Tendering and estimating, 12 November 2026

Is AI tendering accurate? How to test it on your own tender

AI tendering is accurate enough to save hours on reading and measuring, but no tool gets every line right. Test it on a tender you've already priced before you trust it. Give it a tender you've already priced, compare its quantities with yours line by line, and look at why each gap happened. You still own the rates, the risk and the final price.

Is AI tendering accurate enough to use on a live tender?

It can be, once you've seen how it does on your own drawings. Nobody can give you a single accuracy figure for "AI tendering", because results depend on the drawings, the trade, the tool and how it was set up. Any supplier quoting a blanket percentage hasn't measured your work.

The Information Commissioner's Office, which regulates how AI handles personal data, puts it plainly in its AI guidance: statistical accuracy is "the proportion of answers that an AI system gets correct or incorrect", and that "does not mean that every inference has to be correct". The same guidance says you should test throughout the life of the system and keep monitoring it after you start using it. That's the right attitude for a take-off too. Expect some errors, find out where they fall, and set up your checks around them.

AI tendering tools tend to be good at the slow, mechanical parts: reading a long pack, pulling out dates and forms of contract, counting symbols, measuring areas from dimensioned drawings, and drafting qualifications. They're weaker at anything the drawings don't show, which is where your site knowledge comes in.

We've noticed hallucinations.
Director, property developer and builder, four to five live sites · Paid ChatGPT user for 18 months

A general-purpose AI assistant that fills gaps with plausible guesses is the extreme case. A tool built for take-offs should show you where every figure came from, so you can check it. If it can't, treat its numbers as a rough guide only.

How do you test an AI take-off against a tender you've already priced?

Use a past tender you know well, run the AI on exactly the same documents, and compare quantities line by line before you look at any price. The seven steps below take one evening on a small package.

  1. Pick the right test tender.

    Choose one you priced in the last year, in your own trade, with a full set of drawings. A job you won and built is best, because you know the real quantities from the final account and not only your estimate.

  2. Give the AI the same pack, and nothing else.

    Same drawings, same revision, same spec and schedules. Don't tell it your answers or fix its mistakes as it goes. Write down which files you gave it.

  3. Ask for quantities first.

    Ask for each item with its unit, the drawing number and scale it measured from, and whether waste is included. A take-off that can't show its source drawing is hard to check.

  4. Agree the measuring rules.

    The RICS measurement rules for building work (NRM2) measure work net as fixed in position unless stated otherwise, with waste allowed for in the item, not added to the quantity. If your take-off is net and the AI's includes 10% waste, every line will look 10% out.

  5. Fill in the test sheet line by line.

    Your quantity, the AI's, the difference and a reason. Don't stop at the total.

  6. Price both take-offs with your own rates.

    This separates measuring errors from pricing errors, and shows what each gap is worth in pounds.

  7. Repeat on two more tenders.

    Pick different types of work: a refurbishment as well as a new build, or a dense M&E drawing as well as a simple plan. One good result on an easy job proves little.

For how a manual take-off is laid out, see quantity take-off. For which parts of the pack to read first, use the 10-minute tender read.

What does a line-by-line comparison look like?

It looks like the table below: most lines close, a few far out, and one or two missing. This is an illustrative test on an internal works package for an office refurbishment, priced at £64,000 by the contractor before the AI ran.

LineYour quantityAI quantityDifferenceWhy
Stud partitions (m²)412409−0.7%Openings deducted slightly differently
Plasterboard ceilings (m²)286318+11.2%Measured across the stair void
Fire-rated board to corridor (m²)1481480%Matched
Skirting (m)2360MissingOnly on the finishes schedule, not the drawings
Internal doors (nr)1836+100%Counted on the plan and again on the elevations

Partitions: 409 − 412 = −3 m², and 3 ÷ 412 = 0.7%. Ceilings: 318 − 286 = 32 m², and 32 ÷ 286 = 11.2%.

Worked example

What the AI's errors were worth on a £64,000 package

Each difference priced at the contractor's own rates for the package.

Ceilings measured across the stair void+32 m² × £38+£1,216
Partitions short at openings−3 m² × £52−£156
Skirting missed−236 m × £9−£2,124
Doors counted twice+18 × £310+£5,580
Net effect on the price£1,216 − £156 − £2,124 + £5,580+£4,516
Net error as a share of the tender£4,516 ÷ £64,0007.1%
Size of the errors ignoring direction£1,216 + £156 + £2,124 + £5,580£9,076
Gross error as a share of the tender£9,076 ÷ £64,00014.2%
Quantities, rates and the £64,000 tender value are illustrative. The point is the method: the net error (7.1%) understates how wrong the take-off was (14.2%), because overs and unders cancel.

Two lessons. First, a total that looks close can hide large errors in both directions, so never judge an AI take-off on its bottom line. Second, three of the four errors had one cause each that you can fix with a rule: exclude voids, read the finishes schedule, count doors from the door schedule only.

Where does AI usually get a tender wrong?

In the same places a rushed human does, plus a few of its own. Check these first on every test.

  • Scale. Drawings marked "do not scale", scales that differ between sheets, or a PDF printed at a different size from the original. A wrong scale throws out every area on that sheet by the same factor. Check one known dimension per sheet. See how to measure PDF drawings.
  • Hidden or missed pages. Appendices, finishes schedules, door schedules and late addenda. If an item lives only in a schedule, a drawings-only take-off misses it, as the skirting did above.
  • Double counts. The same item shown on a plan, an elevation and a detail.
  • Spec clashes. A spec clause naming one product and a drawing note naming another. A good tool flags this; a weak one quietly picks one of them. Check which product it priced.
  • Provisional sums and "by others". AI may price an item the client has covered with a provisional sum, or miss that something is excluded. Check these against the tender form, not the drawings.
  • Units and slopes. Metres against square metres, plan area against slope area on a pitched roof, or numbers against linear metres for pipework.

For a list of checks to run on any AI-drafted price before it goes out, see AI tender pricing checks.

Is the difference big enough to worry about?

It depends on what the line is worth and whether anything is missing, more than on the percentage. A 10% difference on a £200 line is noise; a 2% difference on a £20,000 line is £400 off your margin. The bands below are a working rule of thumb, not an industry standard. Set your own once you've run a few tests.

Quick check: Is the difference big enough to worry about?
Is the item missing, in the wrong unit, or counted twice?
Is the line worth more than about 5% of the tender value?
Is the AI's quantity more than 2% away from yours?
Is the AI's quantity more than 10% away from yours?
Fix the rule, then re-run

Missing items, unit errors and double counts are process errors. Find the cause (a schedule it didn't read, a duplicated view) and set a rule for it before trusting the next take-off.

Find out which of you is right

On a big line, a 2% gap is real money. Re-measure from the drawing it used. Sometimes the AI is right and your old take-off was out.

Note it and check the pattern

On a small line this won't sink a tender, but if the same item is always out the same way, correct it at source.

Within tolerance

Record it on the test sheet and move on. Keep checking the big lines on every live tender.

What still needs a person on every tender?

Everything that isn't on the drawings. An AI can be right about every quantity and the tender can still lose you money.

  • Site knowledge. Access, ground conditions, the state of what's there, the neighbours, the parking.
  • Buildability. Whether the detail can be built as drawn, in the order the programme assumes.
  • Your output and rates. What your team installs in a day on this kind of job.
  • Risk and scope gaps. What's missing from the pack, who carries it, and how you word it in your qualifications and exclusions.
  • Sign-off. Somebody in your firm reads the price and approves it before it goes.

How do you keep checking once you've started using it?

Spot-check the biggest lines on every live tender and re-run the full test whenever the tool, the trade or the drawing type changes. The ICO's guidance on testing throughout an AI system's life applies to your own use as much as to the people who built it.

A simple routine that works for small firms:

  • On every tender, re-measure the three most valuable lines yourself.
  • Keep the test sheets. After five or six tenders you'll see which items the tool gets right every time and which it doesn't.
  • When you win a job, compare the AI take-off with the final quantities. That's the most honest test there is.
  • If the supplier updates the tool, run one old tender through it again.
AI take-off test sheet
Tender: project name and package
Documents given to the AI: drawing numbers and revisions, spec, schedules
Tested by: name   Date: date
Measuring rules agreed: net as fixed / waste included at %

Line | Unit | Your qty | AI qty | Difference | % | Rate | £ effect | Why
item | m² / m / nr | qty | qty | AI minus yours | difference ÷ yours | £rate | £difference × rate | scale / missed page / double count / spec clash / unit / matched
item | | | | | | | |
item | | | | | | | |

Net £ effect (add with signs): £total
Gross £ effect (ignore signs): £total
Gross effect as % of tender: %
Rules to set before next test: e.g. count doors from the door schedule only

The guide to AI tendering by trade covers what to test for in each trade, and AI tendering for dryliners and AI tendering for builders go into the lines that most often go wrong for those trades.

Where Nexley fits

Nexley is the AI back office for UK trade firms: five AI assistants run your tenders, quotes, buying, scheduling and invoicing. It connects to the tools you already work in, you run it all from WhatsApp, and you approve everything.

To test Nexley, forward a tender you've already priced on WhatsApp and ask for the take-off only. Nexley reads the drawings, spec and schedules, lists each quantity with the drawing it came from and flags spec clashes and items it couldn't find. You compare it with your own figures on the test sheet, and nothing is priced or sent without your approval.

Common questions

What is AI tendering?

Using AI software to read a tender pack, measure quantities from the drawings, draft a price from your own rates and write the qualifications. The tradesperson still checks the take-off, sets the risk and approves the price.

Can AI price a tender on its own?

No. It can draft a price from quantities and your rate card, but it can't judge site conditions, buildability, your gang's output on this job or the commercial risk. Treat its price as a first draft that you check line by line.

Is a specialist take-off tool more accurate than ChatGPT?

Often, because it's built to show the drawing and scale behind each figure, which makes checking possible. But "often" isn't proof. Run the same test on both with one of your old tenders and compare the test sheets.

What if my old take-off was wrong?

That happens. When the AI and you disagree, re-measure from the drawing before deciding who's right. A test that finds a mistake in your own past pricing has still done its job.

Sources

  1. ICO: What do we need to know about accuracy and statistical accuracy?
  2. ICO: Accuracy of training data and model outputs (generative AI consultation response)
  3. RICS: NRM, New rules of measurement
  4. RICS: NRM 2, Detailed measurement for building works (October 2022 update)
  5. Customer quotes and details: Nexley's own sales and onboarding call notes, July to October 2026, anonymised.

Every rule, rate and date above was checked against these sources on the date shown at the top.

Topic: Tendering and estimating

Keep reading

Tendering and estimating

AI tendering for groundworks, drainage and surfacing contractors

A ground risk ledger that sorts every risk into price, qualify or reject, one table of the quantity conversions that cost groundworkers money, and a muck-away example at the 2026 Landfill Tax rates.

15 Nov 2026, 16 min read, Kade Dillon

Tendering and estimating

Let down by your bid writer? Fix this first

Why tenders lose even with a bid writer, five things to fix in what you give them, and when drafting in-house with Nexley is the better step.

15 Nov 2026, 5 min read, Kade Dillon

Try it on your own jobs

7 days free, then £799 a month. No contract.

Runs in WhatsApp Set up on a 20-minute call Cancel anytime
Saves me so much time every week. Does internal organisation and admin, but also when a customer messages in, Nexley replies, books them in against when I’ve got time to do it, and gets the invoice sorted in Xero.”
— Ollie Cable

Book a demo with Kade

A 20-minute call. Pick a time that suits and he’ll show you Nexley on your own work.

WhatsApp Us