Most AI in business software is a chatbot in the corner of the screen. It’s impressive in a demo and forgotten by the second week. We’ve taken a different line in our own products: put AI where the work already happens, connect it to live data, and be very clear about where it stops.
Here’s what that looks like in practice across Graftix and Kenaez, and why we think MCP is one of the most useful things to happen to project tools in years.
Dave, the assistant inside Graftix
Graftix is our site and field management platform for specialist fire, security, electrical, data and AV contractors. Dave is the AI assistant built into it, and the important word is “into”. Dave isn’t a general chatbot. It’s connected to your live project data, and it turns up where the work already is:
- Plain-English answers about your own projects, tasks, sites and costs.
- A RAMS written from a project intake, with likelihood and severity scored one to five.
- Minutes from a meeting transcript, leaving a section out rather than padding it.
- A rewritten section of a tender proposal, grounded in the design and previewed before you accept it.
- The executive summary at the top of a highlight report, written from the project’s own numbers.
Alongside Dave sits symbol takeoff. Quantifying a drawing set is one of the last genuinely manual jobs in our trade: a highlighter, a clicker and a lost weekend. Takeoff reads the legend, you box one instance of a symbol, and every match on the sheet is counted from the drawing’s own CAD geometry, at any rotation. Give it a scan instead of a CAD-exported PDF and it tells you, rather than inventing a number.
MCP: talk to a live project from your own AI
MCP, the Model Context Protocol, is the open standard Anthropic created for connecting AI assistants to real systems. Graftix speaks it. Issue a token in Settings, connect your own Claude or another MCP assistant, and it can read your live project data directly: projects, tasks, blockers, programme, materials, financials, attendance, RFIs and certificates.
On a major project, that changes the shape of the week. Instead of someone pulling a report together on a Friday afternoon, a project manager can just ask:
- “Give me a progress report across every building on the campus.”
- “What has been blocked longest, and who owns it?”
- “How does material spend compare to budget this month?”
- “Who is on site right now?”
The answer comes back in writing, from live data, in the time it takes to type the question. It can act too, within limits: update a status, assign work, set dates, log a blocker or raise an RFI. It only works inside the permissions your account already has, the token can be revoked at any time, and every change lands in the task audit history like any other edit.
The bigger point is that project data stops being locked inside one app. The assistant you already use for everything else can now reason over the programme, the RFIs and the money in the same conversation. We use this way of working ourselves; this website was rebuilt by an AI agent working through WordPress’s MCP connection, with us directing the work and signing it off.
Kenaez: AI that watches cameras, with rules
Kenaez turns the IP cameras an estate already has into a live, searchable view, with 23 AI detection modules and an assistant that answers control-room questions such as “Summarise what happened at the depot between 10pm and 6am”, with links straight to the footage. It works in the app, on WhatsApp and over MCP.
AI that watches people has to meet a higher bar than AI that writes emails, so Kenaez is built around lines that can’t be switched off:
- No facial recognition, ever. An automated test fails any release that tries to introduce face matching.
- Retrace finds people and vehicles by description, what they wore and what they carried, never by face.
- Results come in bands (linked, strong, possible, weak) with the reasons shown, never as a percentage, because a percentage reads like an identification.
- A named operator confirms every sighting. AI proposes leads; people make the decisions.
There’s more on how each of those is enforced on the Kenaez responsible AI page.
Where AI stops
Across everything we build, the same principles apply:
- AI assists; it doesn’t sign anything off. In our world a wrong number ends up on a certificate with somebody’s name on it, so every AI output is reviewable before it’s used.
- Data stays close to home. Dave runs on Claude models through AWS Bedrock with inference in London. Kenaez does the same for anything that escalates, with routine checks kept inside the EU. Customers who would rather bring their own AI provider can. More on that in why we build and host in the UK.
- No training on customer data. Neither product trains AI models on your projects, footage or cases.
- Everything is logged. A change made through AI lands in the same audit trail as a change made by a person.
The detail for Graftix is in its AI policy.
Want this in your own business?
The same approach works outside our products. If your systems hold the answers but there’s no easy way to ask them, we can connect them: MCP servers for your own tools, assistants grounded in your own data, and automations that do the admin without making the decisions. Have a look at integrations and automation, or get in touch.