Skip to content
be-human.ai
← Back to the journal

Human oversight

Superintelligence Risk Register for Company Agents: A One-Page Template

A one-page superintelligence risk register for company agents, with owners, likelihood notes, mitigations and a review cadence tied to model changes.

Be Human typographic cover with the title: Superintelligence Risk Register for Company Agents: A One-Page Template
Conceptual illustration created with AI · Be Human

A superintelligence risk register for company agents is a one-page table that lists each agent, what it can do, the harm it could cause if it fails or is misused, the person who owns that risk, a plain note on how likely the harm seems, and the controls that limit it. The register is only useful if it is reviewed on a fixed schedule and again whenever the model, tools, memory or permissions change. The template below is written for a business owner or operations leader who needs to decide what an agent may do this quarter.

Key takeaways

  • Keep one row per agent, with one named owner and one next review date on every row.
  • Write likelihood as evidence and trigger conditions, not as precise odds you cannot defend.
  • Choose mitigations you can test, such as narrower permissions, approval steps, readable logs and a pause path.
  • Re-score a row whenever the model, tools, memory scope or data it touches changes, not only once a year.
  • Hold the register to one page so people will read it and keep it current.

Why a risk register belongs in agent operations

Most operations teams already keep some form of risk list for payments, staffing or customer data. Agents add a different kind of exposure. They take actions across several systems, and the set of things they can reach often grows faster than the process that first approved them. A register turns that growth into a visible list with a named person responsible for each item. The NIST AI Risk Management Framework describes an ongoing cycle of governing, mapping, measuring and managing AI risk, and a register is the simplest working document that makes that cycle concrete for a company. For a broader view of how these risks appear in daily work, see AI risk register template for business, which covers the general case.

The register is not a forecast about how capable future systems will become. It answers narrower questions: what does this agent touch today, what would a bad outcome cost the business, who decides when the answer changes, and which controls have actually been checked. Keeping the scope that narrow is what allows the document to be finished in one sitting and kept alive afterward.

The one-page template, column by column

Use a spreadsheet or a single document table with these columns, in this order.

  • Agent and role: the agent name and one sentence describing the single job it performs.
  • Actions it can take: read, write, send, pay, delete or publish, and which system each action touches.
  • Harm if it goes wrong: the concrete outcome for a customer, an employee, the books or the brand.
  • Risk owner: one named person with authority over the harm described, not a team alias.
  • Likelihood note: two or three sentences of evidence, described in the next section.
  • Current controls: the permissions, approval steps, monitoring and pause paths in place today.
  • Planned mitigations: the next two or three changes, each with a due date and a tester who did not build it.
  • Re-review triggers: the events that force a re-score, listed in the final section.
  • Next review date: a calendar date no more than one quarter away.
  • Status: open, in progress or accepted, with the name of the person who accepted any residual risk.

Keep one row per agent. If one agent performs two jobs with different permissions, split it into two rows, because the controls and the owners are usually different.

How to write a likelihood note without fake precision

Business leaders are often asked for a probability, and the honest answer is usually that the company does not yet have enough history to estimate one. A likelihood note replaces the number with evidence. It says what has been observed, what has not been tested and what would make the harm more likely. Assign a rating of low, medium or high only after the note is written, so the rating follows from the reasoning rather than the other way around.

Example note: the agent has drafted replies during the pilot with no wrong-recipient sends recorded, but it has not been tested against messages that mix two customer accounts or arrive in a second language. Inbound email is untrusted text, so a message could contain instructions the agent should not follow. Rating: medium, pending a test on mixed-account messages.

Raise the rating when a new channel, tool or data source is added, when the underlying model is replaced, or when the agent reads text from outside the company. Lower the rating only when a test has been run and its result is recorded in the note.

Building the first register in one working session

A first version can be produced in roughly ninety minutes with the business owner, one operations lead and the person who maintains the integrations. Work through these steps in order.

  • List every agent or assistant with access to company systems, including tools that staff set up themselves. Finding those hidden tools first is covered in shadow AI in the workplace.
  • For each one, write the single job it performs and the exact actions it can take.
  • Name the risk owner for each row before discussing risk, so accountability does not drift into the group.
  • Describe the worst credible outcome in business terms: a wrong refund, a message sent to the wrong customer, a file deleted, a record altered.
  • Write the likelihood note from evidence you actually hold, and mark any row where the evidence is missing.
  • Check each existing control against the system itself, not against the policy document that describes it.
  • For each high-rated row, pick two mitigations and assign each one to a tester.
  • Set the next review date and put it on the calendar before the meeting ends.

Illustrative scenario: a customer follow-up agent

Illustrative scenario, not a real company. A distributor uses an agent to answer order status questions and draft follow-up emails for its sales team. The register row names the sales operations manager as owner. The harm column notes two outcomes: a status message that reveals one customer's order to another customer, and a promised delivery date the warehouse has not confirmed.

The current control is that the agent can read order records but cannot send email, because a staff member must approve each draft. The planned mitigations are to restrict the agent's order access to the requesting account and to build a test set of messages that mix account names. The likelihood note cites the pilot results, states that the mixed-account test has not run, and flags inbound email as untrusted. When the warehouse adds a second carrier, the row is re-scored that same week, and the send permission stays off until the new test set passes.

Mitigations that reduce exposure, not paperwork

A mitigation only counts if it changes what the agent can do or how quickly people notice a problem. Four kinds are usually worth the effort: narrower permissions granted per action and per system, human approval at steps where the harm is high, complete logs that staff can read, and a pause path that stops new actions while keeping queued work intact. The OWASP Top 10 for LLM Applications lists prompt injection and excessive agency among the main risk classes, so any row for an agent that reads external text should carry a mitigation that limits what injected instructions could cause.

Test each mitigation by asking concrete questions:

  • If the agent reads a crafted instruction inside a customer message, can it still send anything outside the approved path?
  • If the pause path is used at midday, what happens to work the agent has already queued, and who is told?
  • Can a manager who was not involved see which action the agent took, on what data, and why?
  • If a permission was removed last month, can you show that the agent no longer holds it?

If the answer to any question is uncertain, the row is not yet controlled. For a practical pause design, see kill switches for superintelligent agents.

Reviewing the register as model capability grows

The register is a living document. Re-score a row when a trigger occurs: a new model version, a new tool or connector, a wider memory scope, a new data source, a new customer segment or jurisdiction, an incident, or a change of owner. Each trigger should reopen the likelihood note, not only the rating, because the evidence is what changes. Evaluations are how evidence accumulates. The OpenAI evals guide describes building tests that measure model behavior against defined criteria, and a test of that kind is what turns a likelihood note into something more than an opinion. Autonomy can move in both directions as well, so a re-score may justify reducing an agent's permissions. The principle of granting each step only after evidence is developed in autonomy levels for superintelligent agents.

Be Human designs, implements and operates personalized AI infrastructure for companies, and the register belongs to that operating work. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so every row in the register points to a current control and a person who reviews it. The service is remote and available in English and Portuguese.

If you are unsure which agent task deserves the first row in your register, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

How often should the register be reviewed? At least once per quarter, with a calendar date on every row. Review earlier whenever a trigger listed in the register occurs, such as a new tool, a model change or an incident. A review of rows that have not changed is usually short, which is why a fixed cadence is easier to keep than an occasional deep review.

Who should own the register? A named executive or operations leader should own the document and keep its format consistent. Each row still needs its own risk owner who has authority over the harm described. Assigning the whole register to IT alone is a common mistake, because the harm usually lands on operations, customers or finance.

Is a risk register enough for regulatory compliance? No. A register supports governance, but it does not replace legal review. Depending on where you operate and what personal data the agent handles, rules such as the EU AI Act or Brazil's LGPD may apply, so check the requirements with counsel. The points raised in LGPD and AI agents for Brazilian companies are a reasonable starting checklist.

What if we cannot estimate likelihood at all? Write down the evidence you have and state clearly where it is missing. Mark the row as needing a test, and schedule that test before the next review. A missing test is a legitimate finding, and it is often the most useful line in the whole register.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →