Guide · 10 min read
How to Scope Your First AI Employee
The quality of an AI employee is mostly decided before any code is written. A clear scope makes it buildable, testable, and supervisable. A vague one produces a system that is impressive in a demo and frustrating in practice.
This guide walks through the scoping document we produce at the start of every AI employee engagement. You can use it on your own.
Step 1: Pick one role and one workflow
Choose a function where repetitive work is visible — sales development, support triage, operations reconciliation, invoice processing — and within it, one workflow. Resist the temptation to include everything the human role does.
Step 2: Write the role specification
Write it as you would for a capable new hire on their first week.
- Purpose: what this role exists to achieve
- Responsibilities: the specific tasks, in order
- Inputs: what arrives, from where, in what form
- Outputs: what 'done' looks like for each task
- Rules: the criteria and policies to follow
- Tone: how communication should sound, with examples
Step 3: Define tools and access
List every system the role touches and, for each, the specific operations needed: read contact, update lead status, create draft email, check calendar availability. Anything not on the list is not allowed. This becomes the agent's tool set and permission scope.
Step 4: Draw the approval lines
Sort every action into one of three lists.
- Handles alone — low risk, easy to reverse
- Asks first — consequential, external, or financial
- Never does — outside the role entirely
Step 5: Collect examples and define success
Gather real past cases with the correct outcome. Agree measures: approval rate without edits, turnaround time, escalation rate, and error types. Set the bar for moving a task from supervised to unattended.
Step 6: Plan the supervised pilot
Decide who reviews, how often, and through what interface. Agree how feedback is captured and how often the team meets to review progress. Define what would cause you to pause the pilot.
A one-page template
- Role name and purpose
- Workflow in scope (and explicitly out of scope)
- Systems and operations allowed
- Handles alone / asks first / never does
- Knowledge sources
- Success measures and autonomy criteria
- Reviewer and review cadence
- Owner after launch