AI for Civil Engineering Firms: Proposals and Site Reports First

A civil engineering firm can use AI safely on writing and organising work: proposals, site visit reports, meeting minutes, emails and checklists, each reviewed by a named person before it leaves. It should stay out of stamped calculations because the sealing engineer must be in responsible charge of the design and able to reproduce it without the tool.

That second half is not caution for its own sake. The seal says the engineer directed the work, and a reviewed chatbot answer is not the same thing. The first half is where the hours are: most firms lose far more time to proposals and reports than to calculations. So the practical split is by document type, not by tool: writing and organising work goes to AI with a named reviewer, and design work stays with the engineer and verified software.

Follow me on Instagram@sagnikteaches

Why the stamp changes the rules

Three published sources point the same way.

Connect on LinkedInSagnik Bhattacharya

• The NSPE ethics code, as quoted in its Board of Ethical Review case on AI, says engineers shall not affix their signatures to any plans or documents dealing with subject matter in which they lack competence, nor to any plan or document not prepared under their direction and control.
• In that case, the Board found an engineer's use of AI-assisted drafting tools was not unethical in itself, but that failing to maintain responsible charge over the tool and its output before sealing the document was unethical. It added that technology must never replace engineering judgement, and that reviewing a design at a high level had been insufficient where errors slipped through.
• One state licensing board approved guidelines in July 2025 saying a licensee who uses AI must keep responsible charge of all aspects of the work, exercise direct control, maintain oversight of all data inputs and be capable of reproducing the output independently of the AI program. It also asks for records of inputs, outputs and validation, and says licensees should not use AI on projects where they lack established competence.

The same board's guidelines go further on data: use only inputs you have verified as accurate and complete, keep direct control over what is given to the model, and regularly review and validate outputs against professional standards. Put together, "I checked the answer" is not the test for a sealed calculation. The test is whether you could have produced and defended it yourself. A language model gives you an answer without the working you would need to defend, so it fails the test by design. Rules differ between licensing boards, so confirm yours before you write the firm's policy. This tutorial is not legal advice.

Subscribe on YouTube@sagnikteaches

The safe-use map: three zones

The question to ask of any task is whether a wrong answer would be caught by the next person in line, and whether you could rebuild the answer from your own inputs. If no to either, it is in the never zone.

Proposals: the biggest saving with the lowest risk

Proposals are mostly assembling known material: scope descriptions, team bios, past-project summaries, approach statements and compliance with the client's requirements. A model is good at restructuring that into the client's headings, and the engineering content comes from your own past proposals.

A prompt built on your own material

I am preparing a proposal for a culvert replacement study.
Below are (A) the client's scope text and (B) two paragraphs from our
previous proposals on similar work. Write an approach section of about
350 words using the client's headings. Use only facts in A and B.
Do not state any durations, fees or standards not present in A or B.
Mark anything you need from me as [CONFIRM: ...].
A: [pasted scope]
B: [pasted paragraphs]

Illustrative output: a clean approach section in the client's order, with two flags such as "[CONFIRM: survey method and who carries it out]" and "[CONFIRM: whether permit applications are in scope]". Those flags are the value. The model has shown which parts the client asked for and your material did not answer. What the reviewer fixes: it reads "in accordance with applicable design standards" in a place where the client named a specific standard, so you add the name yourself.

For larger bids, a compliance matrix prevents missing a requirement; the RFP matrix tutorial goes deeper. For tool choices and what proposal software costs, see the proposal software tutorial.

What never goes in the proposal from the model

Fees, hours, staff availability, claims about experience, and any statement of capability you have not checked. A proposal is a representation to a client, and an invented project reference is a serious problem. Treat every factual claim as needing a source in your own files.

Site visit reports: from notes to a clean draft

A site visit produces notes, photos and a deadline. The repeatable part is turning rough observations into a structured report. The engineering judgement, what the observation means, stays in your notes.

Turn these field notes into a site visit report with headings: Purpose,
Attendees, Observations (numbered), Issues requiring action (owner, due
date), Photos referenced, Next visit. Keep my wording where it is
technical. Do not add observations, causes or recommendations I did not write.
Flag anything unclear as [CLARIFY].
Notes:
- 14 May, backfill at headwall east side, 2 lifts visible, compaction tests not on site
- Silt fence down for approx 6 m on north bank, contractor aware
- Pipe bedding looks as drawn, did not see joints, trench closed
- Next visit Thursday, ask for test results

Illustrative output: a tidy report with the four observations numbered, two issues with owners marked [CLARIFY] because the notes name no owner, and "joints not observed, trench closed" kept exactly as written. The risk to check for: a model that writes "bedding conforms to drawings" where you wrote "looks as drawn", a leap from observation to conclusion that may carry liability. Your reviewer compares each numbered observation with your notes and removes any upgrade in certainty.

Photos need care. Do not upload photos showing people, vehicle plates or sensitive infrastructure to a tool you have not vetted. Description in your own words is usually enough for the report. The tutorial on ChatGPT for civil engineers has more report and method statement prompts.

Meeting minutes and action lists

Minutes are a close second. Record the meeting only with participants' agreement, then have the tool produce decisions, actions and open questions from the transcript. Check names, dates and numbers against the transcript, since mishearing a number is the usual error. For example, an action reading "Contractor to submit revised schedule by the 12th" when the speaker said the 21st can cause a week of friction. The person who chaired the meeting is the right reviewer, and the minutes should be marked draft until they have read them.

Five smaller jobs worth having

• RFI and submittal cover letters. Feed your own bullets, receive a clear letter, check clause references yourself.
• Specification tidy-up. Ask for inconsistent terminology or contradictory clauses to be listed, not rewritten, so the specifier decides.
• Checklists. Close-out, pre-construction meeting and QA review checklists generated from your own procedure, then trimmed by experience.
• Plain-English summaries for community or council meetings. Draft from your own technical text, reviewed by the engineer who wrote it.
• Standards digests. Summaries of public documents as a reading aid, always with the original at hand, because a model can turn "shall" into "should".

The pattern across all of them: the model organises and phrases, you supply facts, and a person who understands the engineering reads the result.

A worked example: a twelve-person firm

Take a civil firm with twelve staff: two principals, five engineers, three technicians and two administrators, delivering stormwater, roads and small bridge work. They submit about eight proposals a month, produce roughly 25 site visit reports a month and run perhaps 30 recorded project meetings.

1. Baseline. A principal asks each person to log time for two weeks. Proposals take about six hours each in total across staff, reports about 90 minutes each, minutes about 40 minutes each. That is 48 hours, about 37 hours and about 20 hours a month, or 105 hours.
2. Pick a tool and a rule. One business chat plan for six seats, at the vendor's current list price (ChatGPT Business is listed at $20 a seat annually or $25 monthly, so six seats is about $120 to $150 a month). The rule is the three-zone table above.
3. Trial two document types for a month. Proposals and site reports only, with a review line on each.
4. Measure. Proposals fall to about four hours including review, reports to about 55 minutes. That saves roughly 16 and 15 hours, about 31 hours a month on those two jobs, at a tool cost of about $150.
5. Count the catches. Reviewers found an upgraded observation in 4 of 25 reports and a wrong standard reference in 2 of 8 proposals. Those catches are why review is not optional, and they are the figures the firm uses to decide the tool stays.

Treat the figures as illustrative. Your baseline will differ, and the honest way to know is to time two weeks before and after.

Three firm shapes, three different first moves

The map is the same everywhere, but the best first job depends on where the firm's hours go.

• A land-development firm that bids for many small jobs spends most time on proposals and fee letters. Start there. A library of approved paragraphs and a prompt that restructures them into each client's headings removes most of the retyping. The check is that every fee, duration and standard comes from the firm's own file.
• A municipal sub-consultant producing frequent progress reports and minutes for a lead consultant should start with minutes and monthly reports. Their client's template is fixed, so the model fills headings from notes. The check is that names, dates and action owners match the meeting record.
• An inspection-heavy practice doing construction review will gain most from site visit reports. Short notes in, structured report out, with the engineer who attended as reviewer. The check is that no observation is upgraded into a conclusion.

Quick answers a partner will ask for

Can it check my calculation? It can read a calculation you wrote and list assumptions you did not state, as a prompt for your own check. Treat the list as questions, never as verification. A model that says "looks correct" has verified nothing.

Can it find the code clause? It may name one, confidently and wrongly. Look the clause up in the standard itself. Invented clause numbers are a classic failure and they look plausible in a proposal.

Can it write the design basis memo? The framing and wording, from your bullet points, yes. The design criteria and values, no. They come from your design and your standards.

Can it summarise geotechnical or survey data? A summary of a factual table you supply can save time, but interpretation belongs to the qualified engineer, and the data may be client confidential. Use the stripped-down version of the data, and compare every number in the summary with the table.

What about quantities for payment? Never in the model. A take-off feeding a payment certificate should come from your measured method and be checkable by the contractor.

What if the client asks us to use AI? Ask which tasks. The same zones apply, and responsibility stays with the sealing engineer regardless of whose idea the tool was.

A before and after on a proposal paragraph

Seeing a real edit helps. Before, a paragraph assembled in a hurry: "Our firm has extensive experience delivering high-quality stormwater solutions on time and on budget, and our team will apply best practices to ensure a successful outcome." After, built from the firm's own file and the client's scope: "We completed three culvert replacements in the last five years, each with a hydraulic assessment, a design package and construction review. For this project we propose a three-stage approach: survey and hydraulic modelling, design options with a recommendation, and detailed design with permit support." The first is what a model gives when asked for a paragraph with no facts. The second is what you get when you feed it your own project list and tell it to use only those facts. Check the project count against your records before it goes out, since "three in five years" is exactly the kind of figure that drifts.

One decision rule for any new task

1. Could a wrong answer reach a client, a regulator or a construction site without a competent person reading it first? If yes, it is review-first at best.
2. Could I produce and defend the result myself without the tool? If no, it is never.
3. Does it need client or site data that I cannot strip? If yes, get consent and use a vetted plan, or keep the task out.

Write these three questions on the policy page. People remember them, and they settle most arguments faster than a long document.

Client data and confidentiality

The NSPE code bars revealing facts, data or information without the client's prior consent unless authorised or required by law. In the Board's case, the engineer used the client's information in an AI tool without permission, which was one of the points where the conduct fell short. Check your contracts: public agencies sometimes restrict where project data may be stored or require approval for new tools.

A practical approach is three questions before any upload. Has the client permitted it? Does the plan's contract protect it, with no training on business content by default? Could the same result come from text with names, addresses and site details removed? If the third answer is yes, use the stripped version. What to check in an AI tool's privacy terms shows how to read the terms.

Review sign-off that actually works

A policy that says "review all output" is too vague to enforce. Make the review a named step on the document.

AI use sign-off (project file)
Document:        Site visit report 18, culvert replacement
Tool / plan:     [tool], business plan
Task:            Structured draft from field notes (no photos uploaded)
Reviewed by:     [name], P.Eng., who attended the visit
Checked:         All 6 observations against notes; 1 upgraded statement
                 corrected; 2 [CLARIFY] items resolved
Released:        [date]

Two principles keep this honest. The reviewer must be someone who could have written the document, so a junior does not review a principal's report for engineering content. And the review must compare against source, notes, drawings or the transcript, not just read for sense.

Disclosure and contracts

The Board's opinion was that there is no universal rule requiring disclosure of AI use unless a contract does, yet it said ethical principles favour transparency when AI plays a substantial role in a work product. A sensible position for a firm is to add a standard line to your terms: you may use AI tools for drafting and administration, no client confidential information is entered without consent, and all work is reviewed by a qualified engineer. Where a client contract says otherwise, the contract wins. Public clients can raise AI in procurement questionnaires or contract terms, so keep a short, accurate answer ready.

PDF markup with AI: what Bluebeam's top tier adds

For a first trial, a general business chat plan is enough. If your team lives in PDF markup, Bluebeam's AI features sit on the top tier only. Its pricing page lists Basics at $260, Core at $330, Complete at $440 and Max at $590 per user per year, with AI drawing reviews, AI drawing comparisons, AI multi-view stitching and AI integration via MCP listed under Max, and the Max figure shown as an introductory price. Moving a team from Core to Max is $260 more per user each year, a significant jump for a small firm, so confirm current pricing and test on drawings that do not carry client confidential data first. Treat an AI drawing review as a second reader that flags items for the checker, never as the check itself. The engineering firms tutorial on proposals, specifications and minutes compares more options.

A first month, week by week

• Week 1: time your baseline for proposals, reports and minutes, and draft the three-zone map for partners to approve.
• Week 2: pick the tool, read its terms and set up two seats. Run proposals and one site report with sign-off.
• Week 3: widen to six people, add minutes, log the catches.
• Week 4: compare time and catches, decide what to keep, and write the one-page policy. A template is in the acceptable-use policy tutorial.

If you would rather map this with someone, that is what my AI implementation consultation covers. Whatever you choose, keep the rule that is easy to remember: AI can write about the engineering, but the engineering itself, and the seal, stay with the engineer.

Questions civil engineering firms ask about AI

Can a junior engineer use AI to draft a calculation package?

Not as the working method for anything that will be sealed. The licensed engineer must be in responsible charge of the design and able to reproduce the result without the tool. A junior may use a verified spreadsheet or approved software, with the checker reviewing method, inputs and outputs, not a chatbot's answer.

Do we have to tell clients we use AI?

The NSPE Board of Ethical Review found no universal requirement to disclose AI use unless the contract demands it, but said ethical principles favour transparency when AI plays a substantial role. Check your contracts and any client AI clauses, and put a plain statement in your standard terms.

Is it safe to upload a client's site data to ChatGPT or Claude?

Only if the client has allowed it and the plan's terms protect it. The NSPE ethics code bars revealing client information without prior consent unless authorised or required by law. Use a business plan with no-training terms, strip names and locations, and keep sensitive infrastructure details out.

What is the cheapest sensible first step?

A business chat plan for a small group, a one-page rule and one trial on proposals or meeting minutes. ChatGPT Business is listed at $20 per user per month billed annually or $25 monthly, with a two-seat minimum. Check the vendor page for current prices and terms before buying.

How should we record AI use on a project?

Add a line to the project file: tool, task, inputs used, who checked the output and what changed. Keep it with the document it supports. If a client or board ever asks how a report was produced, you can answer in a minute.

Further reads

• How Engineering Firms Are Using AI: Six Jobs and Their Limits — Six jobs AI does for engineering firms, and their limits.
• HubSpot AI for Heavy Civil Contractors: Does a CRM Fit Bid Work? — How a heavy civil contractor uses a CRM with AI.
• AI for Civil Construction: Plant Pre-Start Checks and Service Schedules — Plant pre-start checks with AI on civil sites.
• Construction Final Account: Variations, Deductions and Sign-Off With AI — A final account worked end to end.
• HR Consulting AI: Sell an AI Policy and Handbook Update Package — How AI policies are packaged for professional clients.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check readiness before rolling out AI in a firm.
• AI for Consulting Engineers: Fee Proposals and Reports, Not Calculations — A green, amber and red rule for AI in a small engineering firm: proposals, reports and minutes with checks, calculations never, plus a sign-off note.
• AI for Land Surveyors: What Helps and What Never Replaces Field Checks — Which survey tasks AI can draft, which it can only check, and which never leave the surveyor, with a worked check-shot test and a one-page staff rule.
• Engineering Company AI: Setup Plan and Costs for a Small Firm — Three cost scenarios for a ten-person engineering firm, an eight-week rollout plan, and hour-by-hour sums for tenders, reports and QA review.
• Best AI Tool for an MEP Engineering Consultancy: Specs, Clashes, Handover — ChatGPT, Claude and Copilot compared on three MEP documentation jobs, with a hard line that design calculations never go through them.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NSPE Board of Ethical Review, Artificial Intelligence in Engineering Practice (BER Case 24-2); a state engineering board's Guidelines on the Use of Artificial Intelligence in Engineering and Surveying Services (approved 16 July 2025); OpenAI ChatGPT Business pricing page; Bluebeam pricing page.