Can ChatGPT or Claude do civil engineering work?
ChatGPT and Claude help civil engineers with emails, calcs, code lookups and reports. Here is where they stop on a real deliverable, and why.
ChatGPT and Claude can do a lot of the work around civil engineering: emails, explaining calcs, drafting report sections, cleaning up spreadsheets, and finding the right code chapter. Think of them as copilots. Handed a full deliverable, like a plan check or a grading change in Civil 3D, they usually stop short, and a licensed engineer has to check everything.
"Can't ChatGPT just do this?"
I hear some version of that question from almost every civil engineer I talk to. Usually it comes right after they describe a week spent on redline comments, or a drainage report that is three days late. It is a fair question. These tools got much better this year, and plenty of engineers already use them every day. So I want to answer it honestly: what general AI assistants are good at in civil work today, where they stop, and what it takes to get past that point.
What ChatGPT, Claude and Copilot are good at in 2026
The general assistants are no longer just chat boxes. Claude can now create real Excel, Word, PowerPoint and PDF files from a prompt, and Anthropic opened that to every user in February 2026 (Claude Help Center). In May, Claude's add-ins for Excel, Word and PowerPoint became generally available (Claude for Excel). Claude Cowork, released in January, can work through a folder on your own computer and chain multi-step tasks (Wikipedia). ChatGPT takes PDFs up to 512 MB, with a text cap around 2 million tokens per file (OpenAI File Uploads FAQ). Microsoft Copilot sits inside the Office apps most firms already pay for.
For a civil engineer, that covers a lot of useful ground:
- Writing and answering emails, including the polite reply to a reviewer comment you disagree with.
- Explaining a calculation step by step, such as why the rational method gives a different peak flow than your hydrograph.
- Cleaning up a spreadsheet of inlet or pipe data and turning it into a table for a report.
- Drafting the narrative sections of a drainage or geotechnical summary from your notes.
- Pointing you to the chapter of a design manual that probably covers your question, so you can go read it.
This is real time saved. I would not tell anyone to stop using these tools for this kind of work.
Where they stop on a real deliverable
The trouble starts when you hand one of them a whole job. A plan check or a grading change is not one question. It is dozens of steps across drawings, standards and judgment calls, and every one of them has to be right.
They don't work inside Civil 3D natively
Neither ChatGPT nor Claude opens a native DWG in chat the way Civil 3D does. The usual workaround is to export sheets to PDF, which throws away the surfaces, alignments and feature lines that make the drawing a model.
This gap is closing. Civil 3D 2027 ships Autodesk Assistant, and its Tech Preview can query and change object styles and analyze surface slopes (Autodesk). There are also open-source plugins that let Claude talk to a live Civil 3D drawing through MCP (Model Context Protocol, a standard way to connect AI tools to software) (Civil3D-mcp on GitHub). However, Autodesk labels its features experimental and says "Mistakes are possible; please double-check responses." Moving a pad, rebuilding the surface, checking slopes against the standard and updating the sheets is still work you drive yourself.
They can cite code sections that don't exist
This is the failure that worries me most. Language models are very good at producing something that looks like a correct reference.
Law tracks this closely. Researcher Damien Charlotin's database had logged more than 2,000 court decisions involving AI-hallucinated content by September 2026 (AI Hallucination Cases database). In 2025, Deloitte refunded part of an AUD$439,000 Australian government contract after its report was found to contain nonexistent sources and a made-up quote from a court judgment (Fortune).
Engineering has no public database like that yet, but the same pattern shows up. Washington University's mechanical engineering research guides warn that these tools "might hallucinate the perfect standard" (GotLaw STL), and building officials keep a running thread of confident, wrong code answers (Building Code Forum). To be fair, web search is now on by default in most assistants, so fully invented sources are rarer. The quieter risk is a real link to the wrong section, the wrong edition, or a code your city amended years ago (Dr. Arun VJ).
You end up correcting them over and over
On a long task, a general assistant fixes the error you pointed out, then introduces a new one two steps later. You catch it, it apologizes, and the loop repeats. At some point the engineer is doing the job twice: once to check the AI, and once to redo the parts it got wrong.
They don't know your firm
Projects and memory features help, but they only hold what you paste in. They don't know your CAD standards, your layer names, the county reviewer who always asks for a separate erosion control sheet, or how you solved the same detention problem on a job two years ago. That knowledge lives in the firm, and it is most of what makes a deliverable pass on the first submittal.
Good at vs. stops at
| Task | Good at | Stops at |
|---|---|---|
| Email and reviewer replies | Drafting clear, polite responses fast | Knowing what your firm actually agreed to |
| Calculations | Explaining the method and checking arithmetic | Owning the design inputs and assumptions |
| Code and standards | Pointing to the likely chapter | Quoting the exact section of your adopted edition and local amendments |
| Spreadsheets | Cleaning, formatting and summarizing data | Tying numbers back to the model they came from |
| Reports | Drafting narrative from your notes | Producing a submittal-ready document on its own |
| Civil 3D | Answering how-to questions, early in-app previews | Carrying a grading change through surfaces and sheets |
| Plan check | Reading exported PDF sheets | Traceable findings tied to the sheet, checked against firm standards |
What a civil-specific tool adds
So why do they stop? It is not that the models are weak. Claude and GPT models are remarkable. The gap is everything around the model: access to the real software, the firm's knowledge, and a review step that fits how engineering liability works.
That is what we are building with Gradon. It works inside Civil 3D natively instead of reading screenshots of it. Every plan-check finding quotes the sheet it came from, so you can check it in seconds instead of trusting it. It learns the firm's standards and remembers past projects. Most importantly, a licensed engineer reviews and accepts everything before it leaves the office. The AI does the grinding, and the engineer keeps the judgment and the stamp.
How to use ChatGPT or Claude on civil work today
If you use a general assistant, and you probably should, these are the habits I would build:
- Treat every code or standard citation as unverified until you have read the section yourself in the adopted edition.
- Give it the source. Upload the actual design manual PDF and ask it to quote from that, rather than relying on what it remembers.
- Keep it on tasks where you can check the output quickly, such as emails, explanations and first drafts.
- Never let it be the last reviewer on anything that carries your stamp.
I am optimistic about where this goes. A year ago none of these tools could touch a drawing, and now the first in-app assistants are shipping. The engineers who learn where the line sits today will be the ones who get the most out of what comes next.
FAQ
- Can ChatGPT read DWG files?
- Not natively in chat. The common workaround is exporting sheets to PDF, which keeps the linework as an image but loses the model data behind it, such as surfaces and alignments.
- Is it safe to use ChatGPT or Claude for building code references?
- Use them to find where to look, not as the final answer. Check every section number against the adopted code edition and your jurisdiction's local amendments before it goes into a report.
- Can Claude work inside Civil 3D?
- Only through add-ons. Open-source MCP plugins can connect Claude to a live drawing, and Autodesk Assistant in Civil 3D 2027 has an experimental Tech Preview. Both are early and need close checking.
- Will AI replace civil engineers?
- No. A licensed engineer is legally responsible for stamped work, and that does not change. What changes is how much of the repetitive drafting and checking an engineer has to do by hand.