Our approach to AI
Useful, predictable, and supervised — in that order.
The principles we apply when deciding what to build, how much autonomy it should have, and how we know it works.
Principles
Six principles we build by
- 01
Start with the workflow, not the model
We map how work happens today — inputs, decisions, systems, exceptions — before choosing any technology. Most of the value comes from understanding the process well enough to know where AI belongs and where it doesn't.
- 02
Use the simplest tool that works
Deterministic rules handle what can be specified. Traditional software handles structure and state. Language models handle interpretation and language. Agents are reserved for multi-step work that genuinely needs planning.
- 03
Autonomy is earned, per task
Agents begin supervised. Individual actions become autonomous only after they meet agreed criteria on real cases. Financial, legal, and customer-sensitive actions keep approval gates unless you explicitly decide otherwise.
- 04
Measure, don't assume
Every system has an evaluation suite built from real examples. Changes to prompts, models, or tools are tested against it before release, and production is monitored for drift.
- 05
Make AI behaviour visible
Interfaces show what an agent did and why, with sources and logs. People can only supervise what they can see.
- 06
Avoid unnecessary lock-in
Where practical we design provider-agnostic layers, keep prompts and configuration in your repositories, and document the system so another team could run it.
How agents operate
From request to action — with control built in.
- 01
Observe
The agent reads only the systems, data, and context it has been authorised to access.
- 02
Plan
It determines the next step against explicit business rules, policies, and instructions.
- 03
Act
It calls approved tools, updates records, drafts communication, or performs an allowed task.
- 04
Verify
It checks the result, records what happened, and escalates anything uncertain or outside policy.
Chatbot vs agent
A chatbot talks. An agent gets work done.
| Dimension | Chatbot | AI agent |
|---|---|---|
| Primary output | Answers and generated text | Completed tasks and updated systems |
| Who drives the work | Waits for the user to ask | Works through a defined task |
| Context | Mostly the current conversation | Authorised systems, records, and policy |
| Tools | Usually none | Approved tools with scoped permissions |
| Effect on your systems | Does not necessarily change anything | Updates records within allowed limits |
| Checking its work | Rarely | Verifies results and logs every step |
| When unsure | Answers anyway | Escalates to a person |
Related
Keep reading
Next step
Bring us the process your team is tired of doing manually.
A strategy call is a working conversation about one workflow: what it involves today, what could be automated, and what a first build would look like.
What happens next
- 1Tell us about one workflow
- 2We review it before replying
- 3A working session, then a first-build proposal