Point your AI at the job: 142 tools and a REST API

Every construction software vendor is announcing AI this year. Almost all of it is the same shape: a chat box inside their app, answering questions about their data, in their words, on their schedule.

The SitesAI App does something less fashionable and more useful. It hands the whole app to whatever AI you already use, and gets out of the way.

What that means in practice

You connect Claude or ChatGPT to your company once. After that you are talking to your business:

"Which inductions expire in the next fortnight, and who do I need to chase?"

"Set up next week's roster the same as this week, but Dave is on leave Tuesday and Wednesday."

"Read the geotech report attached to the Riverside lead and tell me if anything changes our price."

"How many hours did we put into Lot 12 last month, and what did the stock cost?"

The assistant is not being handed a summary. It is using the same operations the app uses — 142 tools across leads, quotes, jobs, contracts, scope of work, variations, progress claims, sites, members, documents and compliance, shifts, worktime, dockets, items and stock, purchase orders, maintenance, tasks, comments, calendar, activity and search.

As far as we can find, no mainstream field-service or construction-ops platform offers this. The category ships chatbots; this ships the keys.

How does an AI assistant connect to the SitesAI App?

Sign in with your AI. Claude and ChatGPT connect over OAuth. You approve the connection, and at that moment you choose read-only or read-and-write. Only owners and managers can grant it.

A key for something you are building. Keys are minted in the app and shown once. Each carries scopes — read, write, or the narrow intake scope — and those scopes are enforced rather than advisory. An intake-only key can create a lead and genuinely nothing else, which is what makes it safe to paste into a website form or an automation.

Plain REST, at /api/ai/v1, for anything that is not an AI at all. It includes a phone lookup that matches on the last nine digits, so an inbound call can be identified as a customer, a contact, a member or an open lead before anyone picks up.

There is also a public intake endpoint, so the enquiry form on your own website can drop straight into your pipeline without anyone retyping it.

It can read what is attached, not just what is typed

Records carry evidence: site photos, PDFs from the engineer, scope documents, a voice note from an operator who could not type with gloves on. An assistant can read all of it — images, the text of a PDF or a Word document, and voice notes as transcripts.

This is the difference between an AI that can discuss your job and one that can actually check it. "Does the SWMS on file cover the work we scheduled?" is only answerable if something can open the SWMS.

The parts that are deliberately restrictive

Nothing sends. There is no tool to send a quote, on any surface, to anybody. That is a decision, not a gap. An assistant can build the quote, price it, attach it to the lead and get it ready — and then a human being presses send. Automated outbound to customers is where AI mistakes become commercial ones.

It is scoped to one company. A key or connection resolves to exactly one company and every query is filtered by it. There is no cross-company view.

It is rate limited and audited. 120 calls a minute per key, and every call lands in an audit log kept for 90 days, so "what did the AI do on Tuesday" has an answer.

Destructive tools say so. Write tools are annotated so the assistant knows which operations change or remove things, rather than discovering it by doing it.

Profiles, because some hosts have limits. Some AI platforms cap how many tools a connection can offer. Rather than making you choose, the connection URL takes a profile: a compact set for phone and inbox automations, a default operating set for daily work, and the full registry for a desktop assistant. Same key, same company, different amount of surface.

Why this is worth having

Most small civil businesses do not lack data. They lack the twenty minutes on a Sunday to go looking through it. The questions that actually cost money — whose ticket expires before the job starts, which quote has been sitting unanswered for a fortnight, which machine is overdue a service, what a job really cost once materials are counted — are all answerable from what is already in the app. They just never get asked.

An assistant that can query the business directly is how they get asked.

What does the AI connector cost?

The AI add-on is a flat $38 a month for the whole company, or $380 a year — not per seat, because a price per seat would punish you for letting more people ask questions. Current details are on the pricing page, and the full tool list is on the AI connectors page.

The app itself is free up to ten licences. Sign in with Google, Microsoft or Apple at app.sitesai.net. If you also price work, the takeoff app has its own connector — see Run your earthworks takeoff from Claude or ChatGPT.