RAISE.AI
THE RAISE METHOD · THE AGENT RUNG

Agentic Control Scorecard

Twelve questions. Four minutes. One honest number for your board, sent to your inbox with your answers, question by question.

When an assistant is wrong, you ignore it. When an agent is wrong, you have to undo it.

Most organisations are switching agents on the same way they switched on Copilot: one team at a time, faster than anyone can write down what each one is allowed to do. An assistant that gives a wrong answer is embarrassing. An agent that takes a wrong action has changed something, and someone has to explain it.

These twelve questions are what the RAISE Method requires at the rung where AI stops advising and starts acting, so your score here is measured the same way a full assessment would measure it.

This is about agents acting in your corporate systems: the ones your staff build, the ones you buy, and the ones switched on inside software you already pay for. It is not about agents in operational or safety-critical systems, which are a different problem with different controls.

Part of the RAISE Method. Read how a RAISE assessment runs.

How to score. For each question: 0 = no, or we don't know · 1 = it is agreed, or written, or approved, or in place for some agents, but an agent could still get past it · 2 = in place for every agent that acts, and an agent cannot get past it. "We don't know" is a real answer and scores 0; guessing is the only unhelpful option.

A policy an agent cannot read is not a control. On this rung the difference between 1 and 2 is whether the thing is switched on.

Answer these about every agent that can take actions: the ones your staff built, the ones you bought, and the ones switched on inside software you already pay for. If you are not certain that list is complete, that is your first finding, and the AI Exposure Scorecard is where it gets fixed.

1

Has your leadership group formally decided, and written down, that AI may take actions on the organisation's behalf, not only give advice?

When the answer is no: A team sets an assistant to send supplier reminders by itself. Nobody can find the meeting where that was agreed, because there was not one.

2

Does every agent that takes actions run under its own identity, separate from any person's account?

When the answer is no: An agent runs under the account of the person who set it up. She leaves in March. The agent carries on, as her.

3

Does each agent hold only the permissions its task needs, and nothing more?

When the answer is no: The agent was given the same access as the finance team to save time. Its task was to file receipts. It can also approve them.

4

Does every agent have one named human owner, accountable for what it does, and is that name still current?

When the answer is no: An agent starts getting things wrong. The register says it is owned by IT. Nobody in IT knows it exists.

5

Is there an approved list of the systems and connections an agent may reach, with anything not on the list blocked?

When the answer is no: A new connection to a file-sharing service arrives switched on. An agent starts saving records there. Nobody decided to allow it; nobody looked.

6

Is everything an agent reads, including documents it retrieves and results it gets back from tools, treated as untrusted before it can lead to an action?

When the answer is no: An agent reads an inbound email that politely asks it to forward the thread to an outside address. It does, because nothing in the email looked like an attack.

7

Is there a written list of actions an agent may never complete without a person, with no way round it?

When the answer is no: An agent changes a supplier's bank details because the request looked routine. Nothing stopped it, because nothing was told to.

8

When a person approves an agent's action, do they see the exact action about to be taken, and is that approval recorded?

When the answer is no: A person approves "update the record" without seeing which record or what the update was. Three weeks later nobody can say what they approved.

9

Are the people who approve agent actions trained and checked for that judgement, rather than simply given the role?

When the answer is no: The approver is the most senior person available. He approves the action because it looks reasonable. Nobody has ever shown him what a bad one looks like.

10

Does your incident plan say what to do when an agent has taken a wrong action, including how to undo it?

When the answer is no: An agent sends the wrong pay figures to forty staff on a Thursday. The incident plan covers data breaches. It does not say how to take back an action already taken.

11

Has the route for stopping one agent, without stopping everything else, been tried by a named person before go-live?

When the answer is no: The stop button for one agent is in a runbook. The first time anyone presses it, every agent stops, and so does the team that depends on them.

12

Is there a hard ceiling that stops an agent that runs away, rather than only raising an alert?

When the answer is no: An agent gets into a loop at two in the morning. By nine it has spent the month's allowance, and the assistants other teams use have gone quiet.

Two questions are scored a little differently. Question 1 is a decision, and a decision cannot be enforced: 2 means it is recorded with a date, 1 means it was agreed in conversation but never minuted. Question 9 is marked one level above what the method requires at this rung, on purpose: the briefing that follows will hold you to "checked" rather than "briefed", so the score does too.

Where should we e-mail your result?

Your score, the band it falls in, and your twelve answers, with the three lowest marked as the ones to raise first.

Your answers and contact details go to RAISE.AI (Gembitz Ltd) so we can send your result. We do not share them with anyone.

About the RAISE Method

A reference model of 60 capabilities an organisation must hold to run AI safely, each traced to a recognised standard (ISO/IEC 42001, the international AI management standard, and NIST's AI Risk Management Framework among them); a nine-rung ladder from foundations to autonomous agents, with a gate before each step; and generated board and technical plans that your board and your IT team can work from.

Built on twenty years of Microsoft enterprise architecture and information governance work. The current version was applied at a large UK charity through 2026. Delivered by RAISE.AI (Gembitz Ltd, est. 2017). Book an intro call or write to ossama@raiseai.co.