always 15 percent over.

An illustrative case of the last point: a workshop that costs a product, builds a first batch and finds labour 20 percent above the estimate has learned something general about its estimating. If it records that, the next estimate starts with a correction factor. If not, the lesson is learned again on every product.

Keeping the estimate current without extra work

Keep the costing in one spreadsheet with the inputs on one tab and the calculation on another, so that a changed steel price flows through. Ask the assistant to build the formulas, then check a few by hand. Keep a column for the source and date of each input. When a supplier changes a price, change one cell, and the sensitivity table follows. For help on structuring such a sheet and on formula prompts, see Google Sheets AI prompts.

The same workbook is the base for a quote. If a customer asks for 250 units rather than 1,000, you change the volume and tooling and you have a number in minutes. And when you set the selling price, cost is only one input; how AI can help a small shop set prices covers the market side.

One caution on cost accounting more broadly: an overhead rate and a loaded labour rate are accounting choices, and an accountant should confirm yours if the figures will feed tax, grants or customer open-book arrangements. For internal pricing, the method here is sound.

A short checklist before you quote the price

1. Is every item on the prototype in the bill of materials, including packaging and fasteners?
2. Does every operation in the route have a time, including handling and transport to outside processes?
3. Were times taken from units 3 to 10, not the first one built?
4. Is the labour rate loaded, and the overhead rate worked out from last year's real totals?
5. Is the margin formula on price, divided by one minus the margin, checked by working back?
6. Has the sensitivity table been run, and do you know the one input that moves the margin most?
7. Have you decided what to measure on the first run, so the estimate gets replaced by facts?

Run through it for your own product, and fix the gaps you find. If you want a second pair of eyes on the method, that is part of what an AI implementation consultation can cover; you can start at the contact page.

Questions makers ask when costing a first run

Can AI estimate manufacturing cost on its own?

It can draft a bill of materials, list the process steps and build the calculation, but it does not know your machine times, scrap or supplier prices. Give it your numbers, make it list every assumption, and check the result against one real run of the product before you commit to a price.

How should overhead be added to a unit cost?

Pick one basis and stick to it: a rate per direct labour hour or per machine hour, worked out from last year's total overhead divided by the hours you actually worked. Do not add a flat percentage you cannot explain. The rate also needs refreshing when your hours or costs move.

What scrap allowance should I use for a new product?

Start higher than for a mature product, because first runs waste more while people learn the job. Use your own scrap records for similar work as the base and test the price at double that figure. Replace the guess with the real number after the first batch.

Should tooling be in the unit cost?

Show it as its own line, spread over the volume you expect to sell in the first year, so you can see its effect and change the volume. If the tooling is paid by a customer, take it out and say so in the quote.

How accurate does an early estimate need to be?

Accurate enough to decide go or no-go and set a price floor. Many small makers treat an early estimate within roughly 10 to 15 percent as good enough for that, then tighten it with supplier quotes and a trial run. The sensitivity table tells you which inputs deserve that effort.

Further reads

• Check a Bill of Materials With AI: Template, Gaps and Revisions — Check a bill of materials for gaps before costing it.
• Capacity Planning Template: Can You Take That Order? Check With AI — Check you can build the volume you just priced.
• AI for Small Manufacturing Companies: Build a Weekly Production Schedule — Schedule the first run once the price is agreed.
• Why AI Saves Time but Not Money, and How to Capture the Gain — Why a faster estimate is not a profit until you act on it.
• Six AI Applications in Manufacturing That Need No New Machines — Six more AI uses that need no new equipment.
• How Much Does Estimating Software Cost? Price Bands With and Without AI — What estimating software costs, with and without AI.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: all figures in the worked example are illustrative workshop numbers, not vendor or market data.