Agents that act inside real boundaries.
AI agents can complete multi-step work across information, tools, applications, and workflows. The value comes from what they can do — and the trust comes from what they cannot.
What an agent might do.
Depending on the task and the permissions it is given, an agent can carry work from the first piece of information to the point where a person decides — or, for well-understood and low-risk paths, all the way to done.
- Gather information
- Compare records
- Prepare a recommendation
- Generate a document
- Draft a response
- Update a system
- Route an exception
- Request approval
- Complete an authorised action
- Monitor a workflow
- Escalate an important issue
Software that can carry a task across several steps
Most automation follows a fixed path. An agent can work through a task that requires several steps, decisions, and tools — gathering what it needs, comparing options, preparing an outcome, and knowing when to stop and ask. That flexibility is genuinely useful, and it is exactly why the boundaries have to be explicit.
- Multi-step work across tools and records
- Access to the systems the task requires
- A defined job rather than open-ended latitude
- A clear stopping point and handoff
Permissions, visibility, and limits come first
An agent should operate with clear permissions, visible actions, and defined limits — with human approval wherever judgement or risk requires it. We decide what an agent may read, what it may change, what it must never do alone, and who is accountable for the outcome before it runs against anything real.
- Explicit read and write scopes
- Actions recorded and reviewable
- Hard limits on consequential steps
- A named human owner for every agent
Prepared by software, approved by a person
For anything that sends, spends, commits, schedules, or changes something hard to undo, the agent prepares the action and a person authorises it. Done well, this is faster than doing the work manually and safer than letting it run unattended — the reviewer sees a finished proposal with its reasoning and sources rather than a blank page.
- Full context attached to every proposal
- Approve, edit, or reject in one step
- Confidence thresholds decide what needs review
- Complete audit trail from trigger to outcome
Not every problem needs an agent
Agents are flexible; fixed workflows are predictable. When a process is well understood and rarely varies, a defined workflow is cheaper to build, easier to reason about, and less likely to surprise anyone. We reach for an agent when the work genuinely varies from case to case — and say so when it does not.
- Fixed workflow when the path is known
- Agent when each case differs meaningfully
- Often both, with clear division of labour
- The simpler option wins ties
Where does work stall waiting on a person to assemble it?
That gap — between information existing and someone having time to act on it — is usually where an agent earns its place.