01Stakes & audience
Who relies on the output or action, and what happens if it is wrong?
Personal note vs. student advising vs. financial decision
05
Six lenses to ask of every agent
The strictest answer sets the control tier. These lenses are reapplied in every scenario. Use the worksheet below to score a real agent.
02Data & reach
What data does it touch, and what systems or actions can it take?
03Autonomy & authority
Does it only recommend, or can it initiate workflows, write data, or send communications?
04Accountability & ownership
Who owns the agent, approves changes, and is responsible when it errs?
05Verification & monitoring
How will we know it remains correct, and how quickly can we detect and stop problems?
06Lifespan & scale
Is this a short pilot or a durable institutional capability? How many users or decisions will it affect?
Score an agent
Choose the answer that fits best for each lens. The strictest answer sets the baseline tier. Your answers stay in this browser tab: nothing is saved, sent, or tracked, and they clear when you leave the page.
—Answer all six lenses
0 of 6 answered · Compare all tiers ·
How to read the tier
The tier is a starting point for a decision. It has not been validated as a risk assessment. Use judgment where the score falls short.
- Read each lens as two questions. Data sensitivity is not the same as write access. How bad an error would be is not the same as how likely it is. Note both.
- Approval before and review after are different controls. An agent that changes records or sends messages on its own needs approval before it acts. Reviewing a log afterwards is not enough.
- Some controls apply at any tier. Anything touching student records, financial aid, academic standing, or health information gets an owner, a log, a way to switch it off, and a privacy review.
- Write down overrides. If expert judgment moves the tier up or down, record who decided and why in the registry.
Help test this model: if a real case scores wrong, send it through the feedback form below.