Why we build and run AI and websites under one roof

What goes wrong when build and run are split?

When one business builds a system and another is expected to run it, nobody owns it when it breaks. The builder has moved on, the IT provider did not design it, and a partner or practice manager ends up as the go-between. There are three common versions of this gap.

• AI consultants often hand over and leave. The workshop, pilot or prototype is often good. What is missing is anyone watching the agent in month four, when a model is retired, a mailbox rule changes or a client replies in a format nobody tested.
• Web agencies rarely own security. Many build attractive sites, then leave hosting, patching, backups and admin access to the client. A firm’s website is public by design, which makes it one of the first systems an attacker sees.
• Managed IT providers seldom build AI. A good IT provider keeps devices, networks and Microsoft 365 secure. Building an agent that chases tax documents, or a report that reconciles WIP from Actionstep, is different work that most IT providers are not set up to do.

Each of these businesses does its own job well. The problem is the gap between them, and firms are the ones left standing in it.

What does “run” actually involve each month?

Run is the unglamorous work that keeps a system accurate, secure and useful after launch: monitoring, fixes, access changes, vendor updates and small improvements. For AI clients, our Managed AI service covers hosting and monitoring of agents, automations and private AI, plus seat administration for public AI tools such as ChatGPT or Copilot.

A typical month includes:

• Watching every scheduled run. Overnight report refreshes, document-chasing sequences and intake flows are checked, and failures are fixed, ideally before anyone at the firm notices.
• Reviewing agent activity. Our agents wait for human approval before anything leaves the firm and keep a full audit log, so we review what was approved, rejected or left waiting.
• Managing access. New starters get the right access, leavers lose it, and permissions are reviewed so staff see only what they should.
• Absorbing vendor change. Models, APIs and plans change underneath every AI system. As at September 2026, Anthropic says it regularly retires older models and gives at least 60 days’ notice before retiring a publicly released one (Anthropic model deprecations). Even names move: Microsoft’s documentation now notes that Microsoft 365 Copilot has been renamed Microsoft Copilot (Microsoft Learn). Each change means testing and updating whatever depends on it.
• Seat admin for public AI. Licences, single sign-on, workspace settings and training opt-outs are checked as staff come and go and vendor defaults change.
• Small changes as the firm changes. A new partner, office, practice area or billing target means the report, the agent or the knowledge base needs adjusting.

None of this is visible when it works. That is the point of it.

Why does an IT background matter?

An IT background means we treat an AI agent or a website like any other production system: controlled access, backups, patching, monitoring and a plan for failure. That discipline is what AI and web projects often lack once the launch is over, and it is the part partners find hardest to check for themselves.

We work in partnership with Key IT, an IT business.

Security also has to be provable, not asserted. Client data in what we build is stored in fully GDPR-compliant data centres, and we don’t describe our own controls as Essential Eight aligned without stating a maturity level. Our about page sets out who we are and our partnership with Key IT.

Why do we sell named products with set pricing?

We sell named products with a set pricing model because a partner can approve a defined scope and a monthly fee far more easily than an open-ended hourly engagement. A named product tells the partnership what it is buying, what it costs and who looks after it once it is live.

Named products also make running them properly possible. Every Practice Reporting build uses the same approach to connections, metric definitions and monitoring, so we can support it for the long haul rather than relearning a one-off system each time something breaks. It is priced the same way: a setup fee plus a monthly fee, quoted in writing after the discovery call. Our agents work on the same principle and are sold by name, for example the Document Chaser, by Pylon Digital.

Fixed prices also align incentives. When we are paid to run what we build, a fragile build becomes our problem to carry, not the firm’s.

What should a firm ask before it buys?

Ask who will run the system in its second year, not only who will build it. The answer tells you more about the risk than any demo will.

Five questions cover most of it:

1. Who monitors it, and how would we know if it stopped working?
2. What happens when the vendor retires the model or changes its API?
3. Who removes a leaver’s access, and how quickly?
4. Where is our data stored and processed, and who can see it?
5. What exactly does the monthly fee cover?

The same test applies whichever tool you choose. A shared AI tool needs seat admin, settings and a policy. A private assistant also needs hosting, document updates and monitoring. Our comparison of a private knowledge assistant vs a shared AI tool sets out which a firm needs and what each takes to run.

If you want to put those five questions to us against your own systems, book a free 45-minute discovery call.