AI for Consulting Engineers: Fee Proposals and Reports, Not Calculations

A small consulting engineering firm can safely use AI for fee proposals, report wording, meeting notes and routine emails, provided a named person checks every figure and the chartered engineer signs off. Design calculations, load assumptions, member sizes and code-compliance judgements should never be produced, checked or approved by an AI tool.

The line does not run between technical and non-technical work. It runs between text that describes what an engineer has already decided, and anything that produces a number or a judgement that a structure, a system or a client relies on. A fee proposal describes your scope; a calculation is the scope. AI can help with the description and has to stay out of the thing described.

Follow me on Instagram@sagnikteaches

Green, amber and red: one rule for every job in the office

Most jobs sort themselves. For the awkward ones, five tests decide the tier, and the strictest result wins:

Connect on LinkedInSagnik Bhattacharya

1. What will the reader rely on? If a client, contractor or regulator will act on the text, it is amber at least.
2. Does a number pass through? Any figure in AI-touched text must trace to a source document or an engineer's calculation. If it cannot, it comes out.
3. What happens if it is wrong? If the failure mode is injury, collapse, flooding or fire, the job is red.
4. Whose information is it? Confidential client or third-party material goes only into a plan that does not train on your content by default, and only where the client's terms allow it.
5. Is checking faster than writing? If verifying a draft would take more than about half the time of writing it yourself, write it yourself. That threshold is a rule of thumb, not a standard, so adjust it once your trial gives you real figures.

Stripping a job down before it goes into a prompt

The fourth test, whose information it is, has a practical form: anonymise first. Replace the client name, the site address, drawing numbers and any figure under a confidentiality agreement with neutral codes, and keep the key in the job file, not in the chat. Whatever the model needs in order to do the job, such as "two-storey extension" or "load-bearing wall", stays.

Subscribe on YouTube@sagnikteaches

Before: Mr and Mrs [surname] at 14 [street] want the rear wall opened up; drawings ABC-204 rev C from [architect's firm]; budget $85,000.
After: Client A, site 1. Rear wall of existing house to be opened up; architect's drawings (drawing set X, latest revision); budget not shared.

It costs a minute and removes most of the exposure if a chat history is ever shared, exported or read by someone else. The model rarely needs the names. It needs the shape of the job.

Fee proposals: scope, assumptions and exclusions from your notes

A fee proposal is a short contract, so the AI's job is clarity, not creativity. It turns shorthand into complete sentences and, more usefully, asks about assumptions your notes never stated. The fee itself stays yours: hours, rates and the price come from your own records and a spreadsheet, never from a model.

The worked example is a six-person practice pricing structural design for a two-storey rear extension, with a new steel beam over an opening in an existing load-bearing wall. The build-up below is illustrative:

The practice rounds that to a lump sum of $6,700, with variations charged at the hourly rates shown. Every figure in the table comes from the firm's records of similar jobs, and the AI never changes one. Add the column in a spreadsheet; chatbots slip on tables of figures, and a fee total that is wrong by a hundred dollars in the client's favour is still your loss.

What the AI does get is the wording. Give it your notes and forbid it from adding anything:

From my notes below, draft the scope of services, assumptions and exclusions sections of a fee proposal in plain English. Use only what is in my notes. After the draft, list any questions where my notes are silent. Do not add services, dates, fees or promises.
Notes: two-storey rear extension; new steel beam over opening in existing load-bearing wall. Client supplies architect's drawings as PDF. Scope: calculations for beam, padstones and foundations; one set of structural drawings; one round of builder queries; two site visits. Assume: ground as neighbouring works, no ground investigation; existing wall construction confirmed at first site visit; architect's drawings do not change after issue. Not included: temporary works design; ground investigation; fire engineering; approval submission fees; changes after drawings issued. Fee: lump sum $6,700, variations at hourly rates; payment 30 days.

A typical result (illustrative):

Scope of services
We will carry out structural calculations for the new steel beam, its padstones and the supporting foundations; issue one set of structural drawings; respond to one round of builder queries; and make two site visits.

Assumptions
- Ground conditions are similar to those found on neighbouring works; no ground investigation is included.
- The construction of the existing wall will be confirmed at the first site visit.
- The architect's drawings will not change after they are issued to us.

Exclusions
- Temporary works design
- Ground investigation
- Fire engineering
- Approval submission fees
- Changes after the drawings are issued
- Liaison with the approving authority

Questions
1. Who arranges access for the site visits?
2. Does "one round of builder queries" include attending site?
3. In what format will the architect's drawings arrive, and who confirms they are final?

One line needs attention and one part deserves credit. The last exclusion, liaison with the approving authority, was not in the notes: the AI added it. Here it takes work out of scope rather than putting work in, but the same slip in the other direction, an inclusion nobody priced, would give a service away for free. The questions are the best part of the output. Access, what counts as a round of queries and who declares the drawings final are the three points that turn into disagreements later, so answer them in the proposal itself.

Read the draft against the notes line by line: every sentence traces to a note or it is cut or turned into a question. For a one-page proposal that takes about ten minutes, and the checker should not be the person who wrote the notes if you can avoid it. The structural version of this job, with more exclusions, is in the structural engineer fee proposal guide, and catching made-up figures in AI-drafted proposals covers the checking habit in general.

Holding the scope line when a client asks for more

The exclusions list earns its keep three weeks later, when the client emails to ask whether the beam design can cover temporary works too. It is a short email, but one that restates what the proposal promised, so treat it as green with an amber check. Paste the relevant lines of the proposal, the client's request and one instruction about tone.

Before (the reply you would send in a hurry): "Temporary works weren't in our scope. We can look at it but it will cost extra."
After (AI draft, then edited): "Hi [first name], thanks for asking. Temporary works design is listed as an exclusion in section 4 of our proposal of [date], so it is not covered by the lump sum. We can quote it as a variation at the hourly rates in the proposal, and I can send an estimate of hours by [day] if you would like one. Please let me know before work starts on site so we can fit it into the programme."

Check three things before sending. The section number and date must be right, because the AI will guess them if you do not supply them. No new commitment should have crept in, such as a deadline you have not agreed with yourself. And the tone should suit that client. The AI added the useful sentence about timing; you decide whether you can honour "by [day]".

Report wording: change the sentences, never the findings

Reports are where AI helps most and quietly does the most damage, because its habit is to make writing sound more certain than the engineer was. The rule is simple: the AI rewrites sentences, and the engineer owns the findings, the numbers and the hedges.

Turn these site notes into one paragraph of a report. Keep every measurement and every qualifier exactly as written (approx, looks, could not see). Do not add findings, causes or recommendations. Flag any sentence you were tempted to strengthen.
Notes: rear wall, north face: diagonal crack approx 3 mm wide, starts at left end of lintel, rises 2 courses. Dirt in crack, looks old, not certain. Tell-tales fitted 12 weeks ago, both read 0.0 mm today. No cracking seen internally at same location (wardrobe in front, could not see all of wall).

A typical first draft (illustrative):

A historic diagonal crack of 3 mm was identified above the left-hand end of the lintel on the north face of the rear wall. Tell-tales installed 12 weeks ago show no movement. No internal cracking was observed at this location.

Three things drifted. "Looks old, not certain" became "historic", "approx 3 mm" became "3 mm", and "could not see all of wall" vanished, which turns a partial observation into a clean bill of health. Each shift is small; together they change what a reader may rely on. The corrected paragraph restores them:

A diagonal crack approximately 3 mm wide runs from the left-hand end of the lintel up two courses on the north face of the rear wall. Dirt within the crack suggests it is not recent, but that cannot be confirmed from a visual inspection. Two tell-tales fitted 12 weeks ago both read 0.0 mm. No cracking was seen internally at this location, although a wardrobe prevented a full view of the wall.

The instruction to flag sentences it was tempted to strengthen will sometimes make the model name its own slips, but do not rely on that. The reliable check is mechanical: lay the notes beside the draft, tick every measurement and every qualifier, and have the checker repeat the ticks. Recommendations are a separate matter. Treat any recommendation an AI offers as unwritten until an engineer has formed the opinion independently. A plain-English summary for the client is fair game once the report is final, provided the checker confirms it promises nothing the report does not; the structural report template with a plain summary you can check shows one layout.

Meeting notes: a draft in minutes, a record only once you issue it

Design-team calls produce decisions, actions and half-decisions, and a summary tends to flatten the half-decisions into firm ones. That is the failure to look for. Ask the AI to sort every item into decided, provisional or discussed only, and to quote the exact words behind any number. Then compare what was said with what the draft records:

Both first drafts are tidy and both are wrong in the way that matters: the first turns a conditional into a decision that someone will later cite, and the second drops the dependency between the two items. The process that catches this is short:

1. Ask everyone present before you switch on a recorder or note-taker, and say where the transcript will be kept.
2. Draft the same day, with the decided, provisional and discussed-only split in the prompt.
3. The engineer who attended reads the draft within 24 hours against their own notes: names against actions, numbers against what was said, every "subject to" still there.
4. Issue it as the engineer's record, and file or delete the transcript under your retention policy.

A tool's summary is not the record until a person has issued it. Many meeting tools offer a summary function, on some plans only, so check what yours includes before adding another subscription, and keep meetings that discuss confidential client information to a plan that does not train on your content by default.

Why the easy sum is the dangerous one

A chatbot will get the textbook bending moment right. For a simply supported span of 4.0 m carrying 12 kN/m, the formula wL²÷8 gives 12 × 4² ÷ 8 = 24 kNm, and any model will say so. That is exactly the danger: it is right about the easy part and sounds equally sure about the parts that depend on the project and on the edition of the standard, such as which load combination applies, which partial factors, which deflection limit and which clause. Four failure modes follow from that. They are the reason calculations are red, and the reason a check by chatbot is not a check.

• Plausible wrong inputs. Load values, factors and limits quoted from memory, sometimes from a superseded edition or another jurisdiction's standard, with no warning that they differ.
• Unit and scale slips. kN against kN/m, millimetres against metres, characteristic against design values, each one plausible in the sentence around it.
• Agreeable checking. Paste a calculation and ask whether it is right, and chatbots often find it reasonable, or invent an objection to seem useful. Neither is a check. A check is a competent, independent person working under your quality system.
• Unverified software. If an AI helps write a spreadsheet formula or script that touches design numbers, you have a new calculation tool nobody has validated.

The last one is the grey area, so give it a rule. The engineer still chooses the method and the inputs. Test the finished tool against hand calculations at three points, the lowest, a typical and the highest case you will use, keep those cases with the tool, and repeat them after every edit. Three cases is a suggestion, not a standard; your quality system may demand more. The red list itself has a longer treatment in ChatGPT for structural engineers: what is safe and what never is.

A one-page AI use note that makes sign-off real

A policy in a folder checks nothing. A ten-line note completed at the time, by the person doing the checking, does. Here is one filled in for the extension fee proposal above:

AI USE NOTE
Job: 24-117, rear extension, new beam
Task: draft scope, assumptions and exclusions for fee proposal
Tool and plan: [business plan name]; no training on content by default
Information shared: anonymised notes; no client name, address or drawings
Output used in: fee proposal, version 1
Figures in AI text: none (fee table prepared in spreadsheet)
Checked by: [name], line by line against the notes; one added exclusion removed, three questions answered
Calculation content: none
Signed off by: [responsible engineer]   Date: [date]

The line that matters is "Calculation content: none". If it ever reads anything else, the note is a flag, and the job goes back to the red list. For the wider policy wording, an AI acceptable use policy for a small professional firm gives clauses you can adapt around this note.

Three practices, three starting points

The same rule gives different first steps depending on what a firm writes most. Every time-saving figure below is an illustration, to be replaced by your trial numbers.

• A six-person structural practice writes about eight fee proposals a month, each taking the partner around 50 minutes. It starts there: a draft from notes in 15 minutes plus 10 minutes of checking, which saves about 25 minutes a proposal, or 3 hours a month. Its ban list is short and absolute: nothing from the calculation folder goes into a prompt.
• A three-person drainage and civils consultancy holds two developer calls a week and spends about 20 minutes writing up each. It starts with minutes, using the decided, provisional and discussed-only split, and with site-visit report wording. Drafting takes about 3 minutes and checking about 9, so each call costs about 12 minutes instead of 20, a saving of roughly 16 minutes a week. Flow rates, pipe sizes and levels never go into a prompt.
• A four-person building-services consultancy starts with logging, not saving time: for four weeks it records every AI use on the note above, and only then chooses which job to keep. Load, sizing and compliance statements are red from day one.

Four weeks, two jobs a week, and a stop rule written first

Run the trial on two jobs a week, one proposal and one set of minutes or one report, for four weeks. Record the same five things each time, so the decision at the end rests on your numbers and not on how impressive the drafts looked.

The stop rule has two parts. Keep a job type only if drafting plus checking takes at least a quarter less time than writing it by hand. And any red-tier incident stops the trial for that job until you know how it happened and have written the fix into the note. The error tags do a third job: a type of error that shows up in three drafts running belongs in the prompt as an instruction. A quarter is a starting threshold. If your jobs are short, the overhead of checking can swallow the saving, and that is a legitimate result.

Questions for your insurer and institution before the first prompt

Some of the rules on your work are not set by you. Ask these in writing and file the answers with your policy:

• Does the professional indemnity (professional liability) wording mention AI, automated tools or software-assisted work, and is any use excluded?
• Must AI use be declared at renewal, or when a claim is notified?
• What does your institution's code of conduct, or any AI guidance it has published, say about responsibility for tool output? The usual principle is that the engineer remains responsible for work issued in their name, but read your own institution's words.
• Do appointment terms or confidentiality agreements restrict sharing project information with outside tools, or require the client's consent?
• Do any of your clients have their own AI policies that apply to your work for them?

The answers decide whether AI is mentioned in your terms of engagement, which how to mention AI use in client contracts and proposals covers, and a qualified adviser should review the final wording. If you would rather build the green, amber and red lists for your own practice with someone, that is what my AI implementation consultation covers.

Questions engineers ask before using AI

Can AI act as a second checker on my calculations?

No. A checker has to be a competent person, independent of the originator, working under your quality system. A chatbot tends to find a plausible calculation reasonable, cannot see the inputs behind it and cannot be held to a standard, so its comment can give false comfort. Use it to explain a concept to a junior, not to check a number.

Which AI plan suits confidential client work?

Prefer a business plan to a personal one. As of October 2026, ChatGPT Business and Claude Team both list no training on your content by default, and both start at $20 a seat a month billed annually ($25 monthly) with a two-seat minimum. Confirm on the vendor pages, and check your clients' terms before any project material goes in.

Do I have to tell clients I use AI?

That depends on your appointment terms, any confidentiality agreement and the law that applies to you. Many clients want to know, and some restrict sharing project information with outside tools. Read your appointment documents, decide what you will say, and ask a qualified adviser to review the wording you put in proposals and terms of engagement.

Can a junior use a chatbot to learn how a calculation works?

As an explainer, yes, with care. Treat the explanation as a lead to follow up in the standard or a textbook, never as the source, and never copy a chatbot's worked example into a project file. The junior's own calculations still go to a senior for checking, exactly as before, whatever a chatbot said about them.

Further reads

• How Engineering Firms Are Using AI: Six Jobs and Their Limits — Six jobs engineering firms hand to AI, with the limits on each.
• AI for Civil Engineering Firms: Proposals and Site Reports First — The same proposal and site-report approach for civil work.
• AI for MEP Engineers: Specification First Drafts and Checks — Specification first drafts and the checks they need.
• Best AI Tool for Construction Design: What It Drafts and What Needs a Stamp — What AI can draft in design work and what needs a stamp.
• AI for Electrical Code: Can ChatGPT Answer Code Questions Safely? — Why edition and jurisdiction errors make code answers risky.
• How Architects Use AI for Fee Proposals, Briefs and Paperwork — How a neighbouring profession handles proposals and briefs.
• Engineering Company AI: Setup Plan and Costs for a Small Firm — Set-up plan and running costs for a small engineering firm.
• AI in Land Surveying: Proposals, Field Notes and Client Updates — A land surveying firm's admin workflow: AI drafts the proposal, the field log and the client update, while the surveyor owns every figure and the sign-off.
• Design-Build Proposal Example: A Two-Stage Fee Letter With AI — See a two-stage design-build fee letter with numbers: a stage-one fee built from hours, a stage-two guaranteed maximum price, and AI-drafted wording.
• Structural Engineer Inspection Report: Photos and Notes to a Draft With AI — A photo-to-report workflow for structural engineers: numbered photos, a register, AI descriptions of what is visible, and checks before the engineer signs.
• 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.
• 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.
• Best AI for Engineering Firms: Proposals, Specifications, Minutes — How a small engineering firm can reuse past bids in proposals, draft specification first drafts and issue minutes faster, with no AI in stamped calculations.
• ChatGPT for Civil Engineers: Reports, Method Statements and Emails — A civil engineer's guide to ChatGPT for reports, method statements and emails, with data-leak safeguards for project files and checks for each output.
• Engineering RFP Responses With AI: Win Themes and Compliance Matrix — A practical engineering bid workflow: extract requirements, choose provable win themes, draft from approved evidence and check every answer.
• AI for Engineering Design: What a Small Consultancy Can Automate — A practical division of design tasks, from drawing checks and specifications to calculations that still require an engineer's judgement.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business and Claude Team pricing pages, checked October 2026. The checking and sign-off approach is general practice, not legal or insurance advice; confirm requirements with your institution, insurer and clients.