verifiedagents.ai
All posts

5 min readAI Agents · Cybersecurity

How do you score an AI agent approval request?

By Michelle Savage, Co-author, Agentic AI + Zero Trust

Hero: How do you score an AI agent approval request?

TL;DR: Score an AI agent approval request on five questions, one for each ATF element. Give two points for a specific written answer and one for a vague one. No answer gets zero. The total out of 10 decides how much freedom the agent gets. Approval stops being yes or no and becomes how much.

"Approve" is a funny word. Everyone uses it. Almost nobody has written down what it means.

In Josh's client calls, one security leader said it out loud. Everybody tells him security needs to approve the agent. Then someone knocks on his door with one, and he doesn't know what approving it would mean. He has no test and no list. He has a line of people waiting and a lot of goodwill.

A scorecard fixes that. It's one page.

Why score a request when you could just decide?

Because deciding takes a meeting. Scoring takes a form.

Without written criteria, every request becomes a debate. The debate goes to a committee. The committee takes months. So teams stop asking and build the agent anyway, somewhere nobody's watching.

A score also gives the requester something useful. They can see exactly what would earn their agent more room. More on what approval should mean is on MassiveScale.AI, in what it means to approve an AI agent.

What are the five questions on the scorecard?

One for each ATF element. Each one lines up with a question on the free assessment, so the two tools speak the same language.

Scorecard question

ATF element

Closest assessment question

Who is it, and who owns it?

Identity Management

Question 6: Who is responsible for each AI agent?

How will we see what it does?

Behavioral Monitoring

Question 7: How are agent actions logged?

What data goes in and comes out?

Data Governance

Question 17: How is data classified for agent access?

What is it allowed to do?

Segmentation

Question 20: How are agent actions constrained?

What happens when it goes wrong?

Incident Response

Question 25: Can you immediately stop a misbehaving agent?

How do you score each answer?

Each question gets up to two points. Here's what earns each.

Scorecard question

2 points

1 point

0 points

Who is it, and who owns it?

Its own identity, and one person's name

Its own identity, owned by a team

It runs on a person's login

How will we see what it does?

Every action logged with inputs and outputs

Some activity in general app logs

No logging

What data goes in and comes out?

Sources and outputs listed, with a label on anything sensitive

A general description of the data

"Customer data" and nothing more

What is it allowed to do?

Named systems and named actions, like "read contacts, no export"

Named systems only

"Access to the CRM"

What happens when it goes wrong?

A named person can stop it, and they've tried

A written shutdown process

"We'd notice eventually"

A requester can fill this in over an afternoon. If they can't, you've found the real problem without holding a meeting.

What does the total score mean?

Add the five scores. The total sets the agent's freedom.

Total

What the agent gets

0 to 4

Not yet. The requester gets the scorecard back with the zeros circled.

5 to 7

A narrow agent. A person approves each action.

8 or 9

More room. The agent acts by itself on low-risk work, and a person approves the rest.

10

The widest scope the job needs, with the usual checks.

One rule sits on top of the math. A zero on the last question caps the agent at the narrow level, whatever the total. An agent nobody can stop doesn't get room.

Those levels follow the same ladder Josh describes in what evidence proves an AI agent is ready to move up a level.

Who decides what?

Security scores the controls. The business owner accepts the risk.

Mixing those two jobs is what creates the pileup. Security's answer should sound like "these controls are met" or "here's what's missing." It shouldn't sound like "you may not do this," unless a hard policy line is crossed.

Writing that split down ends a lot of arguments before they start.

How do you keep the scorecard from becoming a new bottleneck?

Give small requests a fast lane.

Some agents don't need a human decision at all. They use no production data and have no write access. They never contact a customer and can't spend money. Write that down as a class, and let those requests approve themselves.

Then publish the scorecard where builders already look. A wiki page nobody reads doesn't count.

Frequently asked questions

What about agents that are already running?

Offer an amnesty. Give people a window to register agents with no penalty, then score them. You'll find more than you expected.

Should we block AI agents until the scorecard is ready?

No. Blocking moves the work to personal accounts, and you lose what little view you had. A rough scorecard this week beats a perfect one next quarter.

How do we know the scorecard is working?

Watch two numbers. One is days from request to decision. The other is how many agents never came through the process. If the second keeps growing, people are avoiding you.

Is the score the same as an audit?

No. The score is a gate at the start. An audit asks for proof later. Josh lists that proof in what evidence passes each of the five AI agent audit questions.

Key takeaways

  • "Approved" needs a written meaning. A scorecard gives it one.

  • Five questions, one per ATF element, zero to two points each.

  • The total out of 10 sets how much freedom the agent gets.

  • An agent nobody can stop stays narrow, whatever it scores.

  • Security scores the controls. The business accepts the risk.

For a baseline before you write your page, take the free assessment. It asks 30 questions across the same five elements and takes about ten minutes.

Until "approved" means something specific in your company, security will keep getting handed a decision nobody defined. One page fixes that.

See where your agents stand.

The free assessment takes ten minutes and scores you on the five elements of the Agentic Trust Framework.