← All insights

Architecture

Intake, decision, action, control: designing an AI workflow you can trust

If I walked into a business where enquiries were being handled through emails, spreadsheets, WhatsApp and people's memory, I wouldn't start by building an "AI agent".

I'd build something closer to four layers, each with a clear job, and each with its own security questions.

1. One intake layer

Website forms, emails, referrals and other enquiries all feed into the same system. Every record is created once, enriched and assigned automatically. No copying details between four places.

From a security point of view, this is also where untrusted data enters. Everything arriving here (form text, email bodies, attachments) should be treated as untrusted input, never as instructions.

2. A decision layer

The system works out things like: what is this person asking for, are they a good fit, who should handle it, how urgent is it and what information is still missing?

AI is genuinely useful for the messy, unstructured parts: reading a rambling email and pulling out what matters. Normal, deterministic software handles the rules. That split is a security control in itself: the parts that must behave predictably shouldn't depend on a model's interpretation.

3. An action layer

Based on the decision, things happen: records are updated, tasks are created, follow-ups are scheduled, documents are generated, the right person is notified and a response can be drafted.

This is where permissions matter most. Each action should run with the narrowest access it needs, and higher-risk actions should still require approval. Drafting a reply? Fine. Sending money, deleting data or emailing a client on your behalf? A human should press the button.

4. A control layer

This is the bit that normally gets forgotten, and it's the bit that turns a clever demo into something a business can depend on. I'd want to know:

  • Which items are stuck, and where are things dropping off?
  • What failed, and did anyone notice?
  • What did the AI decide, and why?
  • Which actions were approved, and by whom?
  • How much manual work is still happening around the system?

Visibility isn't just operational hygiene. It's how you detect misuse, investigate incidents and prove to a customer that the system behaves the way you say it does.

The real shift happens when a workflow stops being a collection of tools and starts behaving like infrastructure the business can depend on: predictable, observable and controlled, not just one step made faster.

Next step

Designing an AI workflow that needs to be dependable, and defensible?

Explore Secure AI Engineering