Superintelligence in Regulated Industries: Compliance Before Capability
Regulated companies should set compliance, records and human approval rules before more capable AI agents arrive. A practical preparation plan.

Superintelligence refers to AI that could, in time, outperform people across most cognitive tasks. For a regulated company, the practical question is not how quickly that arrives but whether compliance rules, records and human approval points are in place before a more capable agent gets access to customer data, money or decisions. Compliance before capability means deciding who may approve what, which evidence the agent must leave behind, and how a person can stop it, then building the agent inside those limits.
Key takeaways
- Set approval, record-keeping and stop rules before you widen what any agent can do.
- Treat each agent as a defined role with a named owner, limited memory and a short list of tools.
- Keep evidence of permissions, decisions and changes in a form your compliance team can read without help from the vendor.
- Require a named person to approve actions that move money, change regulated records or speak for the company on decisions.
- Confirm sector and jurisdiction obligations with qualified advisers. This article is operational guidance, not legal advice.
Why compliance has to come before capability
Companies in finance, health, insurance, education, legal services and public-facing work already carry duties to explain decisions, protect personal data and show who was responsible for an outcome. A more capable agent does not remove those duties. It enlarges the surface where a duty can be missed, because an agent that reads, drafts, decides and acts across several systems can create obligations faster than a review team can follow them.
Waiting for a capable agent before writing rules means writing them under pressure, with the agent already wired into daily work. Staff will have adopted it, managers will have asked for exceptions, and the vendor's default settings will have become the de facto policy. Writing the rules first gives you a standard the agent must meet, instead of a set of decisions made by whoever moved fastest.
The NIST AI Risk Management Framework is useful here because it frames AI risk as ongoing work: govern, map, measure and manage. Those questions match what an auditor or regulator will ask: what is the system for, who is accountable, how is it tested, and how is it watched after launch.
What changes when an agent acts instead of answering
A chatbot that answers questions mainly creates content risk, such as a wrong statement or a leaked sentence. An agent that opens a mailbox, edits a customer record, sends a message or triggers a payment creates action risk. Action risk needs permissions, logs and stop paths, and it needs them before the first live step.
The OWASP Top 10 for LLM Applications names several risks that matter when agents have tools, including prompt injection and excessive agency. Excessive agency describes a system granted more capability than its task requires. The useful question is rarely whether the model is capable enough. It is whether the agent has more access than its job needs, and whether anyone would notice if it used that access.
For this reason, capability growth should be treated as a request to change permissions, not as a reason to change the policy in advance. Each new ability goes through the same review: what it can touch, what evidence it creates, who approves it, and how it is switched off.
Which records you need before the agent goes live
Records are what let you show what happened and why. Before launch, decide what the agent records for every action: the request, the sources it used, the tool calls, the output, the approver if any, and the final result. OpenTelemetry's observability concepts describe how traces connect steps so one event can be followed across components. The same idea applies to an agent: a single customer request should be traceable from input to action.
Records also contain personal data, so retention and redaction must be decided at the same time as logging. Keeping everything forever is not a compliance strategy, and deleting everything at once makes investigations impossible. Write the retention period for each record type, and name who can read them.
Before launch, your records file should answer these questions:
- What is the owner and purpose of each agent, in one plain paragraph?
- Which data sources can it read, and which must it never touch?
- Which tools can it call, and at what permission level for each?
- What is logged for every action, and where are those logs stored?
- How long are logs and agent memory kept, and who can delete them on request?
- Who reviews blocked or failed actions, and on what schedule?
Where a person must approve, and where they only review
Approval is not a single gate placed at the end of a workflow. It works better as tiers attached to tools. Reads and drafts seldom need approval. Reversible internal updates may only need review after the fact. Actions that spend money, change a regulated record, contact a customer about a decision, or affect a health, financial or legal outcome need a named person's approval before they execute.
A simple decision guide helps staff apply the tiers consistently:
- Read only, with no external effect: allowed, logged and sampled for review.
- Draft for a person to send: the person sends, and the agent never sends.
- Internal record change that can be reversed: the agent acts, and the owner reviews the daily log.
- External message, payment or regulated record change: a named approver signs before execution.
- Anything the agent cannot classify: it stops and asks.
The approval design should also specify what a signer is expected to check. A signature that only confirms someone clicked a button provides weak evidence. Show the source, the proposed action and the reason in the approval view, so the signer's decision is informed and recorded. For a deeper treatment of where people sign, see Human Approval Gates for Superintelligent Agents: Where People Sign.
How Be Human handles this: in a managed engagement, Be Human first defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, before the agent acts on anything live. Operations then include monitoring, recurring evaluations, permission reviews and incident response, so the rules are tested against real activity rather than sitting in a policy document. Permission changes are made in controlled steps and recorded.
A preparation sequence for a regulated team
The following sequence keeps the work small enough to finish and gives you evidence at each step:
- Step 1: Inventory where AI already touches work, including tools staff adopted without a formal request.
- Step 2: Choose one regulated process where an agent could save effort, and name the person who owns that process.
- Step 3: Write the role in one page: purpose, inputs, outputs and forbidden actions.
- Step 4: Classify every action the role could take into the approval tiers above.
- Step 5: Define the logs, retention periods and the reviewer for blocked or failed actions.
- Step 6: Build a stop path and test it with the owner present, so everyone knows how to pause the agent without losing its logs or queued work. The mechanics are covered in AI Agent Kill Switch: Halt Actions Without Losing Logs or Queued Work.
- Step 7: Run the agent in draft mode, compare results with the success criterion, and widen access only in small, recorded steps.
Each step produces a document or a test result. Those artifacts are what your compliance file will need, whatever the final rules turn out to be.
Illustrative scenario: a claims intake agent
The following is an illustrative scenario, not a description of a real company. A small insurance broker wants an agent to sort incoming claim emails by type and urgency. In the first phase, the agent only labels and summarizes messages, and a staff member checks a sample each day. Its access to the policy system is read only.
When the team wants the agent to pre-fill claim fields, the rules require a staff member to confirm each field before it is saved, and every change is logged with a link to the source email. A proposal to let the agent send acknowledgment messages stays blocked until the team has reviewed the blocked actions and agreed on the wording with its compliance adviser. The agent never decides coverage. That decision stays with a licensed person, and the record shows who made it.
The scenario shows the pattern: the capability grows only where the controls already exist, and each expansion is a recorded decision.
Preparing without making legal claims
Compliance preparation should describe controls, not declare that a tool is compliant. Avoid statements such as calling a system approved or certified unless you hold the document that supports the statement. Instead, write what data flows where, which rules you believe apply, who approves which actions and what evidence exists. Obligations differ by sector, location and the kind of service provided. Brazilian companies should review data protection duties with the national authority guidance in mind, and the points in LGPD and AI Agents: What Brazilian Companies Should Check are a practical starting place. Companies serving users in the European Union should check the scope of the [EU Artificial Intelligence Act](https://artificialintelligenceact.eu/) with counsel.
Keep one register that lists each agent, its purpose, its risks and its controls. A short, maintained AI risk register lets a non-specialist see what exists and what is still open, which is often the first thing a reviewer asks for.
The OECD AI Principles offer a shared vocabulary for transparency, accountability and human oversight, and they can help a leadership team agree on terms before arguing about specific tools. Whatever framework you use, the test is the same: can you show, for any action the agent took, who allowed it, what it saw, and how it could have been stopped?
If you are unsure where your first agent should begin, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Does superintelligence affect regulated companies today? The more useful way to answer is that today's agents already create the governance questions that more capable agents will make harder. Permissions, records, approvals and stop paths are needed now for systems that act on company data. Preparing those controls early means the company is not forced to design them during an incident.
Do we need to wait for regulators to publish rules before setting our own? No, and waiting carries its own cost, because agents tend to spread through workflows before policy catches up. Your internal rules can be written against your current obligations and updated as sector guidance and laws change. Keep a dated record of each version so you can show what you knew and decided at each point.
Who should own AI approval rules inside a company? Ownership usually sits with a named executive, with compliance, operations and the business owner of each process taking part. The executive is accountable for the policy, while each process owner approves the role and the tools it uses. A small committee with a written charter can keep the arrangement clear.
Is a human approval step enough to make an agent safe to use? A human approval step is necessary for high-impact actions, but it is not sufficient on its own. The approver needs enough context to decide, the logs must show what was approved, and the permissions must prevent the agent from acting outside the approved path. Approval works best as one layer alongside limited access, monitoring and a tested stop path.
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.


