Most service work is won or lost between the site visit and the quote landing in the customer's inbox. The details are fresh for an hour, fuzzy by the end of the week, and the competitor who sends a tidy quote on Wednesday tends to beat the one who promises something better next Monday. The bottleneck is rarely judgement. It is the friction of turning a head full of notes into a priced document.

Case Study, not Delivered Project.

How we'd approach this kind of problem. See delivered projects →

7x More likely to reach a buying decision-maker when you respond inside an hour, versus one hour later (Harvard Business Review)

You probably have this problem if…

  • Your surveyor leaves the site with photos and a voice note, and the office still has to rebuild the job from scratch before pricing it.
  • Quotes go out three to five working days after the visit, and you're regularly second behind a competitor.
  • The owner is the only person who can reliably write a final quote, so quoting stops when they're on holiday or on the tools.
  • Two surveyors visit similar jobs and send back wildly different notes, and the pricing reflects that inconsistency.
  • You've tried templates and they help a little, but every job has enough variation that the template still needs an hour of rework.

Any two of those is a strong signal. All five and you're routinely losing winnable jobs to whoever quoted first.

How the pattern works

Imagine a careful office estimator who sat in the passenger seat for every site visit. They heard what the surveyor said, saw what the surveyor saw, and know your price book by heart. By the time the van pulls back into the yard, a first draft of the quote is already on the office desk. They never invent prices. They mark anything they aren't sure about and leave the commercial calls to a human. That's the system.

Five things have to work for it to be useful.

Voice + photos Structured scope Price book match £ Draft quote Office review
From a voice note in a driveway to a priced draft quote in the office queue.

1. The surveyor captures the visit without typing

After each visit, the surveyor opens a short mobile form. They pick the customer, choose the job type, take photos, enter the headline measurements, and record a voice note. The voice note carries the bulk of the information. It is the thing they would have said back at the office anyway, recorded once instead of repeated three times. The whole capture takes a few minutes in the van before they drive off.

2. The system transcribes the audio and reads the photos

The audio is transcribed by a speech model tuned for noisy, accented, jargon-heavy speech. Trades audio is rarely studio-quality. There is wind, traffic, a kettle, a saw running next door, and the surveyor's regional accent. Generic transcription often drops 10 to 20 percentage points of accuracy on this kind of input, and a small custom vocabulary of trade terms is what stops "soffit" coming back as "sofa."

Photos go to a vision-capable language model. It is not trying to measure the room from a photo. It is reading what is visibly there. A combi boiler in a kitchen cupboard. An existing tiled splashback. A radiator on the wall the customer wants to remove. Those observations get folded into the structured scope alongside the transcript.

3. It maps the scope to your price book, not the internet's

AI model extracts scope Rules engine applies prices voice note photos measurements Remove old radiator Make good wall Install new combi Labour 1.5h × rate Materials + markup Combi from supplier £
The AI's job stops at a structured scope. Prices come from your own price book through deterministic code, so a hallucinated number never reaches a customer.

This is where most generic AI quoting tools quietly fall over. They guess prices from training data. They do not know your labour rate, your preferred suppliers, your margins, or the line items you always exclude and quote separately. We do not ask the model to price anything. We ask it to map the scope to candidate line items in your own price book. A rules engine then applies your labour rates, supplier prices, markups, VAT treatment, and standard exclusions. If a line cannot be matched with confidence, it surfaces as a flag for the office rather than a guess on the quote.

4. It writes a draft quote in your voice

The system assembles the quote in your standard format. Cover wording, scope of works, line items, inclusions, exclusions, payment terms, optional extras. The tone is calibrated against five to ten real quotes you have already sent, so the draft reads like the business and not like a generic AI assistant. Anything the system is uncertain about is marked, with the relevant chunk of transcript or the relevant photo linked, so the reviewer can see exactly why the flag was raised.

5. The office reviews and sends

The reviewer sees the visit evidence on one side and the draft quote on the other. They correct anything wrong, fill in what is missing, and make the commercial calls (price, margin, optional extras) before the quote goes out. Nothing is sent without a human signing it off. The job of the system is to compress a 90-minute drafting task into a 10-minute review.

The default stack

Site capture Custom mobile web form with audio and image upload
Speech-to-text Deepgram Nova-3 with custom trade vocabulary
Photo interpretation Claude Sonnet (vision)
Scope extraction Claude Sonnet (Anthropic)
Pricing engine Custom rules engine over your price book
Database Neon Postgres
Quote output Branded PDF rendered from HTML and CSS

Two choices that matter most.

A purpose-built speech model over generic Whisper. Off-the-shelf Whisper is fine for tidy, single-speaker desk audio. On noisy, accented, jargon-heavy speech recorded outside a van or in someone's kitchen, real-world error rates can drift well into double digits, and every error compounds into the structured scope downstream. Deepgram Nova-3 holds up better on the bad-audio cases and accepts a custom vocabulary, which is what keeps "RSJ" from coming back as "RSI." It is also fast enough to have the transcript ready before the surveyor finishes the form.

A rules engine over an AI-generated price. The model's job stops at the structured scope. Prices come from your own price book through deterministic code. This keeps quotes consistent with how the business actually operates, makes a margin change a one-line config edit rather than a prompt rewrite, and removes the entire class of risk where a hallucinated number ends up on a customer-facing document.

When this isn't the right fit

The pattern is powerful, but it is the wrong tool for some problems.

Bespoke, judgement-heavy work. If every job is a one-off architectural conversion priced from first principles, the pattern adds friction without saving time. The structured scope still helps, but the pricing engine has nothing standard to match against. Better to keep the pricing manual and use the system only for capture.

Very low quote volume. Below roughly 10 to 15 quotes a month, the build cost outruns the time saved. A good template, a tidy price book, and a habit of dictating notes into a shared doc will get you most of the way.

Fixed-list trades. If you mostly install a small range of standard products from a published menu, you do not need this. An online configurator or a simple form will produce the same quote faster and cheaper.

No price book at all. If pricing currently lives only in the owner's head, the work to extract it has to come first. The system can be built around that, but it cannot be skipped. Without rules, there is nothing to apply.

What to expect

Implementation time 4–8 weeks for a typical first build, depending on how clearly the price book is already documented.
Deployment options Cloud-hosted by default. The mobile capture form works offline and syncs when the surveyor gets signal.
Infrastructure cost Roughly £80–250 per month for typical volumes, covering transcription, vision and language models, hosting, and storage combined.
Typical drafting time 5–15 minutes of office review per quote, down from 60–120 minutes of full reconstruction.
Time and money recovered Harvard Business Review research suggests responding inside an hour makes you 7x more likely to reach a decision-maker than the firm that responds one hour later. For a typical mid-sized operator sending 30–40 quotes a month, the build also returns roughly 30–50 hours of admin time per month.
Secondary benefits Consistent quotes across surveyors, an audit trail from photo and voice note to final price, and a clean dataset for margin and conversion analysis later on.

If this pattern fits your team

A Pare Audit is the way to find out whether it does, and what a delivery would look like in your specific situation. We spend a focused few days with you, look at the real quotes you have sent in the last quarter, the real site notes behind them, and the price logic that already lives inside the business. You leave with a written recommendation, a scoped build, and a costed plan.