AI Drawing Review: Check Architectural Sets Before You Issue Them

Partly. AI is dependable at text checks: whether the drawing index matches the title blocks, whether notes and cross-references agree, whether schedules match plan tags. It is unreliable at reading dimensions, line types and geometry from images. Use it to produce a flagged list, then have a named person check every flag and sign the set off before issue.

That split decides the method. Cross-reference errors, such as a note pointing to a detail that does not exist or a revision letter that differs between a sheet and the register, live in text, so they can be checked without asking a vision model to look at a single drawing. Dimensions and graphic conventions are where a person still needs to hold the pen.

Follow me on Instagram@sagnikteaches

What suits text checks and what needs a person

Sort every check on your pre-issue list by where its evidence lives. If the evidence is text (a sheet number, a note, a schedule row) AI can compare it. If the evidence is geometry or line work, AI can at best point at where to look.

Connect on LinkedInSagnik Bhattacharya

The practice that follows from the table is simple: run the text checks first, because they are cheap and their flags are usually real, then spend the person-hours you saved on the geometry. The tutorial on AI software for architectural drawings looks at products that package some of these checks; this one is about the review itself and who signs it.

Subscribe on YouTube@sagnikteaches

What the vendors' own documents say vision models miss

Both OpenAI and Anthropic publish limitations for image understanding, and the lists read like a catalogue of how drawings are built. This is what each page says, set against what it means on a sheet.

The resizing line deserves a sum. An A1 sheet is 841 mm wide. Anthropic's vision guide lists a longest edge of 1,568 pixels for its standard tier and 2,576 pixels for its high-resolution tier, which covers its newer models (check which tier yours falls in). If you send a whole sheet as one image, even the larger figure gives about 3 pixels per millimetre, so 2.5 mm annotation text is under 8 pixels tall; on the standard tier it is under 5. A dimension reading 3600 at that size is exactly where a 6 turns into an 8. As a working rule, crop until the smallest text you care about is roughly 15 pixels tall after resizing, then test that on your own sheets, because the right figure depends on the tool.

Preparing the set so the tool can read it

Most of the reliability you can buy is bought here, before any prompt is written.

1. Export vector PDFs from the drawing software, not scans, screenshots or print-to-image files. Vector PDFs keep a text layer, which is what the text checks depend on. A consultant's scanned PDF will check far less well; say so on the checklist rather than pretending otherwise.
2. Export the lists as spreadsheets. The drawing register is usually one already. Schedules can be exported too; Revit's File, Export, Reports, Schedule command writes delimited text that a spreadsheet opens, and most other BIM tools have an equivalent. Two exported lists compared in a spreadsheet are more reliable than any model reading them off a picture.
3. Split the set. Keep each PDF under 100 pages so a tool that analyses visuals up to that limit still does so, and group by discipline so flags are easier to read.
4. Keep last issue's PDFs beside this issue's. An overlay or compare tool is the right instrument for "what changed", and Bluebeam's Core plan and above list overlay and compare for drawings. The tutorial on catching changes between revisions covers that side.
5. Scrub what the tool does not need. Replace the client's name and address with the project code, and leave out security layouts and other sensitive sheets entirely. The confidentiality section below gives the full list.

Check one: the drawing index against every title block

A four-person practice is about to issue a 38-sheet set for a rear extension and loft conversion. The register is a spreadsheet, and the sheets come from the drawing software as one PDF each. Every figure below is invented to show the method.

The cleanest version of this check uses no AI for the comparison itself. Export the sheet list from the drawing software (number, title, revision, date) and compare it with the register using a spreadsheet lookup. AI earns its place in two supporting jobs: writing the lookup formula if you do not have one, and explaining the differences in plain words. When the only source is a pile of PDFs from consultants, extract the title-block text from each page and paste it in with the file name.

Below are two lists of drawing sheets: REGISTER (the practice's drawing register)
and TITLE BLOCKS (text read from each PDF's title block, with the file name).
Compare them. Ignore differences in capitals, punctuation and a trailing
"proposed" or "existing" in a title.
Report only:
1. Sheets in REGISTER with no title block.
2. Title blocks with no REGISTER entry.
3. Matching sheet numbers whose titles differ in meaning.
4. Matching sheet numbers whose revision letters differ.
5. Matching sheet numbers whose dates differ.
For each, give the sheet number, both values and the file name.
Do not correct anything.

On the 38-sheet set the comparison returns this (illustrative):

On its first run, before the tolerance rule was added to the prompt, the model also flagged A-102 because the register said "Ground floor plan" and the sheet said "Ground floor plan - proposed". That is noise, and the rule about trailing words removed it. Noise like this decides whether a checker trusts the list on the second set, so tune the prompt on the first.

A mismatch is a symptom rather than the error. The cause is usually that the register was not updated after the last revision, or that a sheet was plotted from the wrong view, and finding which means opening the sheet. In this example the comparison takes about five minutes and the resolution about twenty.

Check two: notes, references and schedules that disagree

Notes are text, so this is where AI is most useful. Export or copy every general note and keynote with its sheet and note number, then ask for groups.

Below is every general note and keynote from the architectural and structural
sheets, each with sheet number and note number.
Group the notes that describe the same element. Use only these elements:
ground floor slab, external wall, party wall, roof, lintels, fire stopping,
cavity barriers, windows, doors.
For each group, list sheet, note number and exact wording.
Then flag any group where numbers, materials or ratings differ.
Do not decide which note is right.

One group from a run, as an illustration:

The element list in the prompt is there because of a failure: without it, the first run lumped external wall insulation and roof insulation into one group and produced a flag that meant nothing. The person decides which note governs, usually by asking the engineer or checking the thermal calculation, and fixes the other two sheets.

References can be handled the same way. Ask for every callout that points elsewhere (3/A-302, "see structural drawings") with the sheet it appears on, then compare against the register and the detail numbers that exist on each target sheet. In the example the check finds a section marker on A-201 pointing to detail 3 on A-302, a sheet that carries only details 1 and 2, and a note on A-401 sending the reader to a lintel schedule that no structural sheet contains. Both are dangling references, and the second is the kind of gap a contractor discovers on site.

Schedules need a spreadsheet match rather than a model: export the door and window schedules, list the tags that appear on the plans, and compare the two columns. The example turns up door D14 on the ground floor plan but not in the schedule, and window W06 listed at 1,200 by 1,350 mm in the schedule while the elevation note says 1,200 by 1,500 mm. Someone chooses which governs and corrects the other.

Check three: dimensions, where a person holds the pen

Missing dimensions are the check people most want from AI and the one it does worst. Whether a given wall has a setting-out dimension depends on seeing every line on the sheet, and the limits in the table above explain why that is unsafe. What works better is a method that makes gaps obvious to a person in minutes.

1. Overall test. On each floor plan, tick that every external face has an overall dimension and every opening is dimensioned to a datum.
2. Chain test. Add each chain and compare it with the overall. Use a calculator, or paste the numbers you typed into a model for the arithmetic only.
3. Cross-floor test. Compare the same wall line on every level. This is the one that catches coordination errors.
4. Scale test. Measure one known dimension on the PDF at its plot size and confirm the scale statement still holds after export.

A quick sum shows why the cross-floor test earns its place. On the ground floor plan the rear wall carries a chain of 3,600 + 2,400 + 1,200, which is 7,200, and an overall of 7,200. On the first floor plan the same line carries an overall of 7,500, so one level is 300 mm out. Each sheet is internally consistent, so a sheet-by-sheet read finds nothing; only comparing levels does, and a table of dimension strings per wall line makes that comparison quick.

If you do want a model to read dimensions, give it a crop at a readable size and make it admit doubt.

This is a crop of the first floor plan at 1:50: bedroom 2 and the landing.
List every dimension string you can read, with its value and what it appears to measure.
Then list any wall or opening in the crop that has no dimension you can see.
Write "unclear" instead of guessing.

Treat the result as a pointer. A realistic failure is the model reading 2,450 where the sheet says 2,540: a transposed pair of digits that looks perfectly plausible. So every value it reads goes through a person before it goes anywhere, and an "unclear" answer is a prompt to look yourself.

A seeded-error test before you trust any tool

Before any tool joins your process, find out what it catches on your own drawings. Copy a finished, already-issued set, plant a dozen known errors in the copy, run the tool, and count. A practice might seed three register or title-block errors, three conflicting notes, two dangling references, two tag or schedule mismatches and two dimension errors. Results a practice might see, invented for the example:

Read it as a product decision. The tool is a worthwhile first pass on the first four categories (8 of 10 found), useless on dimensions (none found, four false alarms), and it created 8 items someone must clear. If clearing a false alarm takes five minutes, that is 40 minutes the tool costs you, and it belongs in the sum. Repeat the test after any change of tool, model or prompt, since a vendor update can move the results in either direction.

Dedicated drawing-review products exist, and the test is also how to judge them. Bluebeam, for instance, lists AI drawing reviews on its Max plan ($590 per user per year on annual billing, shown as an introductory price) and describes its Smart Review checks as including missing sheets, missing door or plumbing fixture tags and gridline consistency. Compare a product's list of checks with the table at the top of this tutorial, then run your seeded set through its trial before paying.

The QA checklist and the sign-off page

This is the checklist as a practice might keep it for each issue, ready for initials. Keep it as a one-page sheet at the front of the issue folder.

The sign-off rule is short. The checker is not the author of the sheets. AI never initials a box. An item is closed only when a person has looked at the sheet, and open items are listed on the page with the reason each is accepted. Once the set is signed off, the issue register and transmittal log records what went out and to whom, which is what you point to if an issue is questioned later.

Keeping client drawings confidential

A drawing set holds more than lines. It shows how a family's home is laid out, where the alarm panel and any safe room sit, and what a business keeps behind a locked door. Consultants' drawings are also their copyright, and the appointment terms often restrict copying. Work through these steps before any sheet goes into an AI tool.

1. Read the client contract and any confidentiality terms. Some forbid third-party processing outright. Your insurer may also have a view; the tutorial on copyright and client data risks of AI in architecture practices covers both.
2. Pick a plan that does not train on your content. Business plans of ChatGPT and Claude don't by default. On a personal plan, find the training setting and switch it off first.
3. Send text, not drawings, wherever the check allows. The register, note lists and schedule exports contain no geometry and can be compared in a spreadsheet on your own computer. Only the plain-language explanation of differences needs a model at all.
4. Replace names with a project code. Crop or redact the client's name and address from anything you send. Leave out sheets showing security systems, and never send the native model or CAD file, which carries user names and file paths along with the geometry.
5. Strip file properties from PDFs using your PDF software's option to remove metadata, since author names, software paths and sometimes the original file location travel with the file.
6. Log it. Record the project code, which sheets or text went where, on which plan, and who sent them. If a client later asks what was shared, the log answers in one minute.

The same rules, written as a practice-wide policy, are in the AI policy checklist for drawings and client data. For a broader view of what a model can and cannot read from a sheet, Can AI tools read construction drawings? goes through what to check afterwards.

What the process saves on a 38-sheet set, and what it does not

Invented times for the example set, for a practice that checks by hand today:

The saving in hours is modest, roughly one hour on a set this size, because the geometry checks stay with a person. The better case for the process is the second pass over the text, where a tired checker skims and a list does not. Costs are small: ChatGPT Business and Claude Team are each $20 a seat on annual billing ($25 monthly) with a two-seat minimum, and a spreadsheet you already own does the matching. If you want overlay and compare, Bluebeam's Core plan lists it at $330 per user per year, while the AI review features sit on the $590 Max plan. Run the seeded-error test before buying either.

Questions practices ask about AI drawing checks

Can ChatGPT or Claude find missing dimensions on a plan?

Not reliably. Both vendors say their vision models can be inaccurate on rotated or very small text, are approximate on spatial tasks and counts, and may shrink an image before analysis. Use AI to add up chains you give it, or to list where a person should look, and let a person check each overall dimension against its chain.

Is it safe to upload client drawings to an AI tool?

Only after checking the client contract, any confidentiality terms and the tool's data settings. Business plans of ChatGPT and Claude don't train on your content by default; personal plans need the training setting switched off. Crop or strip names, addresses and security layouts first, and keep consultants' drawings out unless their terms allow it.

Does AI replace the checker who signs a set off?

No. It produces a list of candidates, and a named person who did not draw the sheets decides what is a real error and what is noise. The sign-off page should record who checked what and when, because that record is what you rely on if an issue is questioned later.

How many sheets can I upload at once?

It depends on the tool. For example, Claude reads PDFs of up to 100 pages with their visual layout, and reads only the text of PDFs from 101 to 1,000 pages, so a 130-sheet set loses its graphics unless you split it. Splitting by discipline or into batches of 30 to 40 sheets also keeps the flags easier to review.

Further reads

• AI Tools for Shop Drawings: Checking Fabrication Drawings Against the Spec — The same discipline applied to fabrication drawings against a specification.
• How to Use Claude AI in Architecture: Specs and Planning Reports — Use Claude on the specification and report side of the job.
• Architecture Firms Adopting AI Design: What to Tell Your PI Insurer — What to tell your professional indemnity insurer about AI checks.
• AI Floor Plan Generation for Architects: Where It Fails a Real Brief — Why generated plans fail a real brief, and what to check.
• AI Policy for Architects: A Ready-to-Use Checklist and Clauses — Turn this checklist into a practice-wide AI policy.
• How Much Does AI Cost a Small Architecture Practice per Month? — Monthly AI costs for a small practice, line by line.
• AI Programs for Architects: Room Data Sheets From a Client Brief — A worked example of turning a messy client brief into room data sheets with AI, and the review pass that catches what the draft invents or drops.
• AI for Architects: Draft Design Schematics From a Client Brief — A brief-to-schematic workflow for small practices: stress-test the brief, draft an area schedule and options, then run the checks that stay with the architect.
• Architect Meeting Minutes: Record, Summarise and Issue Same Day With AI — A same-day workflow for architects: consent to record, a minutes template, a decisions log and an issue sheet, using Otter, Teams Copilot or Meet notes.
• How Can Architects Use AI on Planning Applications? Where to Draw the Line — Where AI helps on a planning or permit application, where it stops, and the checks an architect runs before anything is submitted.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI developer guide on image inputs and limitations; Anthropic vision documentation and file-upload help page; Autodesk Revit help on exporting a schedule; Bluebeam, ChatGPT Business and Claude pricing pages (all checked 4 November 2026).