Regulated Industries and Superintelligence: Set Compliance Before Capability
Regulated firms should set compliance controls for their company AI layer now, so more capable agents inherit the rules instead of bypassing them.

Regulated companies should set compliance controls for their company AI layer now, before agents become more capable. The controls that matter are the ones that travel with each agent: its named role, the memory it may keep, the tools it may call, and the actions that need a person to approve them. Agents built on these controls inherit the rules, while agents deployed without them can only be governed through manual review and good intentions.
Key takeaways
- Compliance has to be attached to the agent's identity, memory, tools and permissions, not kept in a policy document beside them.
- A more capable agent should earn wider authority only after the controls around it have been tested and logged.
- A regulated firm can start with one useful role, a short list of tools and one observable success criterion.
- Approval boundaries and a readable evidence trail matter more to an auditor than a general claim that the model behaves well.
- Permissions and logs need scheduled review, because controls decay as tools, staff and vendors change.
Why the controls must come before the capability
The planning question for a regulated business is not whether a more capable system will arrive. The question is what the company's own agents will be allowed to do on the day they get better at doing things. Capability tends to grow in steps that are hard to predict, so an organization that waits for a clear signal will make its decisions under pressure. Compliance work done early is easy to reuse. Compliance work retrofitted after an incident is expensive, slow and usually visible to people outside the company.
There is a second reason. Regulators, clients and auditors examine what a firm can demonstrate: who had access, what was done, who approved it and how the process was reviewed. A company that cannot produce those records for a simple assistant will not produce them for an agent that writes to customer accounts. The NIST AI Risk Management Framework treats AI risk as a continuing practice of governing, mapping, measuring and managing, not a certificate earned once. That framing fits the company AI layer well, because the layer changes every time a tool, a vendor or a staff role changes.
What it means for an agent to inherit the rules
An agent inherits a rule only if the rule lives in its operating context. A policy stored in a shared folder does not constrain anything by itself. The rule has to be expressed in five places: the agent's identity (a named role with a human owner), its memory (what it may store, for how long and about whom), its tools (which systems it can call and with what scope), its permissions (which actions run alone and which wait for approval), and its logs (what it records about each request and result).
This design matters most when the underlying model changes. If the rules are stored inside a model's behavior, every upgrade forces a new review from scratch. If the rules sit in the agent's identity, memory scope, tool grants and approval queue, the same boundaries carry over to the new model, and the team can check the difference in the evidence rather than in a demonstration. Tool access should follow established authorization practice: deny by default, grant the narrowest scope that completes the task, and check each request rather than trusting an earlier decision. The OWASP authorization guidance describes this least-privilege approach in detail, and it applies to agents as much as to people.
A step-by-step plan for the company AI layer
A regulated firm does not need a large program to begin. It needs a sequence that produces evidence at each stage. A workable order looks like this:
- Name one executive who owns the inventory of agents, tools, vendors and data flows. Without a single owner, the layer fragments across teams.
- Inventory every place an AI system touches company data, including tools that staff already use without approval. Treat that second group as a finding to manage, not a disciplinary matter.
- Write a one-page use policy with a short list of red lines, then translate each line into a concrete tool restriction, approval step or prohibited data category.
- Give each agent a named role, a memory scope and a short list of tools. Remove any tool the role does not need, even if it would be convenient.
- Define approval boundaries in three tiers: actions that run automatically with logging, actions that need one named approver, and actions that need two.
- Set a success criterion a person can observe, such as the share of escalated cases routed to the correct queue in a sampled review, rather than a claim about the agent's general intelligence.
- Keep logs that record the request, each tool call, the approver, the result and the retention rule that applies, and have counsel confirm that retention rule in writing.
- Schedule a permission review and a sampling review on a fixed calendar, and record what changed and why.
- Widen an agent's authority only after the evidence from the previous stage supports it. The permission tier approach describes how that widening can be staged.
Which decisions stay with people
Not every action deserves the same control, and a single blanket rule usually fails in one of two directions: it is either too strict to use or too loose to defend. A decision guide helps the operations team sort actions quickly:
- Read-only lookups of internal, non-sensitive information can run automatically, provided each lookup is logged.
- Drafts that a person will read before anything leaves the company can run with review at the moment of sending.
- Changes to customer records, payments, refunds, pricing or contract terms need a named human approver before execution.
- Any action that uses personal data in a category the company has classified as sensitive needs a documented legal basis and a second approver.
- Decisions that affect a person's employment, credit, health, insurance or access to a service remain with an accountable person, and the agent is limited to preparing the supporting record.
- Anything on the company's red-line list is never delegated, however capable the agent becomes. The red lines discussion explains how to draw that list.
The guide is only useful if the approver can say no. An approval step that a rushed person clicks through every time is a log entry, not a control.
An illustrative scenario: a document intake agent
Illustrative scenario, not a real case: a mid-sized insurance brokerage wants an agent to collect missing documents for open claims. The agent reads files placed in one claims folder, lists what is missing, and drafts a reminder for the account handler. It has no access to the payment system, its memory holds only the claim reference and the list of outstanding items, and every document it opens is logged with the time and the case number.
During the first phase, every reminder waits in a queue for an account handler to send. The team samples a set of queued drafts each week and records whether the agent identified the right missing items and used the right tone for the client. Only after several cycles with clean results does the firm consider letting routine reminders go out automatically, and even then the rule is written as a permission change with a named owner and a date.
Suppose the vendor later releases a more capable model. The scopes, the memory limits, the queue and the sampling schedule stay as they are. The team evaluates the new model against the same criterion and the same logs before any permission changes. The decision records approach describes what to keep so that a later reviewer can reconstruct why a given draft was approved.
How Be Human approaches this in managed operations
Be Human designs, implements and operates personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under the company's control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so the controls are checked again when tools, staff or vendors change. The work is delivered remotely, in English or Portuguese, and the company keeps visibility into what each agent can reach and why.
If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Does superintelligence exist yet, and why plan for it now? Nobody can give a reliable date for systems far more capable than today's, so a regulated firm should not build its plan around one. The controls that help with a highly capable agent also help with the agents in use today, including scoped access, approval gates and readable logs. Waiting for certainty leaves little time to build the evidence an auditor will ask for.
Can we rely on the AI vendor's safety commitments instead of our own controls? Vendor commitments are a useful input, but they do not replace the company's own controls. The firm remains responsible for its data, its records and the decisions its agents support. Review the vendor's contract terms, data handling and subprocessors, and map them to the controls your company runs internally.
Which regulation should we check first? The answer depends on where you operate, which data your agents touch and which decisions they support. Firms handling personal data in Brazil should review their obligations under the LGPD, and firms serving customers in the European Union should check whether the AI Act applies to specific uses. Confirm the scope with qualified counsel before you set the retention and approval rules.
How do we know a control is actually working? A control is working when a person can find evidence of it without help. That evidence includes logs showing which tool calls were made and who approved them, records of permission reviews, results from sampling audits, and notes from incident drills. If the evidence takes a skilled engineer several days to assemble, the control needs to be simplified before it is trusted.
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.


