How Engineering Firms Are Using AI: Six Jobs and Their Limits

Small and mid-sized engineering firms use AI most on text work: fee proposals, specifications and reports, checking documents for consistency, timesheet admin and marketing copy. Those uses are proven because a person can check the output line by line. Calculations, design decisions and anything an engineer signs are hype territory and should stay with people.

The difference between proven and hype is testability. If you can run a use on a document you already finished and compare, it is proven for your firm. If the claim is that AI "designs", "optimises" or "automates compliance", ask for the feature on a vendor page, a limit stated in writing and a way to test it on your own past project.

Follow me on Instagram@sagnikteaches

How to judge a claim: a four-step evidence ladder

Trade-association primers tend to list opportunities and risks in general terms. They rarely tell a principal how to decide on a Tuesday afternoon whether a particular use is worth a trial. This ladder does. Score a use from the top and stop at the first rung that fails.

Connect on LinkedInSagnik Bhattacharya

1. Can you test it on a finished job? Take a report, tender or spec you have already issued, run the use on its inputs, and compare with what you actually produced.
2. Is the feature on the vendor's own page? A named capability with a stated limit counts. A web page saying a tool "transforms engineering" does not.
3. Can a person check every output before it leaves the building? Text, tables and lists pass; a model's numerical answer fails unless independently recalculated.
4. What is the worst realistic failure? A clumsy sentence in a draft is cheap; a wrong load, tolerance or clause in an issued document is not.

Everything below is rated against those four questions. Where a rating depends on tools or prices, the figures are list prices read on vendor pages in November 2026, and they change.

Subscribe on YouTube@sagnikteaches

Six jobs at a glance, with effort and risk

Job 1: proposals and tender responses

What firms do. They paste the invitation, last year's winning proposal and their standard terms into an assistant, and ask for a first draft of the method statement, the scope and assumptions, and a compliance matrix. The firm's engineer still sets the fee and the programme. The tutorials on AI for engineering proposals, specifications and minutes and on building a compliance matrix go deeper on the mechanics.

Evidence. You can test it on a tender you already won or lost. Paste the old invitation and see whether the assistant's compliance matrix matches the one you built by hand. In practice the weak spot is the "assumptions and exclusions" section: models write fluent boilerplate that omits the exclusion this project needs. The fix is to give the assistant your own exclusion list and tell it to use only those.

An illustration. A 12-person civil firm answers about six tenders a month, three hours each on first drafts of the method and compliance sections. With a reviewed assistant draft, the first-draft stage drops to about 90 minutes, so six tenders save roughly nine hours. The firm reinvests that in pricing review, which is the part that wins or loses money.

Effort and risk. Low effort, medium risk. The risk is a fee or scope statement that reads well and is wrong.

Watch-outs. Never let the assistant invent past projects, staff qualifications or references. Give it the real list. Paste the invitation's mandatory requirements into the check and ask for "NOT ADDRESSED" flags rather than a rewrite.

Job 2: specifications and clause drafting

What firms do. They use an assistant to turn a design team's notes into a clause list, to harmonise wording across a spec section, or to compare a draft with an earlier project's spec. The value is in consistency and in a first structure, not in the technical content.

Evidence. The test is a section you wrote last year. Remove the final text, give the notes and your standard clause library, and see what comes back. Models tend to produce plausible clauses that cite standards or test methods by number, and some of those citations will be wrong or out of date. A specification that references the wrong standard can be a contractual problem, so every reference must be checked against the current publication.

Effort and risk. Medium effort because it needs your clause library in a Project or document store; high risk if citations are not verified.

A realistic mistake. An assistant was asked for a drainage-testing clause and produced a neatly worded paragraph citing a test method that had been withdrawn. The engineer caught it only because her checklist had a line "every standard number looked up". Build that line into your process, not your memory.

Job 3: reports, inspection notes and meeting minutes

What firms do. Engineers dictate or type rough site notes and let an assistant turn them into a report in the firm's template, separating observations from opinion. The same technique cleans up meeting minutes and action lists. For inspection workflows see the tutorial on photos and notes to a draft report.

Turn these site notes into a report using the headings below. Use only
the notes. Keep observations, measurements and recommendations in
separate sections. Any measurement or reference you cannot find in the
notes: write [CHECK]. Do not add recommendations I have not written.
Headings: Scope of visit, Observations, Measurements, Assessment,
Recommendations, Limitations.
Notes: [paste]

An illustrative result for a drainage survey returns a tidy report with two [CHECK] flags where a chamber depth was missing and a manhole number was ambiguous. That is the correct behaviour. The common failure is the quiet one: a notes line says "cracking at joint, probably settlement" and the draft hardens it into "settlement has caused cracking". The engineer's assessment is the product; the assistant should never promote a guess to a finding. Read every sentence in the Assessment section against the original note.

Effort and risk. Low effort and medium risk. A report is a signed document, so the signing engineer reads all of it.

Job 4: QA and document checking

What firms do. Two kinds of checking are offered. First, general assistants can compare two documents or check a document against a checklist: does every drawing reference in the report exist in the drawing register, does the report use the same pipe size throughout. Second, some drawing-review tools add AI checking. Bluebeam, for example, lists a Smart Review feature on its top Max tier that looks for items such as missing sheets, missing door or plumbing fixture tags and gridline consistency; Max lists at $590 per user per year, marked as an introductory price, and the lower tiers (Basics at $260, Core at $330 and Complete at $440, also per user per year) list no AI features. Autodesk Assistant now sits inside products including AutoCAD, Revit and Civil 3D; Autodesk's FAQ says there is no general self-service setting to switch the whole experience off, and one of its product guides says chat history is saved for 30 days. Both belong in your confidentiality review.

Evidence. The vendor pages describe the checks; they do not tell you your false-negative rate. Test on a past drawing set where you know of three real errors and see whether it finds them. Count misses as well as catches.

Effort and risk. Medium effort, medium risk. The risk is false reassurance: a tool reports "no issues" and the checker relaxes. Use these as an extra pair of eyes, never as the checker of record.

A quick sum. A Max seat at $590 a year works out at about $49 a month. If it saves one checker six minutes on each of 60 drawing sets a month, that is six hours a month, worth about $540 at a billable rate of $90 an hour. That return holds only if the tool catches things you would otherwise miss while you keep doing the human check; if it merely repeats checks your team already does well, the saving is smaller than it looks.

Job 5: timesheets and admin

What firms do. Engineers dislike timesheets and fill them in on Friday from memory. An assistant can turn a pasted week of calendar entries and a list of job codes into a proposed timesheet that the engineer corrects. It can also draft invoices' narrative lines from the time log and tidy internal emails.

Here is my calendar for the week and our job codes. Propose timesheet
entries in 0.25 hour steps. Where a meeting could belong to two jobs,
list both options and ask me. Do not assign time to any job not in my
list of codes. Show total hours per day.

An illustrative output proposes 37.25 hours, flags two meetings with ambiguous job codes and shows each day's total. You correct the two ambiguities and submit. Time saved is perhaps 10 to 15 minutes a person a week; across eight engineers that is roughly one and a half to two hours a week at almost no risk, which is the timesheet line in the monthly table below. Microsoft 365 Copilot Business, an add-on for firms already on a Microsoft 365 business plan, lists at $21 per user per month paid yearly, with an offer price of $18 that runs to 31 December 2026. It works with your Outlook calendar inside your own Microsoft 365 tenant, and Microsoft says prompts and responses are not used to train its foundation models, which is why firms already on Microsoft 365 often start there.

Effort and risk. Low and low. The main risk is that suggested times look precise and are guesses; the engineer owns the entry.

Job 6: marketing and bid text

What firms do. They draft project case studies, capability statements, LinkedIn posts and website pages. It is the least technical use and the one most firms try first. Brief the assistant with your real project facts and ask for a plain-English description a client can follow.

A before-and-after: a case study first draft says "we delivered a world-class, state-of-the-art solution", which no client believes. Rewrite the prompt to ask for "what the problem was, what we did, what changed, in numbers you have supplied" and the output becomes "the ground-floor slab needed 18 per cent less steel after we revised the load path; programme saved nine days". Only keep a figure if the project file backs it. Check client permission before naming a client or publishing project images.

Effort and risk. Low effort; low to medium risk. The risk is unsupported claims about performance and accreditation.

A month in a 12-person consultancy, with numbers

An illustration to show how the six jobs add up. A 12-person civil and structural consultancy has eight engineers and four technical and admin staff. In a typical month it issues six tender responses, ten site or inspection reports, four spec sections, and about 60 drawing sets; every engineer fills in a weekly timesheet; the principal writes two case studies a quarter.

These are planning assumptions, not measured results, and the "with AI" figures include the time to review every output. Saving 38 hours is a little under one person-week a month. It is worth about $3,400 at $90 an hour of billable value, against a tool bill of perhaps $100 to $150. The numbers only hold while someone reviews; if reviews are skipped, the first wrong sentence in an issued report costs more than all the hours saved. Note too that specs and checking save the least, which is consistent with their higher risk rating.

Two hyped claims, put through the ladder

Here is how the evidence ladder works on claims a firm might actually receive.

Claim: "AI that automatically checks your drawings for code compliance." Rung one: can you test it on a finished set where you know the errors? Yes, so run it. Rung two: is it on a vendor page? A tool may list checks for named items such as missing sheets or tags, which is a narrower claim than code compliance. Rung three: can a person check every output? A list of flagged items can be checked; a list of "no issues" cannot, because absence of a flag proves nothing. Rung four: the worst failure is a missed error treated as cleared. Verdict: use as an extra check on the items the vendor names, never as a compliance sign-off.

Claim: "AI optimises structural layouts and cuts material by 20 per cent." Rung one fails quickly: you cannot test a design-optimisation claim on a finished job without redoing the design with a verified engine. Rung two: ask the vendor for the feature page, the input and output, and the validation method. Rung three fails for numeric outputs unless independently recalculated. Verdict: hold until the vendor shows a validated method, and treat any percentage in a brochure as a sales figure rather than evidence.

Questions to put to any engineering AI vendor

• Which exact step of my workflow does it replace, and which inputs does it need?
• What does it do when the input is incomplete: stop, flag, or guess?
• Where is my project data stored, for how long, and is it used for training?
• Can I export the results and the check log for my records?
• What happens to my work if I stop paying?

A specification test you can run on Friday afternoon

To see how the specification job behaves on your own material, run a ninety-minute test. Take one finished spec section, such as the pavement or drainage section of a recent job. Copy the design team's original notes, remove the issued text and give the assistant the notes and your clause library. Ask for a draft with these rules: use only clauses from the library; where a clause is missing, write [NEW CLAUSE NEEDED] and describe the requirement in a sentence; list every standard or test method number mentioned, in a table.

Then score it on five lines: how many clauses match your issued text in substance, how many invented clauses it added despite the instruction, how many standard numbers it cited, how many of those you can confirm on the current publication, and how long the review took. A usable result is a draft where invented clauses are rare, every cited number can be confirmed, and review takes less time than writing from scratch. If any cited standard fails the check, restrict the assistant to your library and stop it naming standards at all; you add them yourself.

A firm that runs this once learns two things that no article can tell it: whether its clause library is structured enough for an assistant to use, and whether its reviewers catch errors reliably when the text reads well. The second is the real safeguard, so measure it deliberately. Plant two wrong references in a draft and see whether the reviewer finds both; if not, the process needs a stricter checklist before the tool gets near a live job.

What not to try, and why

• Calculations that will be relied on. A language model is not a verified calculation engine. The tutorial on what is safe and what never is for structural engineers covers this at length.
• Code or standard compliance decisions. "Does this comply?" asked of a chatbot yields an answer with the confidence of a final decision and none of the accountability.
• Stamped or signed documents that no engineer has read. Responsibility stays with a named professional.
• Uploading client drawings or data to a personal account. Check the contract and use a business plan.
• Autonomous agents that email clients or submit tenders. A wrong attachment sent without review is a professional embarrassment and possibly an insurance matter.
• Vendor claims of "AI-designed" outputs. Ask which input, which output, what error rate, and which part of the job still needs your engineer.

A small-firm plan: four weeks and about $150 a month

A 12-person firm can try this without buying software for everyone. Week one: pick two engineers who write the most (proposals and reports) and put them on ChatGPT Business or Claude Team at $20 per seat per month billed annually ($25 monthly); both need two seats minimum. That is $40 to $50 a month. Week two: write the clause library and exclusion list into a Project so the assistant uses your language. Week three: run the testing steps above on two finished jobs. Week four: decide which of the six jobs to keep. If the firm is already on Microsoft 365, compare with Copilot Business on a few seats at $21.

The prices are list prices and plans change; confirm them before committing. For a fuller costing, see what small firms pay for engineering proposal software with AI.

Write a one-page policy before the first client document goes in. It needs four lines: which tools are approved, what may be pasted (no client-confidential drawings on personal accounts), who reviews each deliverable, and what is recorded when AI helped. That record is valuable when an insurer, a client or a colleague asks how a document was produced.

If you want to go through your firm's proposals, reports and admin with someone and choose the first two uses, that is what my AI implementation consultation covers.

Questions engineering principals ask about AI

Can AI do engineering calculations?

Treat it as unsafe for anything a person signs. Language models can slip on arithmetic and unit handling and will present a wrong result confidently. Use it to explain a method or draft a calculation template, then run the numbers in your own checked tools and have a qualified engineer verify them.

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

Only if the contract allows it and the tool's terms fit. Business plans such as ChatGPT Business, Claude Team and Microsoft 365 Copilot list no training on your content by default. Personal accounts differ. Check the client's confidentiality clause before uploading anything.

Which of the six jobs should a small firm start with?

Start with reports or proposals. Both are text you can check line by line against your own notes, the saving is in hours per week, and an error is caught before a client sees it. Leave QA tools and anything near a design decision until you have a review habit.

Do professional-indemnity insurers care about AI use?

Ask your broker rather than assume either way, because wordings differ between insurers and change at renewal. Tell them how you use AI, keep a record of the human check on every deliverable, and read your policy and proposal form for any exclusion or condition that mentions automated tools or software.

Further reads

• AI for Consulting Engineers: Fee Proposals and Reports, Not Calculations — Fee proposals and reports, not calculations.
• Engineering Company AI: Setup Plan and Costs for a Small Firm — Set-up plan and costs for a small firm.
• Structural Engineer Report Template: A Plain-English Summary You Can Check — A plain-English report summary you can check.
• Structural Engineer Fee Proposal: Scope and Assumptions Drafted With AI — Draft scope and exclusions in a fee proposal.
• AI for Civil Engineering Firms: Proposals and Site Reports First — Civil practice: proposals and site reports first.
• The AI Tool Stack a One-Person Consultancy Actually Needs — A lean tool stack for a very small practice.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Bluebeam pricing page, Autodesk Assistant page and FAQ, Autodesk Assistant user guide (InfoWater Pro), Microsoft 365 Copilot Business page, Microsoft Learn data-privacy page for Microsoft 365 Copilot, ChatGPT Business and claude.com pricing pages (checked 4 November 2026).