AI Agents with Human Handover
An AI agent can prepare work across several steps, but it should not act on your business unchecked. I design agents for narrow, defined tasks, with limited permissions, a person approving what matters and a clear way to stop them.
New v6 draft from the service brief, awaiting review
The problem
- You have heard that AI agents can run whole processes, and you are unsure what is realistic or safe.
- A task involves several tools and judgement calls, so a simple automation does not cover it.
- You worry about an agent sending, deleting or spending something without anyone checking.
What I do about it
- I narrow the task to steps an agent can do reliably, and keep decisions with real consequences for a person to approve.
- I give the agent its own account with the least access it needs, and log every action it takes.
- I test it on realistic cases, write down how it fails, and build the stop and fallback before it goes live.
Draft copy, to confirm
What you get out of it
AI takes on the preparation work in a defined task, your team approves what matters, and you can see and stop everything the agent does.
Who it’s for
- You have a defined, repeated task that spans several tools, such as preparing a draft reply from a case history.
- You want AI to do the preparation and your team to make the final call.
- You are willing to start small, measure the results and widen the scope only on evidence.
If the task is predictable from start to finish, a plain automation is simpler and cheaper; see AI automation and integration. If you are still deciding whether to use AI at all, start with a digital strategy and roadmap.
What’s included
Use-case definition
One task, its inputs, its expected output and the decisions that stay with people, agreed in writing.
Permissions
A separate account for the agent, with read and write access limited to what the task needs and nothing more.
Approval points
The places where a person reviews and approves before anything is sent, changed or published.
Evaluation
A set of realistic test cases, including hard and unusual ones, run before launch and again after changes.
Failure paths
What the agent does when it is unsure or a tool fails, how your team is told, and the manual way to finish the task.
Monitoring and ownership
Logs of what the agent did, a regular review, a named owner on your side and a simple way to switch it off.
How it works
Define
We choose one task and agree its boundaries, the approval points and what success looks like.
Prototype
I build a limited version on test data, with every action logged and nothing sent without approval.
Evaluate
We run the test cases, look at where it fails and decide together whether it is good enough to use.
Pilot and review
It runs for a period with close review. We widen its scope only if the evidence supports it.
Questions people ask
Draft answers, to confirm
Can an agent run a whole process on its own?
I do not build agents that act without oversight. I design them for narrow tasks, with a person approving any step that sends, changes, spends or publishes something.
How do you know the agent works?
We agree test cases before building, run them before launch and again after any change, and review real results during the pilot. If it does not meet the agreed standard, it does not go live.
What does it cost to run?
AI providers and connected tools charge you directly, usually by usage. I explain the likely running costs before we build, alongside the time it is expected to save.
Who owns the agent?
You do. It runs on your accounts, with documentation, and a named person on your side is responsible for it.
Tell me what you need to achieve.
We can work out the right next step before committing to a solution. You can book this on its own. An ongoing partnership is optional, never a requirement.
Need continuing responsibility rather than a defined project? Ongoing partnerships