The charter is written when you scroll here.
← All tools
Agent Pilot Planner
A pilot leadership says yes to
Scope a 90-day AI agent pilot in six short steps. Set the baseline, the target and the guardrail. Decide what the agent does alone and what a person approves. Leave with a 12-week plan, a one-page charter and the go, adjust or stop rules agreed before the pilot starts.
Wk 2
Wk 4
Wk 6
Wk 8
Wk 10
Wk 12
Target
Baseline
Build
Shadow run
Supervised
Decide
Go decision
Go decision in week 12: the shadow run found data gaps and added a week
Sample dataStart from a sample
Scope the pilot in six short steps
Pick a sample to see a finished plan, then make it yours. Each sample leaves a gap or two on purpose, so you can see the checks react. Everything updates as you type.
Step 1 of 6
The workflow and the agent's job
One sentence each. If you need a paragraph, the pilot is too big.
Why write down what is out of scope
A pilot grows the moment it looks promising. Writing down what the agent will not touch keeps the 12 weeks honest, and it tells risk and audit where the edges are before anyone asks.
Step 2 of 6
Who owns it, and who approves the agent
Roles, not names. The charter goes around the building.
Why name an approver
The owner answers for the result. The approver is the person who says yes or no to each action the agent proposes during the supervised weeks. Without one, either nobody checks the agent or everybody does, and neither works. This is the human in the loop.
Step 3 of 6
Measure today before the agent arrives
Your own numbers from the last month or quarter. A rough figure you can defend beats a precise one you can't.
Why a baseline comes first
Without today's numbers, week 12 turns into a debate about impressions. With them, it is a comparison. Measure the same way you will measure during the pilot, from the same report, so the before and after are like for like.
Step 4 of 6
Decide what success means, in advance
One metric that should improve, and one guardrail metric that must not get worse.
Why a guardrail metric
Most metrics can be gamed by accident. An agent can cut cycle time by closing cases badly. The guardrail is the number that catches that: if it moves the wrong way, the pilot pauses, whatever the success metric says.
Step 5 of 6
What the agent does alone, and what a person approves
List the steps of the work, how much of the effort each takes, and who does it once the pilot is supervised. Mark the steps that are key decisions.
Why not automate all of it
In our experience the last fifth of a workflow is where the judgement lives: unusual cases, money leaving the building, anything a customer or auditor will see. We recommend capping what the agent does alone at around 80% and keeping a person on every key decision, at least until the pilot has earned more.
Step 6 of 6
Risks, and the data the agent needs
Name what could go wrong and what you will do about it. Then list every system the agent reads or writes, so access is requested in week 1, not week 6.
Risks and mitigations
Data and system access
Why read access first
A shadow run only needs the agent to read. Write access can wait until the supervised weeks, and even then it should be the narrowest permission that does the job, with every action in the audit trail.
Step 1 of 6
Your pilot Sample data
Ready to approve0%
Checks passed0/0
Target change–
Agent alone–
DecisionWeek 12
12-week timeline
Earn autonomy one phase at a time
Use the minus and plus buttons to move weeks between phases. The total stays at 12, with the decision in the last week.
The timeline draws when you scroll here.
Why a shadow run before the agent acts
In a shadow run the agent drafts every case while people do the work as they do today. You get a side-by-side comparison on real volume with nothing at stake. Only once its drafts hold up does it start acting, with a person approving.
Readiness checks
What leadership will ask before saying yes
Each check is a gap we see stall pilots. The button takes you to the field that fixes it.
Decision criteria
Go, adjust or stop, agreed now
Written from your metric, target and guardrail. Agree them before week 1, so week 12 is a reading, not a debate.
What this means
Myth
A pilot proves the technology works.
Reality
The technology usually works in a demo. A pilot proves something harder: that this team, on this workflow, with these controls, gets a better result. That is why it needs a baseline, an approver and a decision date, not just a working agent.
Field note
In our experience the pilots that stall are rarely the ones where the agent failed. They are the ones where nobody agreed what success looked like, so week 12 became a meeting about opinions. Write the go, adjust and stop rules first, and the decision makes itself.
Field guide
Three lessons on agent pilots
What an owner, a controller or a CIO should settle before an agent touches live work.
Why pilots need a baseline
"It feels faster" doesn't get a budget approved. Measure volume, cycle time, error rate and hours before the agent arrives, from the same report you will use in week 11. Then the result is a comparison anyone can check, not a claim.
Rule of thumbIf you can't say today's number, spend the first two weeks measuring it.
Shadow runs before autonomy
Let the agent draft while people still do the work. Compare the two on live cases. Then let it act on routine cases with an approver, and widen what it does alone only as its record earns it. Each phase is a gate, and a person decides whether to pass it.
Rule of thumbDraft, then act with approval, then act alone on the routine work. Never skip a phase.
Decide what success means in advance
Write the go, adjust and stop rules before week 1, with numbers. Once people have seen the results, the bar tends to move to wherever the results landed. A guardrail metric keeps the success metric honest.
Rule of thumbOne metric to improve, one guardrail that must hold, one person who decides.
Want help running the pilot?
We scope, build and run agent pilots with finance and operations teams, with a person on every key decision. Tell us the workflow.