Put AI agents to work with controls you can inspect.

Where the problem shows up

Most businesses begin with AI agents by testing what they can do. A prompt is written, an agent is connected to a few tools, and the first results are impressive enough that its use quietly expands to more data and more tasks.

The contradiction is that the agent's capability grows faster than anyone's understanding of its limits. Nobody has written down which actions it may take, which data it may read, what it must show as evidence, or which decisions still need a person. When it makes a mistake, the business discovers that nobody can say who approved the action or how to trace it. What happens when an agent with broad access acts on an incorrect assumption? And who is accountable for a decision that a person never saw?

The better principle is to define an agent's terms before it works: the vocabulary it uses, the actions it may take, the evidence it must show and the decisions that stay with people. That is the work I do.

What I have changed for others

100+ people: specialised agents with narrow tasks

An organisation of more than 100 people needed its Slack, Gmail and team activity turned into usable reports. Instead of one generic prompt, I built a system with several specialised agents for extraction, interpretation, retrospection, trajectory analysis and summarisation. Each agent has a narrow task, which makes its output easier to check. The system produces a Weekly Trajectory Report for each employee and a company summary. Read the case study.

HIPAA-regulated platform: access control and audit trails

A regulated platform needed every sensitive action to be controlled and traceable. I built role-based access control, audit logging and medical review workflows. The platform has had zero security breaches since deployment. Read the case study.

What changes for your business

Each agent has a written scope, its actions are logged, and the decisions that need a person are routed to one. When something goes wrong, you can see what the agent did and why. The work can include agent scoping and permissions, evidence and logging requirements, human review points, and documentation of the rules each agent follows.

Where this leads next

Governed agents sit on top of consistent information systems and dependable automation. If your agents are already scoped and logged, we can begin with a review of their review points.