For a small welding or fabrication shop, a general assistant (Claude, ChatGPT or Gemini) can read a drawing PDF and list what the quote needs, draft cut lists and test a schedule, while a spreadsheet or shop-management tool holds the numbers. It never approves a weld procedure, and every quantity, weld length and date is checked by the shop.
Most search results for this topic are about welding automation: robots, vision systems, monitoring. That is a capital-equipment decision, and it is a different one. The office side of a shop, where quotes go out late and jobs collide on the floor, is where an assistant pays back first, and none of it touches the arc.
Reading a drawing for the quote: a sheet of what drives the price
A quote is a reading exercise before it is a pricing exercise. Prices go wrong when something on the drawing is missed, and the usual culprits sit in notes, not in the geometry. Ask the assistant for a quote-driver sheet: everything on the drawing that changes time, material, outside processing or risk. Then audit the list against the drawing.
You are helping me quote a fabrication job. Read the attached drawing and return a table of everything that affects price. Columns: item, what the drawing says (quote the note or callout), effect on the quote, and "confidence" (high, medium or low).
Cover: parts and quantities, material grade and thickness, every weld callout (type, size, length, intermittent or continuous, all-around, field), joint preparation, finish and coating, inspection and testing requirements, certificates and traceability, tolerances, revision letter, and any general notes.
Do not calculate prices. If anything is unclear, list it under "questions for the customer". If you cannot read a detail, say so instead of guessing.A typical result for a skid frame, trimmed (illustrative sample output):
| Item | What the drawing says | Effect on the quote | Audit result |
|---|---|---|---|
| Parts | 2 rails 2,400 mm; 3 cross members 800 mm; 4 braces 450 mm; 4 base plates 12 mm; 2 lifting lugs | Material and cut list | Matches the takeoff |
| Rail-to-cross-member joints | 6 mm fillet, both sides, continuous (assistant's reading) | Adds weld length | Wrong: the drawing shows an intermittent weld marked 50-150 |
| General note 4 | Seal weld all round on base plates | About 2.6 m more weld per skid | Correct, and not in the estimate |
| Note 7 | Surface preparation and two-coat paint system | Outside processing and a lead time | Confirm turnaround with the coater |
| Inspection | Visual inspection of all welds; weld map required | Documentation time | Correct |
| Certificates | Mill certificates for tube and plate | Traceability and storage | Correct |
| Revision | Revision C | Quote must cite it | Correct |
The audit column is the point. The assistant found a note the estimator missed and misread a symbol in the other direction. Neither would have been caught by trusting the list, and both are caught in ten minutes by someone holding the drawing. Run it as a second pair of eyes, not as the estimator.
A worked quote: ten pump skids, and the weld note nobody priced
Take the shop's first quote for ten of those skids, built from its own history. Each skid takes 142 kg of tube, angle and plate at a supplier price of $1.80 a kg, $14 of consumables and $38 of outsourced painting. Labour is 0.9 hours of cutting and preparation, 1.6 hours of fitting and tacking, 0.8 hours of grinding and cleaning, and welding at 36 minutes per metre for 7.2 m of weld. The shop rate is $85 an hour and the markup 18%. These figures, like every shop figure that follows, are illustrations; use your own recorded times.
| Line | First quote | After the audit |
|---|---|---|
| Weld length per skid | 7.2 m | 9.8 m (adds the 2.6 m seal weld) |
| Labour hours per skid | 7.62 | 9.18 |
| Labour at $85 an hour | $647.70 | $780.30 |
| Material, 142 kg at $1.80 | $255.60 | $255.60 |
| Consumables and outsourced painting | $52.00 | $52.00 |
| Cost per skid | $955.30 | $1,087.90 |
| Price per skid at 18% markup | $1,127.25 | $1,283.72 |
| Order of ten | $11,272.50 | $12,837.20 |
The seal weld adds 2.6 m at 36 minutes a metre, which is 93.6 minutes or 1.56 hours, worth $132.60 of labour per skid. Across ten skids with markup, the first quote was $1,564.70 too low, about 14% of its value. That is the case for the assistant's second look: one missed note on a repeat order costs more than a year of the subscription. The case against trusting it alone is the intermittent weld it read as continuous, which could have pushed the price up in the other direction and lost the job.
Two habits protect the figures. Take material prices from a current supplier quote, since an assistant does not know today's steel price and will quote an old one confidently. And let the spreadsheet do the sums: ask the assistant for the layout, then enter your own rates. For rate-sheet help, how to price welding jobs with a rate sheet covers building one, and whether AI can price a fabrication job from a drawing tests the harder claim.
Where 36 minutes a metre came from: weld time from your own jobs
The labour line in the skid quote leans on one figure: all-in minutes per metre of weld. It should come from your own job records, not from a table in a book and not from an assistant's guess. Pull the weld length and actual hours from past jobs of a similar kind and divide:
| Past job | Weld length (m) | Actual hours | Minutes per metre |
|---|---|---|---|
| J1 | 24.0 | 13.2 | 33.0 |
| J2 | 9.6 | 6.1 | 38.1 |
| J3 | 31.5 | 17.8 | 33.9 |
| J4 | 12.0 | 8.4 | 42.0 |
| J5 | 20.4 | 12.6 | 37.1 |
| All five | 97.5 | 58.1 | 35.8 |
The weighted figure is total hours times 60, divided by total metres: 58.1 times 60 over 97.5 is 35.8, which the quote rounds to 36. A plain average of the five minute-per-metre values gives 36.8, which treats a 9.6 m job as heavy as a 31.5 m one. An assistant can fall back on the plain average unless told otherwise, so ask for the weighted version or write the spreadsheet formula yourself. Once you have enough jobs, split the history by thickness, joint type and position, since out-of-position work in tight corners behaves differently from flat fillets. And use the figure only for work that resembles the history: a new material, a stainless job or thin gauge needs a trial or a margin of caution, not the shop average.
The email that goes with the quote: questions, assumptions, exclusions
The audit sheet produces a second deliverable, which is the list of things to settle before the price is firm. Ask the assistant to turn "questions for the customer" and your assumptions into a short email, with the wording kept factual. A sample draft for the skid job:
Hi [first name], thank you for the drawing (revision C). Our price for ten skids is attached, based on the points below. Questions: note 7 calls for a two-coat paint system, so please confirm the colour and the film thickness you need. The cross-member joints show an intermittent fillet marked 50-150; please confirm that is intended on all six joints. Assumptions: the price covers visual inspection and a weld map only, and any further testing is quoted separately; mill certificates are included for tube and plate. Validity: steel pricing is held for 7 days from the date above. Delivery is two weeks from order confirmation, subject to the painter's turnaround.
Check three things before it goes out. The revision letter must match the drawing. Each question must cite a note number or callout the customer can find. And the delivery wording must not promise a date the schedule has not tested, which is where the next section comes in. Keep internal notes, such as the fact that the first estimate missed the seal weld, out of the customer's copy.
Where assistants misread drawings, and a five-drawing test
Drawings stress an assistant in ways a letter does not. The usual failures are weld symbols, where the arrow side and other side swap or an intermittent weld becomes continuous; dimension strings, where 1,250 becomes 1,520; small print in title blocks and notes; and revision clouds that mark changes without saying what changed. Large sheets add a mechanical problem. OpenAI's help pages say images are resized before analysis, so small text on a full sheet can blur. Crop to the area you care about and ask about that crop alone, at the size you would read it on a screen.
Limits matter too. Claude's help page says it analyses text and visual elements in PDFs of up to 100 pages and reads text only from 101 to 1,000 pages. It takes up to 20 files per chat, each up to 500 MB, and images up to 8000 by 8000 pixels. A normal drawing set sits far inside those limits, so the practical risk is legibility, not size.
Test any tool on your own drawings before it touches a live quote. Pick five past jobs where you know the real answers, run the quote-driver prompt, and score each against the drawing:
| Check | How to score it | Pass mark |
|---|---|---|
| Weld callouts | Count callouts read wrongly (side, size, length, intermittent) | Zero misreads |
| Dimensions that drive cost | Compare lengths, thicknesses and quantities with the drawing | No error that changes material or hours |
| Notes | Count notes on the drawing the assistant did not list | Zero omissions |
| Invented content | Count items the assistant lists that are not on the drawing | Zero |
| Honesty about unreadable detail | Did it say "cannot read" when it could not? | Every time |
If the tool fails the weld-callout or omission checks on any of the five, keep it for summarising customer emails and drafting. Keep the drawing reading with a person. Repeat the test after any big update to the tool, because behaviour changes more often than version numbers do.
Cut lists: draft from the drawing, count in a sheet
The assistant is good at turning a parts list into a cut list: mark, quantity, section, length, end cut and grade. It is less good at nesting that list into stock bars, because the arithmetic needs the cut width and a rule for offcuts. For the ten skids, the cut list is this:
| Mark | Quantity per skid | Quantity for ten | Section | Length |
|---|---|---|---|---|
| R1 rails | 2 | 20 | Tube 100 x 50 x 5 | 2,400 mm |
| X1 cross members | 3 | 30 | Tube 100 x 50 x 5 | 800 mm |
| B1 braces | 4 | 40 | Angle 50 x 50 x 5 | 450 mm |
The tube pieces total 72,000 mm, which is exactly twelve 6,000 mm bars, and the braces total 18,000 mm, exactly three. A quick answer from an assistant may well say twelve and three. Reality is thirteen and four. Fifty tube pieces and forty braces each lose a 3 mm saw cut, which adds 150 mm and 120 mm, so neither total now fits into twelve or three bars. A sensible packing puts two rails and one cross member on each of ten bars (5,600 mm plus cuts), then seven cross members to a bar for the remaining twenty, which takes three more bars. On the tube that is one bar of waste, about 7.7% of what you buy, and the braces leave close to a full bar's length as offcuts, worth labelling for the next job's short pieces. The base plates and lugs come from sheet and are nested separately.
Ask the assistant to build the nesting as a spreadsheet and to state its assumptions about cut width, stock length and offcut reuse. Then compare its bar count with a hand count. The cut-list optimiser I checked, cutlistoptimizer.com, describes its stock as panels, with a cut-thickness option and a $4.90 payment for a three-day pass, and says nothing about bar or tube. For a bar-based shop, the spreadsheet is the safer route, and cutting-list software against AI compares the reliability of each approach for timber.
Scheduling: stress-test the plan before promising a date
A three-fabricator shop with 36 productive hours each has 108 hours a week, or 21.6 hours a day. Over two weeks that is 216 hours. Here are five jobs, with due dates counted in working days and one outside process. The skids carry their corrected hours, 9.18 a skid once the seal weld is in:
| Job | Hours | Due (working day) | Constraint |
|---|---|---|---|
| D: urgent repair | 6 | 2 | None |
| B: gate set | 22 | 6 | None |
| C: handrails | 30 | 9 | None |
| A: ten pump skids | 91.8 | 9 | Painting takes 3 working days, collected the day after fabrication ends |
| E: trailer frame | 40 | 10 | None |
The load is 189.8 hours against 216 of capacity, which is 88%, so it looks manageable. Had the schedule kept the first quote's 76.2 hours for the skids, it would have shown 81% and an easier fortnight than the shop really has. Now ask the assistant for a sequence:
Shop capacity is 21.6 productive hours a day. Here are five jobs with hours, due working day and constraints. Propose a sequence and the working day each job finishes fabrication, and say whether each due date is met.
Constraints: outsourced painting takes 3 working days and the work is collected the day after fabrication ends; delivery is the day painting returns. Do not assume overtime. If a due date cannot be met, say which and why.
[paste the table]A common first answer sorts by due date: D, B, C, A, E. Fabrication of the skids then finishes on day 7, the painter collects on day 8 and returns on day 11, which is two working days after the day-9 due date. The assistant may even say every date is met, because it counted the fabrication days and forgot the paint. The better sequence starts the skids straight after the repair: D, A, B, C, E. The skids finish on day 5, go to the painter on day 6 and come back on day 9, their due date. The gate set finishes on day 6, also its due date, the handrails on day 7 and the trailer frame on day 9. The rule underneath is to sequence by latest start, not by due date, whenever a job has an outside process. Notice too that two jobs now finish on the day they are due, so one late steel delivery breaks the plan, which is worth knowing before you promise the next customer a date.
Then use the assistant for what-ifs. Suppose one of the three fabricators is off for the fortnight. Capacity falls to 14.4 hours a day, 144 over the two weeks, and the load of 189.8 hours is 45.8 hours short. On the same sequence the skids would finish on day 7 and miss their paint slot. Overtime alone would mean three extra hours a day from both remaining fabricators on eight of the ten days. A more workable mix: three extra hours each for the first six days (36 hours) keeps the repair, the skids and the gate set on their dates and finishes the handrails on day 8, and the trailer frame's customer hears on day 1 that it moves into the following week. The assistant can lay out options like these in a minute; check its day-by-day sums the way you checked the bar count. You choose, because only you know who will agree to overtime. For the day-to-day tracking that follows, WIP tracking for a job shop gives a spreadsheet to run on.
Where AI stops: weld procedures, qualifications and acceptance
The line is clear and it does not move. Weld procedures, welder qualifications and weld acceptance are decisions for the people your quality system and the applicable code name for them. An assistant can tidy the paperwork around those decisions, and it must not make them.
| Area | An assistant may | An assistant must not |
|---|---|---|
| Welding procedures | Flag empty or inconsistent fields in a procedure record; reformat records; compare a drawing's weld list with your list of qualified procedures and flag any weld with no match | Approve or qualify a procedure, or choose process, filler, preheat or interpass limits |
| Welder qualification | Track expiry dates and draft renewal reminders | Decide that a welder, or a qualification, covers a job |
| Inspection | Draft checklists from your template; summarise records | Accept or reject a weld, or interpret test results |
| Drawing review | List callouts and ambiguities for the estimator | Be the final reading of a weld symbol |
The unmatched-weld flag is a good example of the right division. Paste the weld list from the drawing and your list of procedures, and ask which welds have no procedure. That is a comparison, the answer is checkable in seconds, and a missing match goes to the qualified person as a question rather than an answer. For the records side, keeping weld procedure records and material certificates in order goes deeper.
Tools and prices for a small shop
Prices below are what each vendor's own page showed on 29 October 2026, in USD. Vendor claims are the vendors' claims.
| Option | What it does | List price | Points to know |
|---|---|---|---|
| Claude, ChatGPT or Gemini | Reads drawing PDFs, drafts quote-driver sheets, cut lists and schedules | Claude Pro $17 a month billed annually ($20 monthly); ChatGPT Plus about $20 a month | You supply every rate and check every reading. Spreadsheets need code execution and file creation switched on in Claude. |
| Paperless Parts | Quoting software aimed at job shops and contract manufacturers, listing sheet metal fabrication among its processes | No price published; a demo is offered | Its home page names no AI features and links to an "AI Principles" page. Ask which steps use AI and watch it run on one of your own drawings. |
| Machestra | A drag-and-drop production scheduling board | Free for 3 machines and 3 users; Starter $190 a year (10 machines, 50 users); Growth $790 a year (100 machines, 100 users) | No per-user fees. It names no AI features, so the assistant does the what-if work. |
| Skody | Finite-capacity scheduling for high-mix CNC and fabrication | Shop plan from $1,500 a month, month to month | Aimed at shops running 20 to 200 machines. It brands itself Skody AI with an "AI Scheduler" but does not say what the AI does, and its results are customer-reported, with no independent validation on the page. |
The pattern is plain. A four-person shop can get most of its value from an assistant at about $20 a month and a sheet, while scheduling software starts to matter only once the whiteboard stops telling the truth. To see how a bigger toolkit fits around that, checking fabrication drawings against the specification covers the review step in more depth.
Before you upload a customer's drawing
Customer drawings are often covered by a non-disclosure agreement, and some carry a restricted-data marking that rules out consumer AI tools altogether. Read the agreement and the title block before the first upload. If the drawing is restricted, use a tool the customer has approved, or read it by hand. Business plans such as Claude Team and ChatGPT Business do not train on your content by default, which is a minimum, not a licence to upload anything. If you cannot tell whether a drawing is restricted, ask the customer before you upload it.
Keep a record for each quote: the drawing revision, the tool used, the prompt, the assistant's list and what you changed. When a customer disputes a price, or the drawing is revised, that record shows what was read and what was decided. It also makes the five-drawing test repeatable when the tool changes underneath you.
Questions from fabrication shop owners
Can ChatGPT or Claude read a weld symbol correctly?
Not reliably enough to price from. Symbols pack the arrow side, size, length, pitch, all-around and field marks into a few lines, and an assistant can swap sides or read an intermittent weld as continuous. Use it to list every callout it finds, then read each one against the drawing yourself before it reaches a quote.
Why does the AI cut list need fewer bars than I do?
It usually ignores the saw cut. Pieces that add up to exactly twelve bars need a thirteenth once each cut removes a few millimetres. Ask for the list as a table, then work out the bars in a spreadsheet that includes the cut width and a rule for offcuts, and compare the two answers.
Is a scheduling tool with AI worth it for a four-person shop?
Usually not yet. A whiteboard or a shared sheet plus an assistant for what-if tests covers a small shop. One finite-capacity scheduler aimed at shops with 20 to 200 machines starts at $1,500 a month. Buy software when you cannot see the floor load any more, and ask to see it on your own jobs first.
Can I upload a customer's drawing to an AI tool?
Only if the drawing's confidentiality terms and any restricted-data marking allow it. Check your non-disclosure agreement and the title block first. If the drawing is restricted, use a tool your customer has approved or do the reading by hand. A business plan that does not train on your content is a minimum, not permission.
Further reads
- Countertop Fabricator Software: Templating, Slabs and Where AI Helps — A vendor question list for templating, slab layout, quotes and cut files, with a scorecard, a yield sum and where ChatGPT helps with quote wording.
- AI-Generated Welder Website Marketing: Case Study Pages for Custom Work — Turn job photos and spec notes into case study pages that bring custom fabrication enquiries, with claim checks, consent wording and a better quote form.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Machestra, Skody, Paperless Parts and cutlistoptimizer.com vendor pages; Anthropic Claude pricing and file-upload help pages; OpenAI image-input help page; shop figures are illustrations (vendor pages read 29 October 2026).