How to Use AI in a Construction Company: Estimator, PM and Office

Give each role its own job and its own reviewer. Estimators use AI to find scope gaps and compare subcontractor quotes, never to produce quantities or rates unchecked. Project managers use it for minutes, RFI drafts and chasers. Office staff use it for coding, matching and reminders. Anything with a figure or a contract clause gets a second reader.

The three roles fail in different ways, which is why one shared prompt guide rarely works. An estimator's slip costs margin on a job you have already won. A project manager's slip costs contractual position, usually through a notice that went late or said less than the contract requires. An office slip costs cash or records: a mis-coded invoice, a lapsed insurance certificate. Set the review step by the cost of the mistake, not by the tool.

Follow me on Instagram@sagnikteaches

Who does what, and who checks it

This table is the whole playbook in miniature, and the sections after it give the prompts, sample outputs and checks for each row. The examples follow a nine-person building firm (one estimator, two project managers, a part-time office manager) doing extensions and small fit-outs. They are illustrations, so swap in your own jobs and figures.

Connect on LinkedInSagnik Bhattacharya

Settle what may be pasted before anyone starts. Quote prices, tender documents and client details are commercially sensitive, so use a business plan with training off rather than a personal account, and write the rules down once. The AI policy tutorial for construction companies covers bids, photos and subcontractors.

Subscribe on YouTube@sagnikteaches

Estimator: scope gaps, quote comparisons and bid queries

An estimator's day is mostly reading: tender documents, subcontractor quotes, addenda. That is where AI helps, because it can read three differently worded PDFs and list the differences in a minute. It is also where the temptation starts, because the same tool will happily produce a rate or a quantity that sounds plausible. The rule I would set is short: the AI reads, the estimator prices.

A worked comparison: three plastering quotes

Say the firm is pricing a two-storey rear extension and has three quotes for skim finish: A at $14,200, B at $11,900 and C at $12,650. B looks like a bargain. The estimator pastes the scope summary and the text of each quote, with the subcontractor names swapped for A, B and C, into a business-plan assistant with this prompt.

You are helping a construction estimator compare subcontractor quotes for one scope.

Scope: skim finish to new walls and ceilings in a two-storey rear extension, 186 m2 of wall and 74 m2 of ceiling. The main contractor provides power and water.

Put quotes A, B and C in one table with these rows: price, what is included, what is excluded, assumptions, conditions (access, scaffold or towers, floor protection, making good), validity period, payment terms.

Rules:
- Use only what each document says. If a row is not stated, write "not stated".
- Do not infer, do not recalculate prices and do not suggest rates.
- After the table, list every item that is included in one quote and missing or excluded in another.

[scope summary]
[quote A text]
[quote B text]
[quote C text]

The reply (illustrative) arrives as a clean table, followed by the list the estimator actually wanted:

• Quote B excludes access towers ("access equipment by others"), the mist coat on new plaster and floor protection.
• Quote C includes protection but excludes making good after electrical first fix.
• Quote A includes all of these and states a 30-day validity period.

Two things need fixing. The assistant gave Quote C a validity of 60 days, but the PDF says 30; the 60 came from the payment terms line printed next to it. And it wrote "not stated" for Quote B's payment terms, which sit on page two, a page that was never pasted. Both errors are typical when documents go in piece by piece, so the estimator reads each "not stated" against the original.

Then the estimator prices the gaps from the firm's own rate book, not from the assistant.

The cheapest quote became the dearest, by $10, and Quote C wins by $1,070 over A. The four allowances are the estimator's own numbers from past jobs; the assistant never produced a price. That separation is the point of the exercise. For the wider toolkit around this, see AI tools for construction estimating, and for using a chat assistant to check a quote rather than write one, ChatGPT for construction estimating.

Clarification lists from a tender pack

Tender documents contradict themselves more often than anyone admits, and finding the contradictions is slow, dull work. Ask the assistant to read the specification and preliminaries and list contradictions, undefined terms, "or equivalent" items, provisional sums with no quantity and anything that needs information the pack does not supply, each with its document reference. Then ask it to label every item as ask, assume or price as risk. A sample of what comes back (illustrative):

• Preliminaries say normal working hours, but specification clause 7.3 needs out-of-hours work in the ceiling void. Which governs? Label: ask.
• Item 4.2 says "allow for making good" with no extent. Label: price as risk, and state the assumption in the quote.
• A provisional sum of $8,000 for sanitaryware, with no fixture count. Label: ask, and state how many fixtures you priced.

The estimator sends the "ask" items as numbered bid queries (tender queries, in some firms) and carries the "assume" items into the quote as written assumptions. Check every reference against the document before it goes out, because the assistant will sometimes cite clause 7.3 when the clause is 7.2.

Measured quantities: count, then recount

Some estimating tools now count and measure for you. JobTread lists "AI Plan Import & Count Tools" that name, number and scale plan sheets and count repeated items. Bluebeam lists "AI drawing reviews" among the extras on its Max plan, priced at $590 per user per year and marked as an introductory price. Whichever you try, test it on a job you have already priced.

Suppose a counting tool reports 214 sockets and outlets across four floors. The estimator recounts one floor by hand and gets 52, against 47 from the tool. That is a 9.6% undercount on that floor (5 of 52). Until three more jobs agree, treat the tool's counts as plus or minus 10% and keep a simple log of tool count, hand count and the difference. If the differences settle under 3%, the tool has earned a place in the process. If they swing both ways, it stays a second opinion.

Project managers: minutes, RFIs, chasers and notices

A project manager's day is a stream of small written outputs. Someone said something on site and it has to become a record, a request or a reminder. AI is useful when the PM was in the room, because they can tell at a glance whether the draft is right. It turns risky when the draft goes out in the firm's name and carries a deadline.

Minutes that end in an action table

Record the progress meeting if everyone has agreed to it, let the meeting tool transcribe it, and paste the transcript, not the audio, into the assistant.

Here is the transcript of a progress meeting on a two-storey extension. Write minutes in this order: decisions, then an action table (action, owner, due date, source line in the transcript), then open questions.

Rules: only record actions that someone accepted out loud. If an owner or date was not stated, write "TBC" and do not guess. Keep names as spoken; if unsure, write the name as heard and add "(check)".

[transcript]

A realistic output is right on the first four actions and wrong on the fifth. It assigned "confirm brick delivery date" to the site manager, when the supplier's representative had said he would ring the PM. The "source line" column is the fix: the PM jumps to that line of the transcript in ten seconds and corrects the owner. In this illustration the PM spends ten minutes checking instead of thirty typing, and the minutes go out the same afternoon.

RFIs that cite the drawing

A request for information (RFI) is only useful if the designer can answer it without ringing you. Give the assistant the PM's rough site note and the drawing references, and ask for one question, written to be answerable.

Draft a request for information to the architect from this site note. One question only. Cite the drawing numbers I give you, state what is unclear, state what we will assume if no answer arrives by the response date, and say what the question is holding up.

Do not invent drawing numbers, dimensions or product names. Where I have not supplied one, write [drawing ref].

Site note: [note]
Drawings: A-201 rev C, S-104 rev B
Response needed by: [date]

The draft comes back tidy with one flaw: it cites drawing "A-210", which does not exist (the PM supplied A-201). That is why the "do not invent" line is in the prompt and why a person reads every reference. A wrong drawing number on an RFI turns a simple question into a back-and-forth. If you keep the RFI log in software, check which plan carries it: Fieldwire's pricing page lists RFIs, submittals and change orders on Business Plus only ($89 per user per month, billed annually). Firms on a lower plan often keep the log in a spreadsheet and use the assistant just for the wording. More on this in AI tools for construction project management.

Chasers that stay specific

Most chasers fail because they ask nothing: "Any update on the windows?" Ask the assistant for one question, one date and one consequence, and to leave the tone alone. Before and after:

• Before: "Hi, any update on the windows?"
• After: "Hi [first name], the window delivery is in the programme for 14 March and the fix-out crew starts on the 18th. Can you confirm by Thursday that the date still holds? If it has moved, please send the new date so I can reset the crew."

The PM reads it, checks the dates against the programme and sends it. It is the lowest-risk task in this section, which makes it a good first one for a PM who is new to the tool.

Notices and variations: draft, never send

This is where AI can quietly damage a position. Contracts carry notice rules: a number of days, a written form, a named recipient. Use AI to explain a clause, never to replace it. Take a clause that reads: "The Contractor shall give written notice to the client's representative within 7 days of becoming aware of any event likely to cause delay." A plausible summary says: "Tell the client promptly about anything that may cause delay." It has dropped the 7 days, the writing and the recipient, which are the three things that decide whether the notice counts.

So the PM keeps a one-page notice card per contract, typed from the contract itself: clause number, trigger, deadline, form, recipient. The assistant can draft the cover wording around that card, and the commercial manager reads the draft against the clause before it leaves. Anything disputed, or any contract with unusual terms, goes to a lawyer or other qualified adviser.

Office staff: coding, matching and reminders

Office work has the most repeatable tasks and the least forgiving errors, because the output is money or a record someone may later be asked to produce. The safe pattern is the same each time: AI proposes in a spreadsheet, a person posts.

Suggesting cost codes for supplier invoices

Give the assistant your list of cost codes (not bank details, not supplier account numbers) and 20 invoice descriptions. Ask for a suggested code, a confidence of high, medium or low, and a one-line reason. In an illustration, 17 of the 20 come back right. The three misses are the same kind of error: a waste-container hire charge posted as materials. The bookkeeper adds a rule to the instruction ("waste-container hire is plant and waste, code 540") and re-runs, which clears two of the three. Every "low" and a sample of the "high" rows are still checked before posting. The accounting side, including job costing, is covered in AI for a small construction company's financials and QuickBooks.

Matching a supplier statement to the ledger

Month-end statements rarely agree with your ledger, and the gap is usually two or three items. Paste the statement lines and the ledger lines for one supplier into a spreadsheet, then ask the assistant to list every line that appears in one and not the other. On a 42-line statement the illustration looks like this:

The assistant finds all three in under a minute. The office manager still checks that the three amounts add up to the gap with a spreadsheet SUM, because an assistant's arithmetic is the weakest part of the job. If the sum does not close, there is a fourth difference that nobody has found yet.

Certificate and licence expiry lists

Give the assistant a list of subcontractors with their insurance and licence dates, and ask for a table sorted by expiry, plus a short reminder email for each one expiring within 30 days. The trap is dates. A certificate showing 03/04/2027 can be read as 3 April or 4 March, and the assistant picks one without telling you. State the format of your list (day first or month first), and ask for dates in the output written as "3 April 2027" so a swap is obvious. A swap in the cautious direction costs a needless phone call. A swap in the other direction hides a lapse until a site audit finds it.

Where figures go wrong, and the three checks that catch them

Every role in this tutorial handles numbers, and numbers are where an assistant's fluent writing hides its errors. These are the failures worth designing checks around:

• Units. Square metres against square feet, millimetres against inches, cubic metres against yards. State units in the prompt and ask the assistant to repeat them in every table header.
• Totals that do not add. The assistant writes a total that looks right and is off by a line item. Recompute every total in a spreadsheet; never paste an assistant's sum into a quote.
• "Typical" rates. Ask for a rate and you get an industry-average figure with no source, which is worse than a blank. Rates come from your rate book or a supplier's price list.
• Tax inclusive or exclusive. A quote's "$12,650" may or may not include tax. If the document does not say, the assistant should write "not stated", and the estimator should find out.
• Working days against calendar days. The same sentence means different deadlines. Have the assistant quote the original wording next to its summary.
• Carry-over between jobs. In a long chat, a figure from the previous job reappears in the next table. Start a fresh chat for each job.

The three checks are cheap. First, every number in the output must trace to a line in a document you supplied. Second, recompute every total outside the assistant. Third, spot-check three random lines against the original. If any of the three fails, check the whole output, not just the line that failed.

Where contract wording goes wrong

I am not giving legal advice here, and you should not take it from an assistant either. What I can tell you is the pattern. When an assistant paraphrases a clause, it tends to change it in one of four ways:

• It softens "shall" to "should" or "may". An obligation becomes a suggestion.
• It drops the exception. The "unless" or "except where" clause often disappears, and that clause is frequently the one that matters.
• It invents a clause number or a standard-form reference. It writes "clause 12.4" with total confidence when the clause is 12.1.
• It merges two clauses with different deadlines into one tidy sentence with one deadline.

The rule that follows: in anything formal, quote the clause, do not paraphrase it, and have a named person read the clause itself, not only the summary. Use the assistant to spot what to read, not to read it for you. If you are weighing software built for contract review, what builders should test before buying it is a good place to start.

Review gates: who reads what, and for how long

A review step nobody has time for will be skipped, so give each one a name and a time budget. These are planning figures from the examples above, not measured results; adjust them after your first fortnight.

Add a line to anything that leaves the building after AI help, so the check is on the record:

AI-assisted draft. Figures and references checked by [name], [date].

A two-week trial for each role

Do not switch all three roles on in the same week. Run each role for two weeks: one week doing the work as you do now while timing it, one week with the AI step and the review gate. The trial tells you whether the review time eats the saving.

Written down like this, the decision at the end of week two takes five minutes instead of a debate about whether the tool "feels" useful. If you would like an outside view on where to start, my AI implementation consultation covers exactly this sort of first-task selection. Otherwise, begin with the project managers: it is the role where the person checking the output also wrote the input.

Questions owners ask before assigning AI tasks by role

Which role should start using AI first?

Start where the person who supplied the input also checks the output. A project manager turning a meeting they attended into minutes is the safest first task, because a wrong action jumps out. Estimators comparing quotes they have already read come second. Office tasks that touch money come last, with a person posting every entry.

Is it safe to paste subcontractor quotes into an AI assistant?

Treat quote prices as commercially sensitive. Use a business plan where your content is not used for model training by default, replace subcontractor names with A, B and C, and leave out bank details and client personal data. Personal and free accounts follow different data rules, so check the settings and write the rule into your AI policy.

Does AI replace our takeoff or estimating software?

No. Takeoff and estimating packages measure drawings and hold your cost data, while a general assistant reads and drafts text. Some packages now add AI features, such as JobTread's plan import and count tools, and those need the same recount check as any measured quantity. Test any of them on a job you have already priced.

Can one shared AI login cover the whole company?

It is better not to. Shared logins mix chat histories, hide who pasted what and clash with per-seat business plans. Claude's Team plan, for example, is for teams of 2 to 150, with a Standard seat at $25 a month or $20 a month on annual billing. Give seats to the three roles first and widen later.

Further reads

• Construction Meeting Minutes: Record, Summarise and Chase Actions With AI — Recording, summarising and chasing actions in more depth.
• How Can AI Help Construction Project Managers? Submittals and Sub Chasing — More on submittals and subcontractor chasing for PMs.
• AI Automation for Construction Companies: Daily Logs to Invoices — Turn daily logs into summaries and invoices.
• Construction Payment Notice Template: Draft With AI, Check the Deadlines — Draft payment notices and check the deadlines properly.
• ChatGPT for Construction: Eight Office Jobs With Worked Examples — Eight office jobs with worked examples.
• How Can AI Help Construction Companies? Audit Your Week in Ten Minutes — Audit your week to choose each role's first task.
• AI Receptionist for Construction Companies: Real Costs and Call Routing — If missed calls are costing you tenders, start here.
• AI Estimator for Contractors or a Human Estimator? How to Choose — A cost sum, a volume-and-complexity grid and a hybrid workflow for choosing between an AI estimating tool and a part-time estimator, with a named sign-off.
• AI Solutions for Construction Companies: One Platform or Point Tools? — One construction platform or several point tools? Four tests (jobs, integrations, data ownership, upkeep) and two stacks priced from vendor pages.
• Benefits of AI in Construction: What a Small Firm Can Actually Measure — Five benefits of AI a small building firm can track with a before-and-after sheet: quote time, site diaries, payment chasing, scope gaps and safety drafts.
• How Can AI Help My Construction Business? Five First Jobs to Try — Five first jobs for an owner-run construction firm, ranked by effort and payback, each with a starter prompt and a six-person worked example.
• How Is AI Used in Construction? Site Jobs and Office Jobs Compared — A small firm's site-versus-office split for AI in construction, with a worked week and a rule for which jobs to try first.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: JobTread's AI features page, Bluebeam's pricing page, Fieldwire's pricing page and Claude's pricing page, all read in November 2026. Worked examples are illustrations, not client results.