Yes, partly. ChatGPT or Claude can catch missing sheets, contradictory notes, schedule rows with no plan tag and callouts to details that do not exist, provided you give them selectable text and exported schedules rather than screenshots. They miss graphical problems and sometimes invent findings, so the architect who signs the set must verify every flag.
The pages that rank for this search are mostly AI design and rendering tools. They make pictures; they do not read your sheets. Checking a set is a different job: extracting text, comparing lists and following cross-references. That is a good fit for a general assistant plus a spreadsheet, and this workflow is written for a practice of two to ten people with no developer on hand.
What the pages that rank for this keyword actually sell
Take one popular example. MyArchitectAI's site lists rendering from an uploaded SketchUp or Archicad design, style transfer, render editing, low-resolution enhancement and video. It does not mention drawing checks or schedules, and that is typical of the category: useful for client images, silent on whether your door schedule matches your plans.
Three other groups do touch drawing review, and it helps to know which is which before you spend anything.
- PDF review software. Bluebeam's pricing page shows Basics at $260, Core at $330, Complete at $440 and Max at $590 per user per year, billed annually (the page shows a dollar sign without a currency code, and flags Max as an introductory price). Only Max lists AI features: "AI drawing reviews", "AI drawing comparisons" and a Smart Review tool described as scanning construction documents for design issues such as missing sheets and missing door or plumbing fixture tags. Bluebeam's own AI page says Smart Review and Overlay are in preview and that Smart Review is optimised for vertical commercial construction, which is a different job from a yoga studio fit-out or a house extension.
- Dedicated review services. Nomic's Drawing Review agent says it checks every page of a drawing set against your firm's own QA/QC standards, across sheets, schedules, details, notes and references, and links each finding to the exact spot on the drawing, with Bluebeam-ready PDF markups. Nomic publishes no price for the agent (its individual plans start at $20 a month, but the page does not say they include it), so ask for a quote, and note that it assumes you have written QA/QC standards for it to check against.
- General assistants. ChatGPT and Claude do not know your standards until you tell them, and they read PDFs within limits. They cost $20 to $25 per person per month at list price, and they are what the rest of this tutorial uses.
The fair way to choose between them is to run the same seeded test on each, which the worked example below shows. If you want the wider picture of which assistant suits an architect, Claude vs ChatGPT vs Gemini for architects compares them on cost and fit.
Prepare the inputs: selectable text, exported schedules, one code
Models are more reliable on text than on pixels. Anthropic's developer documentation warns that Claude "might hallucinate or make mistakes" on low-quality, rotated or very small images, that its counts of objects are approximate, especially for many small objects, and that its location outputs are approximate and should be verified. Drawings are full of small, rotated text and repeated symbols, so give the assistant text wherever you can.
- Test your PDF. Open the issue PDF and try to select a general note with the text tool. If the words highlight, the file contains real text. If nothing highlights, the sheets are pictures (a scan or a flattened plot), and every check below gets less reliable. Re-plot from the model instead of scanning.
- Export schedules from the source. In Revit, open the schedule view, then File, Export, Reports, Schedule. Revit saves a delimited text file and lets you choose the field delimiter, such as tabs or commas, and whether column headers export. Open the file in Excel through its text import tool (Data, From Text/CSV) so the columns split cleanly. Other BIM and CAD packages have their own export commands; check yours. Export the sheet list the same way (in Revit it is a sheet list schedule).
- Keep one issue code. Put the same issue code and date in every file name. An assistant compares whatever you hand it, so a stale schedule from last month's revision produces confident nonsense.
- Mind the limits. Claude's help centre (updated July 2026) says chats take up to 20 files of up to 500MB each, and that PDFs of 1 to 100 pages are read as text plus page images while 101 to 1,000 pages are read as text only. Anthropic's documentation adds that dense PDFs can fill the context window before reaching the page limit and suggests splitting the document. Cap each upload at about 25 sheets. If you upload sheets as images instead, expect small text to suffer: Anthropic's documentation notes that large images may be resized, which can make text less legible, so crop big sheets into sections rather than uploading a whole plot.
- Protect the drawings. Use a business plan (ChatGPT Business or Claude Team, both with a two-seat minimum) so your content is not used for training by default, and follow your practice's AI policy. The architecture firm AI policy checklist covers client data and consultant drawings.
A seeded test set: 22 sheets, 18 planted errors
Do not trust any tool on a live issue until you have seen it fail on a set where you know the answers. The figures here are invented, but the method is the point. A six-person practice takes a closed project, a 22-sheet issue set for converting a ground-floor unit into a yoga studio with changing rooms, and plants errors in a copy. The sheets run from A-000 (cover and drawing index) and A-001 (general notes and abbreviations) through plans, elevations, a section, enlarged plans, details, three schedules and a fire strategy plan.
The planted errors, 18 in all, fall into four groups: seven in the sheet index and title blocks, four in the general notes, five in the door and window schedules and two that can only be seen graphically (a door swinging into a cubicle partition, and a section marker pointing the wrong way). Then the practice runs the four checks below and records what comes back.
| Check | Real issues planted | Found | Missed | False alarms |
|---|---|---|---|---|
| Sheet index and title blocks | 7 | 7 | 0 | 0 |
| General notes | 4 | 4 | 0 | 3 |
| Door and window schedules | 5 | 5 | 0 | 2 |
| Graphical items | 2 | 0 | 2 | 0 |
| Total | 18 | 16 | 2 | 5 |
Sixteen of the 18 found, with five false alarms among 21 flags, is a useful result for a second reader and a poor result for a final check. The time was roughly two hours preparing inputs, 40 minutes on the prompts, an hour and a half triaging flags and half an hour writing the check log: about four hours and 40 minutes. The saving is not the point. The point is a reader that is as attentive at sheet 19 as at sheet 2, and that costs a few dollars a month.
Check 1: the sheet index against the title blocks
The cover sheet says the set has 24 sheets. The PDF has 22 pages. Rather than ask an assistant to eyeball that, make it extract the facts and let a spreadsheet do the comparing, because spreadsheet comparisons are exact and assistants are not. First prompt, with the PDF attached:
The attached PDF is an architectural drawing set. For every page, read the title block and return one table row with: PDF page number, sheet number, sheet title, revision letter, revision date, project number. If a field is unreadable, write UNREADABLE. Do not guess and do not copy values from the drawing index on the cover sheet; read each title block separately.Paste the result into one Excel tab and your exported sheet list into another. Then add a column to each that counts matches in the other:
=COUNTIF(SheetList!$A:$A, A2) (on the extracted tab: 0 means this sheet is not in the index)
=COUNTIF(Extracted!$B:$B, A2) (on the sheet list tab: 0 means the index lists a sheet that is not in the PDF)Anything scoring 0 is a finding, and a lookup column comparing revision letters finds the rest. In the seeded set this produced seven findings: three index entries with no sheet (A-505, A-703 and A-704), one real sheet missing from the index (A-802), two sheets whose revision letter in the index differed from the title block (A-102 and A-501) and one title block carrying the wrong project number. Ask the assistant to explain any mismatch, but make it give the PDF page number each time. If it cannot, treat the claim as unverified.
Check 2: general notes that contradict each other or point nowhere
This is where an assistant earns its fee, because notes are free text and no spreadsheet function reads them. Paste the numbered notes from sheet A-001 and the sheet list into a fresh chat:
Below are the numbered general notes from sheet A-001, followed by the sheet index. Do five things.
1. For each note, list every reference it makes: sheet, detail (such as 3/A-504), schedule, specification section or consultant drawing.
2. List references to sheets that are not in the index.
3. List pairs of notes that give different values or requirements for the same item (ceiling height, fire rating, finish, fixing).
4. List abbreviations used in the notes but not defined in the abbreviations list.
5. For every finding, quote the exact words and give the note number. Do not rewrite any note and do not decide which version is correct.
NOTES:
[paste]
SHEET INDEX:
[paste]A plausible reply names note 26, "Refer to door schedule on sheet A-602", and points out that A-602 is the window schedule in the index. It flags notes 19 and 24, which give ceiling heights of 2,700mm and 2,600mm for the same studio, and notes 7 and 15, which give a 30-minute and a 60-minute fire rating for the same changing-room partitions. It lists note 12's reference to detail 3/A-504 as one to verify, which Check 4 settles. It also lists three problems that turn out to be false alarms: two abbreviations that are defined in a legend drawn as a graphic, which the text extraction never saw, and a "contradiction" between notes 3 and 8 that are both correct because they apply to different zones.
Fixing the real ones is quick. Note 26 before: "Refer to door schedule on sheet A-602." After: "Refer to door schedule on sheet A-601." Notes 19 and 24 need a decision from the architect against the sections, and notes 7 and 15 need one against the fire strategy sheet. The assistant must not make either decision: it has found a conflict, and only the drawings and the architect's judgement say which version is right. The same applies if a note and the door schedule disagree on a rating.
Check 3: door and window schedules against the plans
Schedules and plan tags drift apart when doors are added late. If both come from a BIM model they are less likely to disagree, because the tag and the schedule row read the same object. They still diverge when tags are placed by hand, when sheets come from several consultants, or when practice standards leave notes outside the model. In 2D CAD, the schedule is a table somebody types, so the risk is larger.
Do it in three passes. First, get the marks from the plans: ask the assistant for every door mark it can read on each plan sheet, with the sheet number, and run the request twice. Where the two runs differ, look at those sheets by hand. Second, put the marks in a spreadsheet beside the exported schedule and use COUNTIF as before. In the seeded set, marks D-25 and D-26 appeared on plans but not in the schedule, and D-17 sat in the schedule with no mark on any plan. Third, give the assistant the schedule and the dimension text on the enlarged plans and elevations and ask it to compare only values that are printed:
Here is the door and window schedule as CSV, and the text extracted from sheets A-201 and A-401. Compare only values that are printed in both places (width, height, fire rating, hardware set). List every mismatch with the mark, the schedule value, the drawing value and the sheet. List any hardware set number in the schedule that does not appear in the hardware set list below. Where the data does not let you decide, say "cannot tell". Do not estimate sizes from the drawing.
SCHEDULE:
[paste]
HARDWARE SETS:
[paste]That prompt found two more planted items: door D-09 pointed at hardware set HW-3, which the hardware list does not contain, and window W-05, scheduled at 1,200 wide by 1,500 high, carried an elevation note of 1,500 by 1,200. A swapped width and height is the classic schedule error and one a tired human skims past. It also produced two false alarms. One door was flagged as missing from the plans because its tag appears only on an enlarged plan the extraction skipped, and one window height "mismatch" came from the elevation dimensioning to a different datum than the schedule. Both cost a minute each to dismiss, which is the price of a second reader.
Check 4: callouts that point at details that do not exist
Callouts such as 3/A-504 mean detail 3 on sheet A-504. Broken ones tend to turn into contractor queries. Extract the callout text from the notes and plans, then the detail numbers that exist on each referenced sheet, and compare:
From the text below, list every callout written in the form number/sheet (for example 3/A-504). Then, for each sheet named, list the detail numbers that actually appear as detail titles on that sheet in the extracted text. Report every callout whose detail number is not on the target sheet, with the page number where the callout appears. If a callout has been split across lines by the extraction, say so instead of guessing.
[paste]Callouts drawn as bubbles often extract as a number on one line and a sheet reference on the next, which is why the prompt asks the assistant to say when a split makes it unsure. In the seeded set this independently caught the 3/A-504 reference, because sheet A-504 carries only details 1 and 2, which is a useful cross-check on Check 2. It could not see the other planted problem on this theme, a section marker on a plan pointing the opposite way to the section drawn on sheet A-301, because direction is graphical. Keep graphical checks, including door swings, clashes, escape distances and section directions, with a person.
Where the AI misreads a drawing set
Four failure modes recur, and each has a defence.
- Findings that are not there. A typical case: the assistant reports that "A-505 Details 5 appears on PDF page 17" when it has read the index, not a title block. Ask for the PDF page number and a quoted phrase every time, and open page 17.
- Confident silence. "No issues found" is not evidence. Spot-check ten unflagged items per check, such as ten doors or five notes, before relying on a clean result.
- Counting and positions. Anthropic states that counts of many small objects can be inexact and positional outputs approximate. Do not ask an assistant to count doors from a plan image. Count tags from text, or use your PDF viewer's search, and let a spreadsheet total them.
- Dropped content. A legend drawn as a graphic, text converted to outlines, or a long set split across chats all hide information. Say in each prompt what you did and did not provide, and ask the assistant to list what it could not read.
Revision comparison is a separate job and a better fit for overlay tools; the tutorial on catching changes between drawing revisions covers it. For a broader look at what assistants can and cannot read on plans, see what to check after AI reads construction drawings.
What the architect checks, signs and logs
The assistant produces a list of things to look at. It resolves nothing. Every flag goes into a log with a decision, and the log goes into the project record next to the issue. Here is an extract from the seeded test's log:
| Flag | Sheet | Finding | Decision | Checked by |
|---|---|---|---|---|
| 1 | A-001 | Note 26 cites A-602 for the door schedule | Real. Corrected to A-601 in rev D | Initials, date |
| 2 | A-001 | Notes 19 and 24: ceiling heights 2,700 and 2,600 | Real. 2,600 confirmed from section; note 19 amended | Initials, date |
| 3 | A-102 | D-25, D-26 tagged, not scheduled | Real. Rows added to schedule | Initials, date |
| 4 | A-401 | "FFL" undefined | False alarm. Defined in legend graphic | Initials, date |
| 5 | A-201 | W-05 1,200 x 1,500 vs 1,500 x 1,200 | Real. Schedule corrected, elevation confirmed | Initials, date |
Three rules keep this honest. First, a person opens the sheet for every flag; nobody ticks an assistant's claim from memory. Second, the checker samples items the assistant did not flag, so the log shows what was looked at as well as what was found. Third, keep the prompts and replies with the log, because they show how the check was done if a question comes later. The architect who signs the set is answerable for it, and what to tell a professional indemnity insurer about this routine is a question for your own policy and broker. The tutorial on AI disclosure to your insurer is a place to start the conversation.
What each route costs for a six-person practice
These are list prices as of October 2026, and each vendor can change them, so confirm before you buy.
| Route | What it can check | List price | Quick sum |
|---|---|---|---|
| ChatGPT Business or Claude Team | Text-based checks 1 to 4 as above | $20 per seat a month billed annually, $25 monthly; two-seat minimum | Six seats: $120 a month annually billed, $150 monthly |
| Individual Plus or Pro plan | The same, with consumer data settings to review | $20 a month each | Six plans: $120 a month |
| Bluebeam Max | AI drawing review and comparison; Smart Review in preview | $590 per user per year, annual billing | Two checkers: $1,180 a year |
| Dedicated review service | Whole-set review against your QA standards | Not published; ask for a quote | Unknown until quoted |
| Spreadsheet comparisons | Exact list matching | Already owned | None |
For most small practices the first row plus a disciplined log is the whole budget. Add Bluebeam Max or a dedicated service only if the seeded test shows a gap the assistant cannot close, and test it on the same 22 sheets before paying.
Running it on your next live issue
Treat the first two projects as a trial. On the first, run the workflow beside your normal check and compare what each found. On the second, let the assistant go first and have a person review its flags, then do the usual check. Track three numbers: errors found that the normal check missed, false alarms per set, and minutes spent. If the false alarms outweigh the finds after two jobs, tighten the prompts (for example, give it the abbreviation legend as text) before dropping the idea.
To build the habit, make the issue register do double duty. The tutorial on drawing issue registers and transmittals shows a sheet list you can export each time, which is exactly the input Check 1 needs. If your sets come out of Revit, keep the exported sheet list, door schedule and window schedule in the project folder under the issue code, and the checks become a twenty-minute routine instead of a project.
Questions about AI checks on drawing sets
Can ChatGPT or Claude read a whole drawing set?
Often, if the PDF has selectable text and the set is modest. Claude's help centre says PDFs up to 100 pages are read as text and page images, longer ones as text only, with 20 files per chat. Dense sheets can still fill the context window, so split a big set by discipline or sheet range and run one check per chat.
Is it safe to upload client drawings?
That depends on your contracts, your AI policy and the plan you use. Business plans from OpenAI and Anthropic do not train on your content by default, while consumer plans have a switch you must check. Read any confidentiality clause first, remove client names where you can, and ask a lawyer if a contract is unclear.
Do I need Bluebeam Max or a dedicated review product?
Not to start. Bluebeam lists AI drawing reviews on its Max plan at $590 per user per year, and says its Smart Review is in preview. Run the seeded test in this tutorial with a general assistant first, then trial a dedicated product against the same test set and compare found, missed and false alarms.
Will this replace the architect's own check?
No. It is a second reader that does not tire at sheet 19. It misses graphical problems such as door swings and section markers, and it can invent findings. The person who signs the set still checks every flag and a sample of the items it did not flag.
How many errors should I plant in a test set?
Fifteen to twenty in a copy of a finished set is enough. Spread them across the sheet index, general notes, schedules and at least two purely graphical items. Count what the workflow finds, misses and wrongly flags, and repeat the test whenever you change the prompts or the tool.
Further reads
- Copyright and Client Data Risks of AI in Architecture Practices — What to protect before you upload drawings anywhere
- AI Tools for Architects 2026: A Toolkit From Brief to Handover — Where drawing checks sit in a brief-to-handover toolkit
- AI Tools for Shop Drawings: Checking Fabrication Drawings Against the Spec — Checking fabrication drawings against the specification
- AI Programs for Architects: Room Data Sheets From a Client Brief — Room data sheets give your schedule checks a source to compare against
- AI Floor Plan Generation for Architects: Where It Fails a Real Brief — A test protocol for architects: write one real brief as a ledger, score a generated plan against it, and set the sign-off rules before clients see anything.
- Architect Site Visit Report: Notes and Photos to a Clear Report With AI — A worked notes-to-report example for architects: observation, instruction and action columns, photo captions, and a rule that keeps unseen items out.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Claude help centre (uploading files to Claude, updated July 2026), Anthropic developer documentation (PDF support, vision limitations and image resizing), Bluebeam pricing and Max pages, Nomic automated drawing review page, MyArchitectAI website, ChatGPT and Claude pricing pages.