AI Agent Readiness Review
How do I show my insurer or customer that my AI agents are under control?
You need two things you can hand over. A clear record of what each agent can reach, and one place that points to the records showing it’s controlled. We put both together for a fixed $2,500 in 1 to 2 weeks. You get a least privilege change list, a dated memo, and an evidence pack that points to your records instead of copying them. It’s readiness work. It isn’t an audit, a certification, a guarantee or security testing.
Start with what your agents can do, not what they’re told to do
Most teams can tell you what an agent is supposed to do. Fewer can show what it’s able to do. Which systems can it reach? What can it send or change without a person approving it? If it ignored its instructions, what would actually stop it?
That gap is where a careful reviewer looks. Written rules in a prompt are a good start, but a reviewer will ask what enforces them.
Then put the proof in one place
When a customer, auditor or insurer asks about your agents, they usually want to see the same few things. A list of the agents you run, with an owner for each. The access each one holds. How actions that leave your company get approved. How changes to an agent’s rules are tracked. And when all of that was last checked.
If those records exist but are scattered, the job is to point to them clearly and put a date on it. If some don’t exist yet, you want to know that before anyone else finds it.
That’s what this review does.
What you get
Access review
- For each agent in scope, we write down what it can reach, what its instructions say it must not do, and whether anything in your setup enforces that.
- You get a least privilege change list. Specific changes, ranked by how much they matter. Your team decides which ones to make.
- You get a dated memo, signed by the person who did the work, that you can share with a customer, auditor or insurer.
Evidence pack
- One dated pack covering the records that show how your agents are controlled: the agent list, access, approval before outbound actions, and change control for agent rules.
- It points to where each record lives. It doesn’t hold copies of your logs or your data.
- Gaps are marked plainly, so you see them first.
Coming later: behavior review
This would run your agents through agreed everyday and edge case tasks and check whether they stay inside their limits. It’s non adversarial. It doesn’t try to trick, attack or break anything. We’re not offering it yet, and it isn’t part of the fee.
Price and timing
$2,500, fixed.
It’s the same fee as our published Readiness Diagnostic. No hourly meter.
1 to 2 weeks
from the day we can read your configuration.
How the work happens
We read configuration. That means agent definitions, tool and permission settings, and any approval or guard rules. We don’t run your agents and we don’t change anything.
You can share those settings with us directly, with keys and passwords removed first. Or you can give us read only access in your own systems. If you do, it’s issued by you, scoped to this task, logged on your side, and removed when we’re done. We never take an owner, admin or billing role, and nothing gets exported.
All of the work is remote.
What this is not
- It isn’t an audit, an attestation or a certification. If you need a formal report, it has to come from an independent firm, and we’ll help you find one.
- It isn’t a guarantee. We can’t tell you what your insurer or customer will accept, and we don’t promise any outcome.
- It isn’t security testing. We don’t do penetration testing, red teaming, vulnerability scanning or monitoring. If you need testing, we’ll help you scope it and pick a vendor you hire directly.
- It isn’t a statement that your agents are safe or compliant with anything. It’s a dated read of how they’re set up.
Who does the work
Control and Function is a SOC 2 readiness practice. A named person does the work, writes the findings and signs the dated memo you can hand to someone else.
If we’re not the right fit for what you need, we’ll say so.
How we use AI
We use AI tools, including large language model assistants, for research, drafting and analysis. A person reviews, corrects and approves every deliverable before you get it, and we stay fully responsible for what’s in it. We don’t send your material to any AI service whose terms let it train on that material. If you’d like certain material kept out of AI tools entirely, tell us in writing and we will.
Questions people ask
Does SOC 2 cover AI agents?
SOC 2 doesn’t have a separate section for AI agents. But if an agent can reach systems that are in your SOC 2 scope, the same access questions apply to it as to any other account. Who approved its access, is that access limited to what it needs, and when was it last reviewed? This review gives you the answers in writing.
Will my insurer accept this?
We don’t know, and we won’t guess. Ask your insurer or broker what they want to see about AI agents. We’ll lay the evidence pack out in the form they ask for. The memo is dated and signed, and it says plainly what it is and what it isn’t.
Do you need access to our systems?
Not necessarily. Most of this can be done from settings you share with us, with any keys or passwords taken out first. If read only access is easier for you, you issue it, you scope it, and you remove it when we’re done.
Will you test our agents or try to break them?
No. This review reads how your agents are set up. It doesn’t run them, and it doesn’t attack them. If you need adversarial testing, we’ll help you scope it with a vendor you hire directly.
Does the pack line up with an AI control framework?
If a customer, auditor or insurer names a framework they use, we can add a cross reference to it. A cross reference shows where your records sit against their list. It doesn’t mean your agents meet that framework, and we won’t say they do.
What do we need to have ready?
A list of the agents you run and who owns each one. If you don’t have that list yet, building it is the first step of the review.
Who fixes the gaps?
Your team decides which changes to make and makes them. If you’d like help, we’ll scope it separately and agree on it before any work starts.
How is this different from a compliance platform or an AI security tool?
Tools collect data and flag issues. This is a person reading your setup, writing down what they found, and signing a dated memo you can hand to someone else. Plenty of teams use both.