Before you build an AI “employee,” write its job description. This turns each role into a one-page card: the single outcome it owns, what it can read and do, how much freedom it has, who answers for it, when it hands off to a person, and how you switch it off. It checks every role against a compliance gate and tells you what to fix first.
Each AI employee gets a role card. You fill in the parts; the worksheet checks the card against a compliance gate and a freedom ceiling, then tells you what to fix before anything is built.
The worksheet sets a ceiling on the level for each role, based on what it touches:
Giving an AI a narrow job, a limited level of freedom, and a named person who answers for it isn’t caution for its own sake. It follows how automation, risk management and medical regulators already frame the problem.
The designer turns those into one rule: give each AI employee a narrow job, a freedom level it has earned, and a named person who answers for it.
Illustrative composites — not any one practice’s setup. They show how the same card changes as you add roles. Use them to calibrate your own below.
One role, one owner, level 2. The aim is to learn what “good” looks like before the role has any freedom.
Several roles now share inputs. The risks are two roles answering the same message, and one role’s mistake feeding another.
At level 3 you stop watching each action, so the review sample and the pause switch carry the weight.
This is where you design your own. Three example roles are already filled in, so you can see finished cards — one ready, one fixable, one blocked — in a single tap before you change anything.
Illustrative design tool for strategy use — not legal or compliance advice. Confirm BAA and patient-information handling with your own counsel before building anything that touches patient data. · The AI Employee Role Designer™ · Dr. Renée Allen Consulting
1. One job per employee. Write the job as a single outcome: “Reply to every new inquiry within the hour.” If you need two outcomes, you need two roles — and two owners. The narrower the job, the easier it is to check and to switch off.
2. Start lower than you think. Begin at level 1 or 2 even where the ceiling is higher. A role earns a higher freedom level after several weeks of review in which you changed nothing important.
3. The hard stops are the real job description. What a role must never do matters more than what it does. Choose the stops before you build, and write them into the instructions you give the AI in plain words.
4. Every role has a person’s name on it. Not “the team.” One owner who reviews its work, can pause it, and answers for it. In a clinical practice that accountability stays with you and your clinicians.
5. Plan the pause before you launch. Write down exactly how you would switch the role off, in under a minute, from a phone. If you can’t describe it, the role isn’t ready.
6. Re-run the card whenever something changes. A new data source, a new action, a new vendor. A card describes the role you built; the moment it reads something new, it’s a different role.
Bring your role cards to our next session. We’ll build the first one in your own Claude, at the level it has earned.