One placeto quote the job, track the money and run the crews, from a phone on site.

One operational hub for a façade-maintenance business. Every job, quote, contract, client and photo in one place, priced off the company's own history.

It is drawn as three floors: sales, finances, and operations.

The sales floor runs on eleven working surfaces. A command deck computes the picture live: pipeline value, proposals out, conversion rate, a money weighted funnel, the largest jobs, the quote tier mix, and what to do next. Leads carry every enquiry with its origin, service type, status and next action date, with overdue work flagged, which is the answer to the question the messaging app could never answer. The pipeline puts every job on a board by funnel stage and writes each stage move to a log as it happens, so the funnel has a history rather than a current mood. Clients hold the database: per client, their jobs, their quotes, their photos, their history.

Quoting is two surfaces, because this business needs both. One click quoting builds a quote from the shape of the job itself, its area, height, building condition, service type, complexity, access method, pathologies, materials, margin and tax. It is previewed first, every value is editable, and nothing sends on its own. The full quote builder does the detailed version, priced, taxed, rendered as a branded document and sent. The pricing underneath is not a guess and not a language model: it is the company's own parametric model, its rates per square metre and its complexity and access multipliers, ported into arithmetic.

Then the surface that changes how the business earns. While a quote is being written, comparable jobs shows what similar past work actually came to per square metre, from the company's own completed jobs, banded by type of work rather than flattened into one rate, and flagged when the sample is too small to trust. The company's history becomes its pricing reference, and the flag is as important as the figure.

The rest of the floor: contracts generated in three steps from an awarded job and a reusable template, pre filled and reviewed before they exist. Tasks tracked across commercial, site, admin and after sales, with deadlines and client linkage. Job photos from before, during and after, attached to the right job automatically. And a duplicate review queue that proposes merges, waits for a human decision, keeps every one reversible, and disappears when there is nothing to review.

The finances floor covers money in and out against each job, with each job treated as its own small company. A tax map split by rate, ready for the accountant. A cash flow forecast for the weeks ahead. And real margin per job, what was quoted against what was actually spent.

The operations floor covers scheduling per job across subcontracted crews, a jobs calendar by week and month, a site diary with before and after photos, each job's state at a glance, crews and subcontractors with their certifications, insurance and expiry dates, a company document archive, monthly executive reporting, and one combined panel showing sales, jobs and finances together.

All of it installs on a phone like an app, because the work happens in front of a building, not at a desk.

Now the part worth more than the word AI. Every number the system worked out for itself is marked as derived, on screen, with a distinct mark when its confidence is low. Nothing auto derived is ever quietly blended with a real figure. For a business that had just been carrying a pipeline number it could not defend, that property matters more than any feature: you can always see which figures are yours and which ones the system inferred.

Where AI does the work is the history. A language model read the text of 223 candidate quote documents and recovered the values and line items that rule based parsing could not reach, with constrained output, plausibility bounds and idempotent re runs, filling blanks only and never overwriting a real number. It only ever saw scrubbed text: tax IDs, bank details, phone numbers and emails were removed before anything left, so no personal data reached a model.

Underneath everything, the data foundation, because none of the above is worth building on numbers you cannot defend. All 3,030 drive items were read view only and triaged into 2,668 business, 326 personal and excluded, and 36 quarantined for personal data. The three sum exactly. Every one of the 387 business documents was extracted, 387 of 387, zero failures. The real entities behind them were rebuilt: 187 clients, 187 works, 328 quotes, 22 contracts and 214 photos.

Then the pipeline was audited against itself, duplicate by duplicate. The ambiguous cases were not guessed at. They went back to the owner as direct questions, and eighteen reversible merges were applied on his answers. The result was cross checked and matched to the cent.

It runs on the client's own database account, with the data under the client's control, authenticated and row level secured per table, with encrypted storage. The connection to the document drive is read only, so the system cannot write to or delete the source of truth it learned from.