An MEP consulting engineer can use AI to draft a specification section from the design brief and then check it against a named standard, as long as the engineer supplies every design figure, pastes the relevant standard extract into the prompt, and signs off the result personally. AI writes the surrounding text and flags gaps. It does not decide capacities.
Vendor pages for MEP AI tools tend to promise a finished specification at the push of a button. In practice the useful split is narrower: the model handles structure, boilerplate and consistency, and you handle numbers and judgement. Here is the sequence I would use on a real section, with prompts, a sample output, the checks that catch the usual errors, and the sign-off record that makes the draft safe to issue.
Where AI fits in specification writing, and where it does not
A specification section does three jobs. It tells the contractor what to supply, how it must perform, and how it will be installed and tested. Most of the drafting time goes on the second and third jobs, and much of it is text your practice has written before. That repetition is where AI saves time. The first job, choosing what to supply and at what rating, is design, and it stays with you.
| Task | AI role | Engineer role |
|---|---|---|
| Section skeleton (general, products, execution) | Draft from your template and the brief | Choose the template and scope |
| Performance figures and ratings | None; ask it to insert placeholders | Take from calculations and schedules |
| Standard references | Compare draft text against the extract you paste | Confirm the edition the contract names |
| Consistency across sections | Flag mismatched terms, units and cross-references | Decide which wording wins |
| Coordination notes | Turn a meeting list into action items by trade | Confirm what was actually agreed |
| Issue and sign-off | None | Review, approve, take responsibility |
Practices differ on numbering and layout. Many use a three-part structure (general, products, execution) and a division-based numbering scheme such as MasterFormat, where HVAC sections sit in Division 23 (Heating, Ventilating and Air Conditioning); public master specifications file metal ductwork under 23 31 13 and testing, adjusting and balancing under 23 05 93. Whatever your house standard is, give the model your template. A draft that matches your format saves more time than a clever one that does not.
Set the account up before the first brief goes in
Design briefs usually carry client names, site details and commercial information, and many appointments include confidentiality clauses. Settle three things before you paste anything.
- Which account. A business plan (ChatGPT Business or Claude Team) does not train on your content by default, according to both pricing pages in November 2026, and list prices are $20 per seat a month billed annually or $25 monthly on both, with a two-seat minimum. Consumer plans differ, so verify the training setting on whichever plan you use. For a longer comparison of seat costs, see ChatGPT Business vs Claude Team for a small consultancy.
- What you strip out. Replace the client name, the site address and any fee or programme information with neutral labels before upload. A brief reads as well with "the client" and "the building".
- What your client allows. Check the appointment for clauses about third-party tools or data. If it says nothing, ask the client for a one-line written confirmation rather than assuming.
If your practice has more than a handful of engineers, put these three rules in a one-page note, as covered in what to check in an AI tool's privacy policy and terms.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep 1: turn a design brief into a section outline
Take a worked example, illustrative throughout. A six-storey office building needs a ducted air-conditioning system with variable air volume terminals. The brief lists the zones, occupancy assumptions, the standard the contract names for ventilation, the duct leakage class, and the testing and commissioning requirements. You want a first draft of the specification section for supply and return air ductwork.
Start with a prompt that asks for an outline only. This forces the model to commit to a structure you can correct before it writes any detail.
You are assisting a mechanical engineer. From the design brief extract below, produce an
outline for a specification section titled "Ductwork: supply and return air".
Use three parts: General, Products, Execution. For each part list the clause headings only.
Do not insert any numbers, ratings or dimensions. Where a value is needed, write [ENGINEER TO
CONFIRM: what is needed].
Brief extract:
[paste, with client and site names removed]A typical output (illustrative) comes back with around 20 headings: scope, related sections, reference standards, submittals, quality assurance, materials, fittings, supports, insulation interface, installation, leakage testing, commissioning support, and so on. What you would fix: remove headings that belong to other sections (insulation often sits in its own section), add the headings your practice always includes (for example a clause on access doors), and check the model has not invented a clause for a requirement your brief does not contain. The placeholders matter. A model told not to insert numbers will mostly obey, and the [ENGINEER TO CONFIRM] tags give you a ready list of inputs to chase.
Step 2: draft the clauses with your own figures in place
Now supply the figures yourself. Paste the schedule rows or calculation outputs and ask the model to write the clauses around them without altering any value. Be explicit, because models "tidy" numbers, converting units or rounding without telling you.
Draft the clauses for Part 2 Products, "Ductwork". Use only these values, exactly as written:
- Duct leakage class: [from brief]
- Maximum design velocity, main supply: [from calculation sheet]
- Sheet gauge table: [paste table]
Keep units as given. If a value you need is missing, write [ENGINEER TO CONFIRM] rather
than estimating. Use "shall" for requirements. Plain, short sentences.Illustrative output for one clause: "Ductwork shall be constructed to the leakage class stated in Clause 1.4, with all longitudinal and transverse joints sealed. Sealant shall be compatible with the duct material and rated for the operating temperature range." What you would fix: the phrase "rated for the operating temperature range" is vague and unenforceable, so replace it with the range from your design. Then search the draft for any digit the model produced that you did not supply. A simple technique is to copy the draft into a text editor, search for numbers, and tick each against your input. Any digit with no matching source is an error to remove.
Why placeholders beat estimates
Models are fluent, and fluent text looks authoritative. A made-up figure written in the confident tone of a specification reads exactly like a real one. A bracketed placeholder cannot be mistaken for anything. Every [ENGINEER TO CONFIRM] that survives into an issued document is a visible error. Every plausible invented number that survives is a hidden one, and it is the hidden one that costs money on site.
Step 3: check the draft against the standard the contract names
This is the check most vendor pages skip, and it is the one that makes the draft safe. Take a ventilation clause. The contract names a standard for ventilation, say ASHRAE Standard 62.1, which covers ventilation for acceptable indoor air quality. Editions change. ASHRAE's bookstore pages, read in November 2026, list ANSI/ASHRAE Standard 62.1-2025 and ANSI/ASHRAE/IES Standard 90.1-2025, the energy standard for sites and buildings except low-rise residential buildings. ASHRAE's own summary says the 2025 edition of 62.1 adds, among other things, mandatory adjustments for air density, demand control ventilation sequences and new calculation methods for exhaust airflow rates, which is exactly the kind of change a model answering from memory will miss. Your contract may still fix an earlier edition, and the edition in the contract is the one that counts. Confirm it in your appointment documents.
The method works the same for any standard. Do not ask the model what the standard says; it will answer from memory and may be wrong on clause numbers and limits. Instead, paste the relevant extract from your licensed copy and ask for a comparison against only that text.
Below is (A) a draft specification clause and (B) an extract from the standard named in
the contract. Compare them. Use ONLY text in (B); do not use outside knowledge.
List: 1) requirements in (B) that the draft omits; 2) statements in the draft that
conflict with (B); 3) draft statements with no basis in (B). For each item quote the exact
words from (A) and (B). If you cannot find a basis, say "not found in B".
(A) [draft clause]
(B) [extract from licensed copy]The "quote the exact words" instruction does most of the work. It turns a vague opinion into evidence you can verify in thirty seconds. Illustrative result: the model reports that the draft says outdoor air shall be "not less than the design rate" without stating that the rate must be derived using the method in the extract, and that it has no wording on which of the extract's calculation procedures applies. You then add a clause pointing to the calculation basis. It also reports a sentence about filter grade with "not found in B", which is correct, because filter selection sits elsewhere in your design.
A word on figures from the standard. The sum below uses dummy numbers chosen only to make the arithmetic easy to follow. They are not from any edition of 62.1, and the units are left deliberately generic so nobody copies them into a design. Take real rates, units and any required adjustments from the tables and equations in your own licensed copy.
Suppose your edition calculates a zone's outdoor air as a people-based rate times the number of people plus an area-based rate times the floor area (check the exact equation and any corrections in your copy). With a dummy people rate of 10 units per person, 20 people, a dummy area rate of 1 unit per square metre and 150 square metres, that is 200 + 150 = 350 units. Ask the model to repeat the sum with your real inputs, then do it yourself or in your spreadsheet. If its answer differs from yours, trust the spreadsheet and ask why. Often the cause is a unit slip, for example per-person rates written in the wrong units, or an adjustment your edition requires, such as the air-density adjustment ASHRAE lists as new in 2025, left out. Arithmetic checks like this are cheap and catch real mistakes. But note the model is checking your numbers against your numbers. The check does not prove the design is right.
Step 4: a consistency pass across sections
Specifications fail at the joins. Section 23 31 13 says "Class B sealing", section 23 05 93 says "leakage class 2", the drawings say something else. A model that reads long text without tiring is useful here. Paste two or three related sections and ask for contradictions only.
Read the three sections below. List only inconsistencies: terms for the same thing that differ,
unit mismatches, cross-references to clauses or sections that do not exist in the text
provided, and requirements stated twice with different wording. Table output: location,
the conflicting text, suggested fix. Do not rewrite the sections.Illustrative output rows: a cross-reference to "Clause 3.7" that does not exist in the pasted text; "VAV box" in one section and "variable air volume terminal" in another; a test pressure given in pascals in one place and in millibar in another. None of these is clever, and each is the type of error that prompts a contractor query or a variation. For long documents, split the check into chunks by division. Models can lose detail in very long pastes, and chunks are easier for you to verify. To see whether a given tool can read your file types at all, can ChatGPT read PDFs, spreadsheets and photos sets out what breaks.
Step 5: coordination notes from a meeting list
Coordination meetings produce long lists of clashes and agreed actions. AI can sort a raw note into action items by trade, with the owner and date, so nothing hides in paragraph six. Use your own typed notes, not a recording, and remove names of third parties before you paste.
Turn these coordination notes into an action table. Columns: item, trade responsible, location
or grid reference, action, due date. If the notes do not say who owns something, write
"owner not stated". Do not infer owners. Notes: [paste]Illustrative result: of 14 notes, 11 get a clear owner, and 3 come back "owner not stated". Those three are the ones to raise at the next meeting, and they are exactly what a human skimming the notes tends to miss. What you would fix: the model sometimes merges two clashes at one grid location into a single action. Re-read the original against the table before circulating. The table is a draft for the coordinator to confirm, not minutes.
Choosing the model and the way you use it
You do not need a specialist MEP product to start. A general assistant with a long context window and the ability to read pasted tables is enough for the five steps above. Specialist products add things such as live drawing links or template libraries, and they bring a data question: where do your clients' briefs and drawings live, and who can see them? The comparison in the best AI tool for an MEP consultancy by job sets general and specialist options side by side. A sensible trial uses one section type and one engineer, as in the four-week plan further down.
Judge a model on three things in your own work, not on a demo. First, does it keep to "use only these values"? Second, does it quote accurately from an extract when asked? Third, does it say "not found" when the answer is not there, or does it bluff? Test the third deliberately: paste an extract with a clause removed and ask about it. A model that bluffs gets restricted to drafting boilerplate only.
Model names and versions change monthly, so check the current options on the vendor pages and name the version in your internal note with the date you tested it. When a new model arrives, repeat the same three tests on the same sample rather than assuming it is better.
The sign-off record: making a human check visible
The draft is not the product. The product is a specification section you are willing to put your name to. A short record per section makes that explicit, and helps if anyone later asks how the text was produced. Here is a filled-in example for the ductwork section.
| Item | Entry |
|---|---|
| Section | Ductwork: supply and return air, draft 2 |
| AI use | First draft of clauses and a consistency pass, on a business plan with training off |
| Inputs supplied by engineer | Leakage class, velocity limits, gauge table, test pressures from calc sheet rev C |
| Standard checked | Edition named in the appointment; extract pasted from licensed copy; 3 omissions and 1 conflict found and fixed |
| Numbers audit | Every digit in the draft traced to a source; 2 invented values removed |
| Reviewed by | Named engineer, date |
Keep this with the project file. If the appointment, your professional indemnity policy or a client brief says anything about AI, follow it. If it is unclear, ask your insurer or a qualified adviser, and see whether your business insurance covers AI mistakes for the questions worth asking.
Four ways it goes wrong in a live project
- The helpful rounding. A model asked to tidy a clause converts a velocity from metres per second to feet per minute and rounds it. The unit is right; the figure is no longer the design value. The fix is the "exactly as written" instruction and the digit audit.
- The imported clause. A draft includes a paragraph on fire dampers that the brief never mentioned, lifted from typical specifications. It reads well. The test is to ask of every clause: which part of the brief or design asked for this?
- The stale edition. A reviewer asks the model what a clause requires and gets an answer from an earlier edition, with none of the 2025 changes. The fix is the extract-only prompt in step 3.
- The silent gap. The draft looks complete but omits commissioning witness requirements, because the brief mentions them in a different section. Check the draft against the brief's requirement list, not only against itself.
Each of these is caught by process, not by hoping the model improves. If you want the same discipline for the adjacent calculation work, ChatGPT for structural engineers draws the same line between safe uses and ones that never are.
What a first month looks like for a five-engineer practice
Week one: set up the account, write the one-page data note, and pick one section type that your practice writes often, such as general ductwork or low-voltage distribution. Week two: one engineer drafts it the usual way and, alongside, with the steps above, timing both. Say the usual draft takes 4 hours and the AI-assisted route takes 2 hours 30 minutes including checks (illustrative numbers). Week three: compare the two drafts and note what the AI route missed and what the manual route missed. Week four: decide, with a record, whether to extend to a second section type or stop.
The saving is not 100 per cent, and it should not be. The checks take real time, and they are the reason the draft is usable. If a trial shows no net saving after checks, that is a valid result and worth knowing before anyone buys seats.
If you would rather map this for your own practice with someone, that is what my AI implementation consultation covers, through the contact page. The AI guide lists the rest of the series.
Questions MEP engineers ask about AI-drafted specifications
Can AI quote the clauses of a standard accurately?
Not reliably. Models paraphrase from memory and can state a clause number or limit that is wrong or from an older edition. Paste the relevant extract from your licensed copy into the prompt and ask the model to compare against only that text, then verify every figure yourself.
Is it safe to upload a client's design brief to a chatbot?
Only on an account whose training setting you have verified, and only if your client's terms allow it. Remove client names, site addresses and commercial figures first. Many briefs carry confidentiality clauses, so read yours before any upload.
Should AI choose equipment sizes or ratings in the spec?
No. Capacities, ratings and design criteria come from your calculations and the design brief. Let the model draft the surrounding text and flag missing inputs. Every number in the final section should trace to a calculation, a schedule or a quoted clause.
Who is responsible if the AI-drafted spec is wrong?
The engineer who signs it off, exactly as with a junior's draft. Check your professional indemnity policy and your appointment terms for any wording about AI, and ask your insurer or a qualified adviser if the position is unclear.
Further reads
- AI for Consulting Engineers: Fee Proposals and Reports, Not Calculations — Use the same tools on fee proposals and reports.
- How to Use Claude AI in Architecture: Specs and Planning Reports — How architects handle specification writing with Claude.
- AI for Mechanical Contractors: Can It Read Specs and Drawings? — The contractor's side of reading specs and drawings.
- AI Tools for Shop Drawings: Checking Fabrication Drawings Against the Spec — Checking fabrication drawings against the spec.
- The AI Tool Stack a One-Person Consultancy Actually Needs — A lean tool stack for a small consultancy.
- 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.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ASHRAE bookstore pages for Standard 62.1-2025 and Standard 90.1-2025 (ashrae.org); a public master specification index for Division 23 section numbers; ChatGPT Business and Claude Team pricing pages, all checked November 2026.