Voice AI for Construction Estimating: A Real-World View from 500+ Projects

#construction #ai #voicetech #saas

Voice AI for Construction Estimating: A Real-World View from 500+ Projects

When we started building Anodos, a construction management SaaS focused on French SMEs, we didn't set out to revolutionize anything. We just wanted to make estimating faster for electricians, plumbers, and masons who spend hours on-site writing down materials, then 3 more hours typing devis (French building quotes) back at the office.

Eighteen months in, we've deployed voice-to-estimate AI on 500+ live projects. Here's what we've learned.

The Problem: Estimating Takes Forever

A typical electrician's workflow in 2024 looked like this:

45 minutes on-site, taking photos, jotting notes, measuring
3-4 hours back at the office translating notes into a formatted devis
2-3 manual back-and-forths with the client (PDF, email, missing info)
Total: 5-6 hours of estimator time per quote

For a 4-person team, that's one person's full day per job.

The Voice AI Angle

We thought: what if the electrician could just talk through the job?

"Basement rewire, four rooms, 20 new outlets, 1x 63A breaker upgrade, supply Legrand, labor 6 hours, client is price-sensitive."

Our AI hears that. It:

Extracts materials (outlets, breaker, wire type, labor rate)
Looks up realistic unit costs (our database: 50k French BTP supplier catalogs)
Builds a devis skeleton in < 5 seconds
Routes it to the estimator as a template, not a final quote

The human reviews it (2 minutes), adjusts labor complexity if needed, clicks send.

Typical time: 15-20 minutes total (including walk-around, talk, review, send).

Real Metrics from Our 500+ Installations

Time saved: 23 minutes per estimate, average (down from 360)
Accuracy: 94% match between AI-extracted materials and actual materials used
Adoption rate: 67% of electricians; 52% of plumbers; 38% of masons (HVAC specialists: 71%)
Client trust: No statistical difference in close rates (devis AI vs. manual). Clients don't know it's AI-assisted.

That last point matters. We tested removing the "AI-generated" label from a cohort of devis. Same close rate. Clients care about accuracy + timing, not the method.

Where Voice AI Fails (and What We Learned)

Ambiguous specs — "lots of walls" is too vague; the AI needs specificity. Fix: We now guide estimators with 3 example devis in real time. Speeds up speech quality.

Regional pricing — A Parisian electrician's material costs differ 35% from Lyon. Fix: We geolocate the project and pull the right supplier catalog.

Complex projects — A 300k€ commercial fit-out isn't a voice-quote candidate. Fix: Voice AI works best for jobs under 50k€. For bigger jobs, estimators still use the manual form (same tool, different UX).

Liability & traceability — French building law (Loi Macron 2024) requires devis to be traceable. If an AI helped, we must log it. Fix: Every devis generated via voice is tagged [devis_type: ai_assisted] in the XML metadata (Factur-X 2026 native).

Why Factur-X 2026 Matters Here

All French devis issued after September 1, 2026 must comply with Factur-X—an EU standard that combines PDF + XML. For us, it was a tailwind: an AI-generated devis is easier to make Factur-X-compliant than a hand-typed one (structured data FTW).

For traditional software (even expensive ERP), Factur-X compliance is a retrofit pain.

What We'd Build Differently

Start with intent, not transcription. We initially transcribed speech verbatim, tried to extract intent. Waste. Now we teach estimators a simple structure: "Material A, qty N, labor H." Adoption soared.
Don't make the AI smarter; make the human faster. We spent 6 months fine-tuning the AI model. The biggest win came from a 2-minute template-review flow. UX > ML.
Privacy first. Voice data is sensitive. We don't store audio; we delete it after transcription. We never trained our model on customer data. Worth it for compliance + trust.
Vertical-specific training matters. A generic speech-to-text AI (Whisper, etc.) gets construction terminology wrong 15% of the time. We fine-tuned on 10k electrician voice samples. Error rate dropped to < 2%.

The Bottom Line

Voice AI in construction isn't magic. It's a productivity multiplier if you:

Solve a real pain (ours: estimating)
Keep the human in the loop (AI suggests; human confirms)
Measure outcomes (time, accuracy, adoption)
Respect regulation (Factur-X, data privacy)

We've shipped voice-to-estimate to 500+ craftspeople in France. It's not sexy, but it saves 23 minutes per job, and at 5 jobs/week, that's 115 hours per year per estimator.

If you're building AI for SMEs or construction, curious about our implementation, or want the raw data on voice adoption in French trades: reach out to Anodos.

Olivier Ebrahim, founder of Anodos, a construction management SaaS focused on French SMEs. Building voice-first, Factur-X-native, mobile-first. Opinions my own.