Case: installation quotes with AI

Home Cases Installation quotes: from measurements to Presto…

Ship Shift · Cases

Installation quotes: from measurements to Presto in minutes

Last updated: 2026-11-17

Reads measurements, finds items in the price bank and delivers a Presto-ready quote. The technician only reviews and confirms.

Installation company (name not published).

The challenge

A technical office at an installation firm prepares dozens of quotes each month. Typical input is a measurements PDF or an incomplete BC3: long descriptions, mixed units and often no item codes. Hand-entering that into Presto (or Arquímedes/CYPE) and looking up every line in BEDEC, BCCA or an in-house bank consumed whole afternoons.

Without a code, text search fails: synonyms, site abbreviations and the designer's wording do not match the bank card. Technicians ended up juggling tabs, copying quantities and fixing units. Tender deadlines do not wait for a «perfect» BIM model.

The pain was not «generate text with AI» but rigorous classification: hit the right item, anchor price to the bank breakdown, and leave a .bc3 the estimating software opens cleanly. Any hallucinated price was unacceptable.

What we did / The solution

We built a hybrid matching engine (text + semantic) on a curated bank, with AI proposing and the technician deciding. Prices come from the bank breakdown — the AI never invents amounts.

  • PDF/BC3 reading: code (if any), unit, quantity and normalised description.
  • Hybrid search in BEDEC/BCCA/own bank with embeddings and unit rules.
  • Top-5 suggestions with confidence scores; technician picks or corrects.
  • Bank prices and breakdowns; block «creative» invented items.
  • Export .bc3 (FIEBDC-3) ready for Presto / Arquímedes / CYPE.
  • Review queue for low-confidence lines or no clear candidate.

The flow was designed with the technical office: side-by-side measurement/item view and traceability of which candidate was chosen. Adoption depended on the first real quote being faster than the old method.

We logged typical matching errors (bad units, material synonyms, compound items) and turned them into rules or training examples. Each weekly cycle shrank the «no candidate» queue without opening the door to invented prices.

Results

MetricValue
First-hit accuracy83%
Correct in top-591.8%
Items in curated bank1,045
Test project159 uncoded items

Measured on a real public project classified with the system. The technician remains responsible for the signed quote: AI proposes; the signature is human. The qualitative jump is from «an afternoon per quote» to a minutes-long review when matching fits.

With 83% first-hit accuracy and nearly 92% in the top-5, human effort concentrates on the hard lines — where it adds most value — instead of typing the easy ones.

How we rolled it out

01

Phase 1 · Bank & sample

Curated the installation-relevant item bank, defined field schema and selected a public test project with 159 uncoded items.

02

Phase 2 · Engine + review UI

Tuned hybrid search, built the top-5 UI and measured first-hit and top-5 coverage before scaling volume.

03

Phase 3 · Controlled production

Pilot on real office quotes, short training and rules for when to force manual review. Weekly iteration on matching errors.

Lessons & who it fits

Without a curated bank, semantic matching degrades. The metric that matters is first-hit accuracy + calibrated confidence — not «pages generated». Exporting to the software the team already uses (Presto) avoids an adoption war.

It fits technical offices in installations/engineering with bid volume and BEDEC/BCCA/own banks. It fits less for generic civil works without a stable typology, or if the real process already births a structured BOQ from BIM.

FAQ

Why no company name?
Client request; the case and matching metrics are real.
Replace Presto?
No — feeds it with a ready-to-open .bc3.
Generic civil works?
Pilot was installations; other types need their own bank and sample.
Own price bank?
Yes — preferred; matching improves with the team's real vocabulary.
How fast by hand?
Hours; the system proposes in minutes and leaves human review.
Does AI invent prices?
No. Amounts come from the bank breakdown; no candidate means review.

Want the same pattern in your technical office?

Send an anonymised measurements excerpt. We'll assess matching fit.

Let's talk