The drift I see on most subcontractor tender submissions is 10-20%. A few swing by way more than most would care to admit in print.
‘Give 10 estimators the same job to price, you will receive 10 different estimates’
This is very true.
I'm a Senior QS. Fifteen years in.
I should know better. So should most of us.Turns out we don't.
WHY THIS MATTERS
Take a 15% drift on a £500K tender. That's £75,00.Do that five times a year and you're leaking £375,000. And the money is only the surface cost.The real damage is the next tender. You've had two close calls recently. Now you're second-guessing every number. You pad. You lose. You get frustrated. You pad less next time. Estimating is often a mental game.It's a doom loop. And it starts with not knowing how wrong you actually are.
WHAT CONTRACTORS USUALLY DO
1. The 2015 spreadsheet. Still pricing off a template you stole a decade ago. It's fine, until material costs shift, labour rates move, cell cals stop working, or the job is anything unlike the last one.2. Gut feel. Works when the job is familiar. Falls apart the moment you step outside your wheelhouse. You don't know what you don't know.3. Generic ChatGPT. "Help me price this job." Vague role. No structure. Whatever the AI spits out, you accept. No sceptical pass. No flagged assumptions. Just a number.
Estimating is about stress-testing the number before it goes out.
The 2015 spreadsheet breaks itself. Gut feel can't argue back. Generic ChatGPT agrees with everything.
Build a number and then try to break it. The 2015 spreadsheet can't break itself. Gut feel can't argue back. Generic ChatGPT agrees with everything.You need a system that pushes back.
AI ESTIMATING 101
I call it The 5-Prompt Chain. Five prompts, run in order, on every estimate I care about. Each one stress-tests the one before it.I aim for 90% accuracy.I don't always hit 90%. But the chain catches things that used to slip through.These are templates. Fill in the brackets with YOUR business - your project types, your region, your qualifications, your target margins. The prompts only work when the AI knows who you actually are.
PROMPT 1- THE ROLE LOCK (EXPERT FRAME)
Sets the register. Without it, AI defaults to a hedging, generic assistant.
You are a Senior Estimor with 30 years of UK construction experience across [project types — e.g. commercial fit-out, refurbishment, new build, residential]. You are [qualifications — e.g. NRM2-trained, degree-qualified, RICS], and have priced everything from [smallest typical project — e.g. small works under £50k] to [largest typical project — e.g. Cat A office fit-outs over £2m].
You understand construction technology, sequencing, and the commercial realities of [your role — e.g. main contractor, subcontractor, consultant].
You will produce an estimate in [output structure — e.g. NRM2-aligned element breakdown, CAWS, trade-by-trade] format. You will always challenge assumptions, push back on thin scope, and flag anything that could blow margin. You will treat UK market rates as a benchmark and use historical rates I provide as the primary source when available.
You do not hedge. You do not agree to be agreeable. If something is wrong, say so. If scope is unclear, demand clarity before you price.
Fill in YOUR specifics in the brackets. The AI isn't useful until it knows who it's pretending to be.
PROMPT 2 - THE SCOPE
The biggest source of bad estimates is incomplete information. This prompt demands clarity before numbers.
Before any pricing, review every document I've uploaded — drawings, specification, and any pricing document. Produce three outputs:
1. Scope Gaps. Every element, trade, or item that is missing, ambiguous, or inadequately specified. Name each one and state what question I need to ask [recipient — e.g. the client, the employer's agent, the design team] to close it.
2. Client Expectation Risks. Anything in the spec or drawings that suggests [the client / employer] expects more than what is written. Examples: [example 1 — e.g. 'high-end finish' without spec reference], [example 2 — e.g. premium brands implied in drawings but not specified], [example 3 — e.g. quality standards referenced that conflict with the scope]. Flag each one.
3. Subcontractor Pricing Risk. Every package where a subbie could misinterpret scope and price the wrong thing. Name the package, the ambiguous item, and the exact wording I need in my enquiry to remove doubt.
Do not produce any numbers yet. Do not invent assumptions to fill gaps. If the information isn't there, I want it flagged, not guessed.
Three separate outputs, three separate risk categories. This is where senior QSs earn their money — catching what the drawings don't say.
PROMPT 3 - THE TAKE OFF QA PASS
Your takeoff goes in. Its job is to review it with a sceptical eye.
Produce a full [structure — e.g. NRM2-aligned, CAWS, trade-by-trade] takeoff from the drawings and spec. Then do a second pass as a sceptical reviewer.
For every trade package, output:
- The measured quantity and unit
- The source (drawing reference or spec clause)
- Any quantity that looks wrong against the stated dimensions — challenge it and show your working
- Every standard item for this type of work that is NOT in the takeoff — be specific (e.g. [example 1 — 'no allowance for protection to existing finishes'], [example 2 — 'no skip hire'], [example 3 — 'no out-of-hours premium for occupied floor'])
- Every item where supply-only vs supply-and-fix is ambiguous — flag and state the commercial impact of getting it wrong
- Buildability risks — sequencing, access, existing conditions, or coordination issues that affect the quantity or cost
Challenge me. If my implied scope contains buildability issues, say so. Do not give me a clean takeoff if it isn't one.
The AI stops being your assistant and becomes your sceptical reviewer. Harder to work with. Better output.
PROMPT 4 - RATE SENSE CHECK
Rates drift. A unit rate you've used for months ago might be 7% under today's market. This prompt catches it and flag for your review.
For every rate in the estimate, produce a four-column comparison:
1. My historical rate (if I've provided one — use as primary)
2. UK 2026 market benchmark for this trade, project type, and region ([e.g. Plumbing rates for commercial office fit out, London, South East])
3. Flag: LOW / TYPICAL / HIGH
4. What to be suspicious of
For LOW rates — state what's likely missing. Common causes: [cause 1 — e.g. waste factor too low], [cause 2 — e.g. wrong material spec assumed], [cause 3 — e.g. access or prelim allowance omitted], [cause 4 — e.g. subcontractor base rate without main contractor mark-up].
For HIGH rates — state the likely reason (e.g. unfamiliar scope driving caution, outdated rate, specialist trade premium) and suggest whether to hold, challenge, or find alternative supply.
Always push back. If a rate looks wrong, say so with a reason — not a hedge. Where my historical data is older than [X] months, apply a market adjustment and show the calculation.
Your local rates might differ from market. Your historical rates might be stale. This pass reconciles both.
PROMPT 5 - MARGIN PROTECTOR
The final pass. The one most contractors skip. The one that decides whether you submit with eyes open.
Produce a formal risk register for this estimate. List the top [number — e.g. 5–7] risks ranked by commercial impact. For each:
- Risk description (specific, not generic)
- Trigger — what would cause this risk to materialise
- Worst-case cost impact in £ or %
- Likelihood: LOW / MEDIUM / HIGH with reasoning
- Recommended contingency allocation (either as a line item or % uplift)
- The one assumption that, if wrong, would blow the margin — state it plainly
Then produce a final margin-protection summary:
- Total base estimate
- Recommended contingency % based on scope clarity and risk register
- Minimum overheads and profit % to maintain margin given the risk profile (my typical target is [X]%)
- The three questions I should answer for myself before this estimate leaves my desk
Do not soften the risks. If the scope is weak, say the estimate should not be submitted until specific gaps are closed. Challenge me — this is the pass that separates a 95% accurate estimate from a 75% one.
A formal risk register. A plainly-stated margin assumption. Three questions to answer before the estimate leaves your desk. This is what separates a 95% accurate estimate from a 75% one.
The Short Version
Most estimates are 10-20% wrong. You don't know which way until it's too late
The 2015 spreadsheet, gut feel, and generic ChatGPT all skip the same step - stress-testing the number before it goes out
The 5-Prompt Chain bolts scepticism onto every estimate: Role Lock → Scope Validator → Takeoff QA → Rate Sense-Check → Risk Register
The prompts are templates. Fill in YOUR project types, YOUR region, YOUR qualifications. They only work when the AI knows who you actually are
Estimating is scepticism.
Which of these five are you currently skipping?
P.S. There are two more prompts I use that didn't fit in this issue:
1. The Submit Bolt-on — takes your final estimate through a last-pass QA before it leaves your desk.2. The Subcontractor Bid Comparison — weighs competing subcontractor quotes against each other on scope, not just price.
Both are in the free 7-Prompt Chain PDF.
If this was useful, the newsletter is free and weekly. One system per week you can set up on Monday morning.
