Skip to content
Field notes

Point of view · 6 min read

What AI should and should not be allowed to do in an estimate

A language model is very good at remembering that roofing needs labour. It is very bad at knowing what a bag of cement costs in Tema this week. Build the boundary accordingly.

Obed Buadey · Founder, EstimatePro Africa ·

There is a version of construction software being marketed at the moment that promises to tell you what a building costs from a sentence of description. It is a bad idea, and the reason has nothing to do with whether the technology is impressive.

The failure mode is that it looks right

A model asked for a price will produce one. It will be plausible, it will be formatted confidently, and it will be derived from text rather than from your supplier’s invoice. A wrong quantity is caught by anyone who has been on site. A wrong price that is within twenty percent of reality is caught by nobody until the job is half built.

Where models genuinely help

  • Recall. Roofing sheets without roofing labour, tiles without waste, an estimate with no transport line — these omissions are patterns, and pattern recall is exactly what these systems do well.
  • Drafting scope. Turning “20 metre wall, 2.4 high, plastered and painted both sides” into a structured list of trades and quantities saves a real half hour.
  • Language. Turning your shorthand into a description a client can read without turning it into marketing copy.
  • Comparison. Flagging that this quotation sits below the margin you normally achieve on this type of work.

Where they must be kept out

  • Rates. Every price comes from your price book, your supplier, or it is left blank and shown to you as blank.
  • Arithmetic. Quantities, waste, markup, overhead, tax and totals are computed in code that can be unit-tested, not generated as text.
  • Silent edits. Nothing enters an estimate without a person accepting it.
  • Cross-company inference. Your rates are not training material for someone else’s quotation.

What that looks like in practice

You describe a job. You get back a scope with quantities, each line mapped to a material in your own price book with the source shown. Lines that could not be matched arrive unpriced and clearly marked, because an honest blank is more useful than a confident invention. You review, adjust, accept — and only then does the calculation engine produce a number.

The result is less magical than the demonstrations you have seen, and considerably more useful: an estimate that is fast to produce and that you can defend line by line when the client asks how you arrived at it.

I do not need software that guesses what a building costs. I need software that stops me forgetting the crane.

A contractor in Tema, during an early product interview

Put this into practice.

The platform applies the arithmetic in this article automatically — markup against margin, overhead before profit, waste inside the quantity — using your own supplier and labour rates.