Artificial intelligence
It suggests. You decide. Always.
There is no setting that lets a model write a rate into a bill you are about to issue. Every AI feature in Quantifiable produces a suggestion, shows what it is confident about, and waits. The estimator's name stays against the number, which is the only arrangement a professional indemnity insurer would recognise.
The rule
A structural constraint, not a preference
Suggestions are stored as suggestions — separate rows with their own pending, accepted or rejected state. A pending suggestion contributes nothing to a quantity or a total. It is not that the interface asks you to confirm; it is that until you accept, the number does not exist in the priced document at all.
Nothing lands unattended
No background job, scheduled run or integration can convert a suggestion into a priced figure. Acceptance is a deliberate act by a person with a name.
Confidence is shown, not hidden
Suggestions carry a confidence indicator and, where the answer came from a source, a citation you can check. A low-confidence suggestion looks different from a high-confidence one.
Rejection is recorded
Turning a suggestion down is part of the audit trail. What the model proposed and what the estimator did about it are both on the record.
Structured output only
Model responses are validated against strict schemas before they are shown to you. Numbers cross that boundary as decimals, never as floating-point values, so a suggestion cannot introduce rounding drift.
Where it helps
Applied to the tedious parts, not the judgement
The work AI is good at here is the counting, the transcription and the first draft — the parts of estimating that consume hours without exercising any professional skill.
Symbol detection and counting
Point at a symbol on a drawing and get every other instance found and counted across the sheet, as suggested count measurements you review before they become a quantity.
Plan analysis
A first read of a drawing, identifying what is on it and what might need measuring, so you start from a checklist rather than a blank sheet.
Scale from the title block
Reads the title block and suggests a calibration. It is never applied on its own — you confirm the scale, because everything downstream depends on it.
Bill lines from measurements
Proposes NRM2 lines that fit what you have measured, with descriptions written in the right idiom, for you to edit and accept.
Rate prediction
Suggests a unit rate for a line from its description and context, with confidence shown, when your library has no entry to resolve against.
Labour assistant
Works from crew size, expected output and pay rates to a labour cost, so a build-up can be reasoned about rather than recalled.
Material specifications
Researches the specification and purchase unit behind a material description, with sources cited, so an item can be linked to a measurement in different units.
Programme generation
Drafts a programme from the priced work, with durations and dependencies derived from the bill. Nothing is created until you accept the plan.
Report narrative and email drafts
Produces first drafts of the commentary around a report and the correspondence around a project, into a form you edit before anything is sent.
Your data
What leaves, and what never does
The uncomfortable question about AI in a professional practice is what happens to the client information you feed it. These are the specific measures in the product, not a general assurance.
Prices are stripped from research
When a feature searches the web, the query is scrubbed of your rates and any identifier that could locate your workspace before it leaves. Web research never carries your pricing out with it.
Matching prompts are price-free
The prompts that link materials and descriptions are constructed without rates, so your commercial position is not present in the request at all.
Search indexes exclude rates
The embeddings that power library search are built from descriptions and specifications. Prices are not part of what is indexed.
Every call is logged
Each AI call records its purpose, its token usage and the workspace it belonged to. You can see what has been used and on what, in your own settings.
Model training
Where the guarantee comes from
Your project data is not used to train foundation models. That protection is contractual — it rests on the terms we hold with the model providers, and it is set out in our AI data-usage statement alongside the list of which providers process what. It is not something our code can enforce, and we are not going to imply otherwise. The measures above, which our code does enforce, are how we keep the amount of your information that reaches a provider as small as the feature allows.
Control
Metered, visible and yours to limit
AI usage is measured per workspace against a monthly allowance, with a warning as you approach it and a hard stop at the limit rather than a surprise on an invoice. Usage is broken down by feature in your settings, so you can see what is actually being used and by whom. Firms that would rather not use a particular capability simply do not use it — nothing in the estimating workflow depends on AI being switched on.
Faster where it should be, unchanged where it matters
Use AI for the counting and the first drafts. Keep the professional judgement, and the accountability, exactly where they belong.