AI in Steel Detailing: Gains for a Small Office, and Tasks to Avoid

A small steel detailing office can use AI for the writing around the drawings: RFIs, transmittal emails, drawing check lists and weekly status updates. It should never touch modelling, connection choices, dimension checks, bills of material or any approval and release step, because those need a detailer or checker who can answer for them.

The test is simple: AI can draft words, but it cannot be responsible for a number, a revision letter or a decision. A wrong purpose line on a transmittal, "for approval" against "released for fabrication", can send steel to be cut from drawings nobody has approved. The sections below sort the tasks, show four drafts with their faults, and end with a checklist the office can use.

Follow me on Instagram@sagnikteaches

Sorting the tasks: gains, grey zone and no-go

The usual workflow starts from the design drawings, raises RFIs when information is missing or unclear, produces shop and erection drawings, loops through checking and approval comments, and releases the drawings to fabrication. That list is the map. AI belongs in the writing around it and nowhere near the model or the signatures.

Connect on LinkedInSagnik Bhattacharya

Gains. These are words around the drawings, with the numbers and references coming from you.

Subscribe on YouTube@sagnikteaches

• Drafting an RFI from the detailer's note: the AI does wording and layout, the detailer owns every reference and number.
• The transmittal cover email: the courtesy text is safe, but the purpose and sheet list come from the issue log and stay locked.
• A project check list built from the spec and the office standard: drafting a list is not checking a drawing.
• The weekly status email: the numbers come from the tracker and the AI writes the sentences.

Grey zone. Useful, with a check built in.

• Turning approval comments into an action list: the summary can drop a condition, so keep the original reply beside it.
• Summarising the specification: ask for clause numbers and check each one, because addenda are easy to miss.

No-go. A named person answers for each of these.

• Modelling, member selection, connection design and clash decisions.
• Dimension or geometry checks on drawings, because AI misreads dimensions and symbols and states the result with confidence.
• Bills of material, weights, CNC and fabrication data, which come from the model and not from text.
• Approval, release for fabrication and sign-off.

An RFI from a detailer's note, and the three things it got wrong

RFIs are the best first job because they are short, frequent and easy to check against the detailer's own note. The illustrative setting is a six-person office detailing a three-storey frame. A detailer jots this down at the model (all references and dimensions here are made up):

"B14 L2 C/3-C/4. M-204 shows 150 sleeve thru web, about 1150 from col C3 centreline. S-201 shows no opening in B14. beam 450 deep. need engineer to say ok + any reinforcement, or move the sleeve. B14 and its 2 end plates on hold in model. need answer by 14th or L2 package misses release."

Turn the detailer's note into an RFI. Format: subject; drawing references exactly as written;
the question in one or two sentences; what is on hold; response needed by.
Do not add any drawing reference, revision letter, dimension or number that is not in the note.
Write [ ] for anything missing. State no assumptions unless the note states one.
Note: [paste]

A typical draft (illustrative):

Subject: Level 2 beam B14, grids C/3 to C/4: web opening shown on the mechanical drawing only.
References: S-201 (structural, level 2 framing); M-204 revision C (mechanical, level 2 services).
Question: M-204 shows a 150 mm sleeve through the web of B14, 1,150 mm from the centreline of the column at grid C/3. S-201 shows no web opening in B14, which is 450 mm deep. Please confirm whether the opening is acceptable, and if so its position and any reinforcement required, or advise that the sleeve is to move.
On hold: B14 and its two end plates. We have assumed the opening can be formed in the shop.
Response needed by: the 14th, to keep the level 2 package on programme.

It reads well, and it has three faults, all of the kind a detailer must catch. First, "revision C" for M-204 is not in the note. The AI supplied a plausible revision letter, and a question that points at the wrong revision can come back answered for a drawing nobody is using. Second, "about 1150" became "1,150 mm", which turns an estimate into a stated dimension. Third, "we have assumed the opening can be formed in the shop" is an assumption no one made, and it would sit in the record as if the office had committed to it. The detailer deleted the revision letter (the drawing register gives the real one), restored "about", and cut the assumption line.

The rule that makes this routine safe is not in the prompt, it is in the habit: every number, reference and letter in the draft is checked against the note and the drawing before sending, and the sender's name goes on it only after that. The prompt's "do not add" line cuts the faults but doesn't remove them. For the log that goes with it, ask for a row (number, date sent, recipient, subject, response due, items on hold) and a three-line chase for anything unanswered after your agreed number of days. The tutorial on AI for RFIs, logs and reports covers the project-management side of the same log.

The sum, illustrative: 40 RFIs a month, 25 minutes each by hand to find references, word the question and format it, and 10 with a draft plus the check, saves 600 minutes, or 10 hours a month. The saving that matters more is quality: an RFI that states the two drawings, the conflict, the options and the date gets a usable answer first time, where a vague one costs a second round.

Transmittals: the AI writes the courtesy, the log writes the purpose

A transmittal tells the recipient which sheets, at which revisions, are being issued, and for what purpose. The purpose is the dangerous part. "For approval" asks someone to review; "released for fabrication" says steel can be cut. Never let the AI choose it.

Keep the purpose and the sheet list in the issue log and build the first line there, for example with a spreadsheet formula: ="Transmittal "&A2&" - "&D2&" - "&B2, which joins the transmittal number (A2), the project name (B2) and the purpose (D2) into one line. Paste that line at the top of the email yourself, then give the AI only the polite text:

Write the body of a transmittal email under 90 words. Do not state the purpose, the sheet list or
any revision; I will paste those. Include: response requested within 10 working days; comments to be
returned marked up on the PDF; a reminder that the drawings are not for fabrication unless the
purpose line says so. Plain, polite, no exclamation marks.

The mistake this prevents is easy to picture (illustrative). Let the AI write the whole email from the log rows and a draft can open "Please find attached the drawings for fabrication" when the log said "for approval". It reads naturally, and the first line is the one a busy fabricator acts on. By locking the purpose into a formula and asking the AI to say what the drawings are not for, the wrong word has nowhere to enter. A tutorial on the design side, keeping a drawing issue register and transmittal log, shows the same log from the architect's chair and is worth reading for its columns.

Check lists: the AI drafts the list, the checker ticks the boxes

Every office has a standard check list. What it lacks is a version tuned to each project's specification, and that is a drafting job. Paste the office list and the drawing-related clauses of the spec, and ask for a project list where every line carries the clause it came from, and any line with no clause is labelled "office standard". That label stops the AI inventing requirements and shows the checker where the list is theirs rather than the client's.

A short extract, filled in (illustrative; clause numbers and tolerances are placeholders for your project):

Two points keep this honest. The "Checked by" column is a person, never the AI. And when the list is built, a human reads it against the spec once, because the AI can miss a clause that sits in an addendum or a different section, and a list that omits a requirement is worse than one that has none, since people stop looking for it. For readers who want the longer version of comparing drawings with a spec, checking fabrication drawings against the spec goes into what AI can and can't do there.

It is tempting to go one step further and paste a drawing PDF into the assistant, asking whether the dimensions are right. Don't. Models can describe what they think a drawing shows, and can be badly wrong about dimensions, symbols and small text. The tutorial on whether ChatGPT can read engineering drawings sets out what to protect and check, and the short version is that reading is not checking.

Weekly status: the numbers from the tracker, the sentences from the AI

A status email for the general contractor is a pure summary job. Keep a tracker with one row per sheet and a status column (not started, in modelling, in checking, submitted for approval, comments received, approved, released), and compute the counts there, not in the AI. Then paste the counts and the RFI list and ask for the email.

Write a weekly status email to the general contractor from the figures below. Use only these figures;
do not recalculate or round them. Order: progress in one sentence; open RFIs and what each is holding;
what we will issue next week; what we need from them. Under 150 words, plain, no adjectives.
Figures: [paste the counts and the RFI list]

With these illustrative counts for week 17: 120 sheets in scope; 64 approved, 22 submitted and awaiting comments, 18 in checking, 16 not started; 9 RFIs open, 4 of them older than 10 working days and holding 11 drawings. A usable draft: "Week 17: 64 of 120 sheets are approved, 22 are with you awaiting comments, 18 are in checking and 16 have not started. Nine RFIs are open. Four are over 10 working days old and hold 11 drawings between them; the oldest concerns the level 2 web opening and holds the level 2 package. Next week we will issue 14 sheets of level 3 framing for approval and resubmit six with comments incorporated. We need answers to RFIs 23, 27, 31 and 34 by Friday to keep the level 2 release date."

Check the sums before sending: 64 + 22 + 18 + 16 is 120, so the counts reconcile. The draft left out the percentage (64 of 120 is 53%) because the prompt said not to recalculate; add it yourself if the contractor likes one. What you'd fix is any invented reassurance. A draft will sometimes close with "we remain confident of the programme", which is a commitment nobody made, so the prompt bans adjectives and the check removes anything that sounds like a promise.

Turning approval comments into actions without losing a condition

Drawings come back from approval with comments, and turning a page of markups into a numbered list of changes is tedious. The AI can do the tidy-up if you give it the typed comments (not the drawings): one line per comment, sheet number, type of action (change, clarify, note), and who does it. For example, "S-12: add stiffener ring at B14 opening per engineer's sketch" becomes a change on sheet 12 for the detailer, with the sketch reference preserved.

The fault to watch for is the summary that drops a condition. Say the engineer replies to an RFI (illustrative): "Opening acceptable provided reinforcing plates are fitted as sketch SK-12 and the sleeve is not moved." A summary that reads "engineer confirms opening acceptable" is shorter, tidier and wrong, because it loses both conditions. A detailer working from that line models the opening without the plates, and the omission is found at checking if you are lucky and after fabrication if you are not. Two rules follow. Never replace a response with a summary: the full reply goes in the log, linked from the action list. And have the AI quote the reply verbatim in a column, so the condition sits next to the action.

Sample action list rows (illustrative):

Even with the quotes, the action list cannot see sketches or markups on the PDF. The detailer opens the marked-up sheet for each line. The tutorial on what is safe and what never is when a structural engineer uses ChatGPT is a useful look at the same boundary from the engineer's side.

The tasks that stay with a person, and why each one does

The sorting list gave verdicts. This is the reasoning, so the office can apply it to tasks that are not on the list.

• Modelling, member selection and connections. This is the detailer's craft and the point of the job. Where the design drawings don't give enough, a detailer can propose a connection, but the engineer approves it. An AI suggestion adds a step, not an answer.
• Dimension and geometry checks. A drawing is evidence, and an AI's reading of it is a guess. A wrong dimension on a shop drawing becomes a wrong piece in the yard.
• Bills of material, weights and fabrication data. They are exported from the model so that they match it; Trimble says model data from Tekla Structures, a structural BIM package, drives fabrication machinery, and the principle holds whichever package you use. Text produced any other way can disagree with the model without anyone noticing.
• Approval, release and sign-off. The checker signs for checking, the engineer or architect for approval, and a named person in your office for release. Each signature means a person looked.
• Interpreting design intent. If the notes are ambiguous, the answer is an RFI to the people who own the design, not the AI's best reading.
• Declaring compliance. "Complies with the spec" is a claim somebody must be able to defend. The AI may list the clauses; a person decides whether the drawing meets them.
• Anything confidential. Models, drawing images and client names are often covered by contract. Don't paste them into a consumer AI account without permission, and where you have a team plan, read its data terms. Claude's pricing page, for instance, lists no model training by default on its Team plan; check the page for whichever plan you use.

Three questions will place a task you have not seen before. Is the output words or a decision? If it is wrong, who is accountable, and is that person the one reading it? And how, and how soon, would a mistake be found? A weekly status email is words, the sender is accountable and the contractor will spot a wrong figure the same day, so it passes. A check mark on a drawing is a decision, the checker is accountable and a mistake surfaces at the fabrication shop, so it doesn't.

A pre-send checklist the whole office can use

Print this and keep it next to the screen. Anything drafted with AI passes all six before it leaves the office:

1. Does every drawing number, revision letter, grid reference and dimension match its source, the issue log or the drawing itself?
2. Is the purpose of the issue (for approval, for information, released for fabrication) copied from the log and not typed or drafted?
3. Is every decision, conclusion or commitment in the text one that a named person made?
4. Has a person read any reply from the engineer, architect or contractor in full, with its conditions intact, rather than a summary?
5. Is anything in the prompt or the draft confidential under the contract, such as drawing images, client names or addresses, that should not be there?
6. Is the sender a person who has read this version, and will they answer for it?

If the answer to any is no, fix it before sending. If you can't answer one, the task belongs on the keep-human list.

What a six-person office saves in a month, and what it costs

Put the four gains and the grey-zone job together, using illustrative volumes and times, and replace them with the figures from your own tracker after a month of noting them down.

At an illustrative $50 an hour, 23 hours is about $1,170 a month. Six seats of Claude Team at its listed $25 a seat, billed monthly, would be $150 (it lists $20 a seat billed annually), and a free or $20 individual plan costs less still, though a team plan is the more sensible home for anything under a client's confidentiality terms. Check the plan's own page before you buy, because prices and limits move.

The real return is rarely the hours. It is fewer second rounds on RFIs, fewer status calls and no wrong purpose line on a transmittal, so measure those: count the RFIs that needed a follow-up question, the transmittals corrected after sending and the status questions from the contractor. If those fall over a quarter, the routine is working. Keep the keep-human list on the wall, and when a new tool or a keen colleague proposes adding a task to the AI side, run it through the three questions first.

Questions detailing offices ask about AI

Can AI check shop drawings for errors?

Not reliably. It can draft a check list and compare pasted text, such as notes or a list of revision letters, but it misreads dimensions, symbols and scanned markups, and it will sound confident when it is wrong. Use it to prepare the list and let the checker tick every box against the drawing and the model.

Is it acceptable to paste client drawings into an AI tool?

Only if your contract and the client allow it. Most of the useful jobs here need text, not drawings: a detailer's note, an issue log row, a tracker table. Paste those, leave out client names where you can, and read the data terms on your plan before you start.

Does this replace the checker or the detailer's judgement?

No. It takes over drafting time for emails and logs. The detailer still decides what goes on the drawing, the checker still verifies it, and the engineer or architect still approves it. If a task needs someone to answer for a number, a revision letter or a decision, it belongs on the keep-human list.

Which job should a two-person office try first?

RFIs. They are short, frequent and easy to check against the detailer's own note, and a clear RFI saves a round of emails. Add weekly status updates next, once your tracker is tidy, then transmittal text and check lists. Leave anything near modelling or sign-off alone.

Further reads

• AI Software for Construction Drawings: Catch Changes Between Revisions — Catch what changed between drawing revisions.
• ChatGPT for Electrical Contractors: Bid Clarifications and RFIs — RFI and clarification wording from another trade.
• How to Quote Fabrication Jobs: Can AI Price One From a Drawing? — Whether AI can price a fabrication job from a drawing.
• How Can AI Help Construction Project Managers? Submittals and Sub Chasing — Submittal and chasing routines the other side of your drawings.
• AI Tools for Welding Shops: Quotes, Drawings and Job Scheduling — The fabricator's side: quotes, drawings and scheduling.
• Best AI for Welding: Can It Keep Weld Procedure Records in Order? — How a welding shop can use AI to match material certificates to jobs, flag missing documents and build a customer audit pack, with a qualified person signing.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Trimble's Tekla Structures product page and a public overview of the steel detailing workflow (design drawings, RFIs, approval, checking, release); claude.com/pricing (Team plan data terms). All office figures, logs and drafts are illustrative.