Approach
Agentic AI, with the human kept in
An AI agent doesn't just answer. It plans, uses tools and acts, step by step, until a goal is reached. That makes it useful, and it moves the risk: from what the model says to what it is allowed to do. My approach starts there.
My big-picture approach: The big picture
How an agent works
An agent pursues a goal in a loop: set the goal, plan, act with tools, observe the result, decide what comes next — and repeat until the goal is met or a limit stops it.
Building blocks
- Language model — the reasoning: what makes sense to do next.
- Tools — what the agent can actually do (search, write, call, store).
- Context — what it knows about the task, the person and the limits.
- Orchestration — the code that runs the loop, sets limits and records every step.
Autonomy is a decision, not a default
- Assistant: suggests; people do everything.
- Human in the loop: the agent acts, a person approves each step that matters.
- Human on the loop: the agent acts; a person monitors and can step in.
- Autonomous: no routine human involvement.
Organisations should earn autonomy task by task. Most public-sector use cases start in the loop, and some should stay there, especially anything that affects decisions about people.
What I design for
The risks shift once a system is allowed to act:
- Wrong actions rather than wrong words
- Small errors compounding across steps
- Hidden instructions in documents or emails the agent reads (prompt injection)
- More access than the task needs
- Personal data in prompts and logs
- Nobody able to say who decided what
Safeguards in three layers
Technical
- Least privilege
- Approval gates before anything irreversible or external
- Separate environments
- Treating everything the agent reads as data, never as instructions
- Step, time and cost limits
- Full logging
- Rollback
Organisational
- A named owner per agent
- A use-case assessment before building
- Testing with real cases
- Monitoring after launch
- Clear hand-over to people
- Training so staff know what the agent does and where it stops
Regulatory
- EU AI Act risk class and human-oversight duties
- Data protection, including special rules for social data in public institutions
- Information security
What I learned building one
EquitAI's job agent searches for roles and drafts CVs and cover letters. I built supervision into the data model: nothing is sent without the person's approval, every item carries a status and an authorship label (AI-generated, AI-edited, written by the person), and the pre-agent version is kept for rollback.
And a lesson from the build itself: an AI coding agent I worked with had more access than it needed and deployed to production on its own. Nothing was designed wrong in the product, but the guardrails around the agent were missing. I separated development from production, removed its credentials and made every release a deliberate human step. Least privilege isn't theory to me; I've seen what happens without it.
Where to start in an organisation
Start with low-risk internal tasks (knowledge search, drafting), keep a person approving anything that leaves the building, and measure before widening autonomy.