Code decides, AI writes

Most organisations want AI to take work off people. Most are also nervous about letting it decide anything. Both instincts are right, and you don't have to choose.

The problem

On our sports club platform, volunteers spend evenings chasing: renewals, unanswered enquiries, unconfirmed match scores, waiting lists. It's repetitive, but it's also personal. A wrong message to a member, or any message to a child, is a real problem.

What we did

We split every assistant into two parts. Plain code decides who to contact, when and why, using rules anyone can read. The AI model only writes the words, and code adds every link, so the model can't send anyone anywhere. Around that sits one shared engine with the same guardrails for every assistant:

  • Each message is sorted by risk. Routine ones can go out on their own; anything unusual waits for a person; anything that can't be undone, like publishing a post, always waits.
  • Limits per person: a member hears from the assistants at most twice a week, never at night, never if they've opted out, and never if they're a child.
  • A monthly spend cap per club. At the cap, plain templates take over.
  • An off switch that stops everything at once.
  • Example cases for every assistant, run as automated tests on every change.

In an enterprise

The same split answers the question every risk team asks: "how do we explain what the AI decided?" The answer is that it didn't. The rules decided, and they're in the code. The model drafted, and a person approved where it mattered.

What we'd do differently

Write the example cases before the assistant, not after. They are the specification.

Book a call

Tell us what you need and we'll get back to you.

Tell us the problem Call 07746850139WhatsApp