AI Training for Construction Companies: Workshop or Self-Taught?

Pay for a workshop when eight or more office staff need shared rules and practice on real documents; site crews need only a 20-minute briefing. Self-teaching costs about four paid hours a head, often with no fee. A workshop adds its fee per head, say $260 on an illustrative $2,600 quote for ten, plus similar hours.

Smaller office groups, or one or two curious people, can teach themselves, and site supervisors need about ninety minutes with a phone rather than a course. The fee is rarely the big number. The wages for hours away from work are, and on a building site a missing supervisor can leave a crew waiting for instructions.

Follow me on Instagram@sagnikteaches

What each role on a building firm needs from AI training

Training plans go wrong when they teach everyone the same thing. A construction firm has several different relationships with AI, and the useful levels follow the work, not "beginner" and "advanced". An estimator who has never opened a chatbot needs a different session from a supervisor who has used one for a year to write texts. The table sets out what each role should practise first and, just as important, what they must learn to check.

Connect on LinkedInSagnik Bhattacharya

Two things follow from the table. First, the estimator and project manager rows are where a workshop earns its fee, because the exercises need your own addenda, RFIs and programmes, which generic courses don't have. Second, the tradesperson row is the cheapest and the most neglected, and it is where most privacy slips begin. For the wider picture of how long and how much a first round of staff training takes in any small firm, there is a separate tutorial on AI staff training time and cost; this one stays with construction roles.

Subscribe on YouTube@sagnikteaches

A cost-per-head sheet for a 24-person contractor

Here is an illustration with invented figures, so swap in your own payroll. A building contractor has 24 people: one owner, two estimators, three project managers, four site supervisors, three in the office and accounts, one health and safety lead and ten tradespeople. Loaded hourly cost means wage plus employer on-costs: $90 for the owner, $60 for an estimator, $65 for a project manager, $55 for a supervisor, $38 for office staff, $50 for the safety lead and $40 for a tradesperson.

The plan has three parts. The ten desk-based people (owner, estimators, project managers, accounts and the safety lead) attend one half-day workshop on site. The four supervisors get a 90-minute session at the yard on a rain-off day, run in-house by one of the estimators. The ten tradespeople get a 20-minute briefing inside a normal toolbox talk.

The desk group's wages are built like this: the owner 6 hours at $90 is $540, the two estimators 12 hours at $60 are $720, the three project managers 18 hours at $65 are $1,170, the three office staff 18 hours at $38 are $684 and the safety lead 6 hours at $50 is $300, which adds to $3,414. I counted the two practice hours at full cost even though people will be doing real work in them, because that is the cautious way to read a training budget.

The sheet shows something easy to miss. Even on these assumed figures, the $2,600 fee is under 38% of the total, and the biggest line is wages for hours that nobody is invoicing. That changes the question from "which course is cheapest?" to "which route gets ten people to the same skill in the fewest paid hours?".

Now the break-even sum. The ten desk staff cost $569 an hour between them ($90 plus $120 plus $195 plus $114 plus $50). Divide the fee by that and $2,600 comes to 4.6 hours. If teaching themselves would take each of those ten people about four and a half hours longer than the workshop to reach the same point, the workshop is the cheaper route. If they would get there in the same time, the workshop is an extra $2,600 spent on shared rules, your own documents and a fixed date. Neither answer is wrong. The pilot below is how you find out which one applies.

Licences are not in the sheet, because the sensible time to buy them is after the pilot. A business seat of ChatGPT or Claude lists at about $20 a user a month on annual billing and $25 monthly, with a two-seat minimum; Claude's pricing page describes its Team plan as for 2 to 150 people. Ten seats at $20 for twelve months is $2,400 a year, which is roughly the size of the workshop fee and recurs every year. Buy seats for people who prove they use them.

COST-PER-HEAD SHEET (one row per group)

Group:            ______   Heads: ___
Route:            workshop / self-taught / in-house session / briefing
Fee:              quote ÷ attendees          = $___ a head
Hours each:       session ___ + practice ___ = ___ hours
Loaded rate:      $___ an hour per role (wage + employer on-costs)
Wages for hours:  hours × rate, added up over the group = $___
Site cover:       crew idle hours × crew rate (0 if sessions are planned around the work)
Licences, year 1: seats × monthly price × 12 (decide after the pilot)
TOTAL PER HEAD:   (fee + wages + cover + licences) ÷ heads
BREAK-EVEN:       fee ÷ the group's combined hourly cost = extra hours that self-teaching can take before the workshop wins

Self-taught routes, with the hours and prices I could check

As of October 2026 these are the routes I could confirm on the providers' own pages, along with what each one leaves out for a building firm.

Two points about this table. "Free" means free of fees, not of hours: four hours of an estimator's time is $240 at the sheet's sample rate. And none of these routes will teach your firm's own rules, which accounts are approved, or how to check a quantity against your drawings. That is the gap the in-house session and the error hunt later in this tutorial fill. If you want a framework for judging providers that do sell live training, how to choose an AI training provider for your team covers the questions in more general terms.

What a workshop quote has to say before you compare it

Workshop prices are quotes, not list prices, so two numbers that look alike can mean different things. Get these answers in writing before comparing:

• Is the price per head or flat for the group, and what is the cap on attendees?
• How many hours of teaching are there, as distinct from the length of the day?
• Do they build exercises from your redacted addenda, RFIs and daily logs, or use generic examples?
• Is a prep call included, and who attends it?
• Do the exercises need a paid account, and who pays for it?
• What follow-up is included: a second session, office hours, an email address that answers?
• Do you keep the prompts, exercises and any recording?
• What happens if weather or a delivery moves your site programme on the day?

Two invented quotes show why the answers matter. Quote A is $2,600 flat for up to 12 people, half a day on site, a 30-minute prep call, your redacted documents and one follow-up call. Quote B is $290 a head, minimum eight people, three online sessions of 90 minutes each, with generic examples. B looks cheaper per head at first glance, but the maths turns on how many people turn up.

A becomes cheaper from nine attendees upwards, and it also runs for 4 hours against 4.5, on your own documents. B wins only if the room stays small. This is also why I would put the safety lead and the accounts staff into the desk group even though their jobs differ: each extra head lowers the fee per head, and the first 30 minutes on rules is the same for everyone.

What the half-day for the desk group should contain

If a provider can't deliver the outputs in the right-hand column, or you could produce them yourselves with a free course and an afternoon, you don't need the workshop. Here is a four-hour agenda to hand to anyone who quotes.

The champion matters more than it sounds. After the workshop, one estimator or project manager runs the supervisor session and keeps the shared prompts tidy. If you need a format for that session, the tutorial on running a 60-minute AI workshop for your team gives a plan you can adapt to a yard cabin.

Keeping time off site cheap

Time off site is where the sheet can go wrong, because site staff cost more to move than office staff. At the sheet's sample rates, a supervisor at a three-hour session isn't only $165 in wages; if a crew of four stands waiting for direction for those three hours, that is another 4 × $40 × 3 = $480. Cover costs can exceed the fee per head, so plan around them:

• Use rain-off days, when work has stopped anyway, or the last hour of a Friday.
• Hold supervisor sessions in the yard cabin, so travel is zero.
• Run two supervisors at a time, never all four, and leave the most experienced on site.
• Book around concrete pours, inspections and key deliveries, and ask any provider what their weather-delay terms are.
• Keep site sessions to 90 minutes and build them around a phone, since a phone is the only tool a supervisor will have on site.

The sheet above assumed no cover cost because the sessions are planned around work. If you can't do that, add the line: crew idle hours multiplied by the crew's loaded rate.

A 90-minute pilot before you spend anything

Before booking a workshop or buying licences, run one 90-minute session with three volunteers: one from estimating or project management, one site supervisor and one person from accounts. Each brings one real task from the last fortnight, with client names and prices removed, or a finished job that is no longer sensitive. Use a business-plan seat, or a personal plan with the model-training switch in privacy settings turned off.

1. Minutes 0 to 10: the rules. Which account, what never goes in, what to do after a slip.
2. Minutes 10 to 25: each person states the task and how long it took last time, from memory of a real instance.
3. Minutes 25 to 65: each person does the task with AI, keeping a note of the prompt used and anything the AI got wrong.
4. Minutes 65 to 80: swap results. Each person shows the mistake they caught.
5. Minutes 80 to 90: score the three tasks and decide.

Here is how a scorecard might look after a session. These are illustrative results.

The decision rule I would use: go ahead if at least two of the three save 15 minutes or more after checking, and every tester can say what they checked. In the table all three pass, though accounts only just. That result would argue for the desk-group workshop and against a broad licence purchase: the accounts task saved the least, so that group might start with free tools. If the result is one pass or none, stay with self-taught routes and try again in a quarter. A failed pilot costs about $230 of wages (three people for 90 minutes each at the sample rates above), which is the cheapest information in this whole sheet.

An error hunt that works for every role

The most useful 20 minutes in any session is practice at catching confident mistakes. Hand out a short source and an AI summary of it that has errors planted in, and ask people to mark every place the summary says more or less than the source. A realistic example: a prompt, the source lines and the kind of output a tool might give. The output below is illustrative.

Here is Addendum 3 to a tender. List every change that could affect
price, programme or scope, one line each. Quote the item number.
If something is unclear, say so instead of guessing.

[paste the addendum text]

The addendum says: item 2, the allowance for external paving is reduced from $20,000 to $15,000; item 5, bid close moves from 13 March to 20 March; item 6, a substitute for the specified fire-rated door sets will be considered only if approved in writing before bid close. An illustrative summary reads:

1. Item 2: external paving allowance deleted.
2. Item 5: bid close extended by one week.
3. Item 6: alternative fire-rated door sets are allowed.

Three lines, two errors. Item 2 says "reduced", not "deleted". An estimator who believed the summary would carry nothing for paving where $15,000 should still sit in the build-up. Item 6 turns a condition (written approval before bid close) into a permission. Item 5 is right. Nothing in the output flags any doubt, which is exactly why the briefing teaches people to ask for item numbers and then read them. When estimators take this further, the tutorial on whether AI can produce accurate estimates for a small building firm goes into where the numbers themselves go wrong.

A realistic slip: tender pricing in a free chatbot

This is an illustration, not a report of anything that happened. An estimator, short of time on bid day, pastes a subcontractor's priced quote and the client's name into a personal free account to tidy a comparison. Nothing visibly goes wrong, and that is the problem. The slip surfaces weeks later, if at all, as a question about where figures that were meant to be confidential have been. Deleting a chat doesn't necessarily remove the provider's copy straight away either, so tidying up afterwards is not a safe plan.

The 20-minute briefing for everyone, tradespeople included, should say this in five lines:

1. Use only the accounts the firm has approved.
2. Never paste client names under confidentiality, subcontractor prices, unreleased drawings, staff or client personal details, or photos showing faces or number plates.
3. Check what the AI writes: numbers, units, dates and names.
4. AI doesn't sign anything. Toolbox talks, method statements and risk assessments need a competent person's sign-off.
5. If you slip, tell the office manager the same day. No blame; the point is to review settings and delete what can be deleted.

Check your client contracts and your insurer's wording too, because some restrict sharing project documents with outside tools, and a short briefing can't settle that for you. For the settings side, how to stop staff pasting client data into free tools lists what to switch on and what to ban.

Deciding per role: workshop, self-taught or a short briefing

Three things justify paying for a workshop: exercises built on your own documents, a fixed date that people can't postpone, and a shared set of rules that everyone hears together. Three things don't: a certificate with a famous badge, a motivational "AI for everyone" day with no exercises, and a recorded talk you could have found for free. If a quote can't tell you which of the first three you are buying, ask again until it can, or save the money and start with the pilot.

Questions owners ask before booking AI training

Do site crews need AI training at all?

Most do not need a course. They need a 20-minute briefing, ideally inside a normal toolbox talk, covering which accounts are approved, what never goes into a chatbot (client names, subcontractor prices, unreleased drawings, photos showing faces) and who to tell after a slip. Supervisors who write daily logs benefit from about ninety minutes with a phone and a voice note.

Is a free course enough for an estimator?

It is enough to learn the tool, but not to learn your firm's rules or how to check a quantity against a drawing. Pair a free course with the error-hunt exercise: take a past addendum, ask for a change summary, and mark every place the summary says more or less than the source. Repeat it on your own documents until checking becomes routine.

How many people make a workshop worth paying for?

On a flat quote the fee per head falls as the room fills. The sheet's sample quote of $2,600, an assumption rather than a market rate, works out at $260 a head for ten people and $520 for five. A useful test is the break-even sum: divide the fee by the group's combined hourly cost, then ask whether teaching themselves would really cost each person that many extra hours.

Should I buy AI licences for everyone after the training?

No. Buy seats for people who showed real use in the pilot or the follow-up weeks, then review at renewal. Business seats of ChatGPT and Claude list at about $20 a user a month on annual billing ($25 monthly) with a two-seat minimum, so ten seats cost about $2,400 a year, which is close to the workshop fee in the example.

What should I ask a workshop provider before paying?

Ask whether the price is per head or flat, the cap on attendees, the hours of actual teaching, whether they use your redacted documents, what follow-up is included and what happens if weather moves your site programme. Ask for a sample exercise, then check that it matches a real task your estimators or supervisors do every week.

Further reads

• AI Training for Architects: What a Practice Workshop Covers and Costs — The same decision for a design practice, with workshop content and cost.
• How to Build a Staff Training Matrix With AI — Track who has been trained on what once the sessions are done.
• AI Literacy Requirements: What Your Staff Need to Know — The baseline knowledge staff need before they use AI at work.
• How Small Builders Use AI to Price Variations and Change Orders — A first real task for project managers: pricing variations.
• AI Scheduling for Small Builders: Subcontractors and Deliveries — A second task worth practising: subcontractor and delivery scheduling.
• Free vs Paid AI Tools for a Small Trades Business — Which tools are worth paying for once staff are using them.
• How to Stop AI Tools Training on Your Business Data — The privacy settings to check before anyone uploads a tender.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Procore Learning, Anthropic's Claude Academy, OpenAI Academy, the Google AI Essentials page, the Coursera Plus and Coursera for Teams pages, the AI CERTs certification page and the Claude pricing page, all read in October 2026. Wage rates, workshop quotes, pilot timings and the sample addendum are invented illustrations, not market rates.