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

Human oversight

Superintelligence Policy Watch: What US Companies Should Track Now

Which US policy signals matter for companies using capable AI, and how to turn each one into an internal review item with an owner, a trigger and a record.

Be Human typographic cover with the title: Superintelligence Policy Watch: What US Companies Should Track Now
Conceptual illustration created with AI · Be Human

As of this writing, no US statute defines superintelligence or sets rules for it specifically, so companies do not comply with a single superintelligence rule. What they should watch instead is a set of existing signals: federal guidance on AI risk, state AI and privacy laws, sector regulators, buyer requirements and the terms of vendors that handle company data. Each signal becomes useful only when it is written down as an internal review item with a named owner, a trigger and a record of what was decided.

Key takeaways

  • Treat superintelligence policy as a bundle of obligations you already face: privacy, consumer protection, employment, sector rules and contracts.
  • Track each signal by where it binds your company: the data your agents hold, the decisions they influence and the vendors they depend on.
  • Give every signal a review item with an owner, a trigger, a control, evidence and a next review point.
  • Tie the watch to what your agents actually do today, not to predicted capability milestones or dates.
  • Start with one agent role and its approval boundaries, then widen the watch list as that role changes.

Why there is no single rule to follow yet

Public debate about very capable AI systems is active, and it includes proposals for oversight of frontier development. Most of those proposals concern model developers, and they change as governments revise their positions. A business owner deciding what to do this quarter cannot wait for a final definition, and does not need one. The obligations that already apply to a company running AI agents come from privacy law, consumer protection, employment rules, sector regulation and the contracts the company signs.

Start from those obligations, then add capability-specific watch items as they become concrete. For a fuller discussion of why forecast dates should not drive your calendar, see Superintelligence Regulation Watch Without Relying on Predicted Dates.

Five signal families to monitor

The following families cover most of what reaches a US company that uses capable agents. The mix will differ by sector and by where your customers are.

  • Federal guidance on AI risk. Executive and agency documents often describe expectations for documentation, testing and human oversight. The NIST AI Risk Management Framework is a voluntary reference that buyers and auditors already ask about.
  • State AI and privacy laws. These often apply based on where your customers, employees or consumers are located, not where your company is headquartered.
  • Sector regulators. Existing rules on credit, hiring, health, insurance, advertising and financial communication already reach automated decisions, whatever the system is called.
  • Buyer and procurement requirements. Customer security questionnaires, government contract clauses and partner terms often set the practical standard before any statute does.
  • Cross-border rules. If you serve users in the European Union, the EU Artificial Intelligence Act can reach your agents directly, so check exposure against its overview.

The useful question is not which family sounds most serious. It is which of them touches an agent you run now or plan to run in the near term.

Turn each signal into an internal review item

A signal is only a review item when someone can act on it. Write every item with the same fields so that items can be compared, filed and checked later.

  • Signal: the rule, guidance or requirement, described in one sentence and linked to its source.
  • Binding point: the agent, data set, memory store, tool or vendor the signal touches.
  • Owner: one named person who decides, with a deputy for absences.
  • Trigger: the event that reopens the item, such as a new vendor, a new data source, a new customer contract or a change in the agent's tools.
  • Control: the boundary you set, such as an approval step, a retention limit or a restricted tool.
  • Evidence: the record you would show a customer, auditor or regulator, such as logs, a signed terms page or a test result.
  • Next review: the date or event for the next look, set by the owner.

Keep the register short. A one-page register with a few carefully maintained items is more useful than a long document nobody opens. The AI Risk Register Template shows one simple way to lay this out.

Illustrative scenario: a billing agent and a new state privacy rule

Illustrative scenario, not a real case: a services company runs an AI agent that answers billing questions for customers and can apply small service credits within limits set by staff. The agent keeps a memory of past conversations so it can recognize returning customers. A state privacy amendment then strengthens customers' rights to request deletion of personal information a business holds about them.

The operations lead opens a review item. The binding point is the agent's memory store and its logs. The owner is the operations lead, and the privacy officer reviews the control. The trigger is confirmation from counsel of when the amendment applies. The control is a memory structure that tags entries by customer, a deletion process that reaches both memory and logs, and a rule that credits above a set limit go to a staff member. The evidence is a test showing that a deletion request removed the customer's memory entries and that logs followed the retention policy. Nothing in this scenario requires a new model or a new vendor. It requires the review item to exist before the amendment takes effect.

Data, memory and vendor terms need the earliest attention

Most of the signals above converge on data. Agents create copies of company information in prompts, memory, logs and tool outputs, and each copy has a location and a retention period. Ask vendors where those copies go, who else processes them and how they are removed. Then compare the written answers with what the agent actually does, because the gap between the two is where most audit findings start.

  • Where does each agent store memory, logs and tool outputs?
  • How long is each kept, and who can change that setting?
  • Which subprocessors touch that data?
  • Can a deletion request reach every copy?
  • Is customer content used to train models, and is that refusal recorded in writing?

Authorization deserves the same discipline. The OWASP authorization guidance is a practical reference for deciding which identity an agent acts under and which actions it may take.

How Be Human handles this in a managed setup

Be Human designs and operates AI agents with a defined role, memory, tools, permissions and approval boundaries, all kept under human control. In a managed engagement, policy signals attach to those same boundaries. When a signal touches a role, its review item is linked to that role's configuration, recurring evaluations check the agent's behavior against the control, permission review runs on a schedule, and changes to memory or tools are recorded. The first engagement defines one useful role, the memory and tools it needs and the evidence that will show whether it works, so the watch list starts from a concrete agent rather than from general commentary. Be Human does not provide legal advice, and your counsel decides how each signal applies to your business.

A monthly routine that does not depend on predictions

A simple cadence keeps the watch current without turning it into a research project.

  • Each month, the owner of each review item checks its source for changes and records whether the item is affected.
  • Each quarter, review agent permissions, memory retention and vendor terms against the register.
  • When a vendor changes terms, a model is replaced or a new data source is added, reopen the affected items that week.
  • When a customer or auditor asks for evidence, pull it from the item instead of rebuilding it from memory.
  • Once a year, ask counsel to confirm which items still reflect current law.

If you are not sure which agent task should be the first one covered by this routine, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Is there a US law that regulates superintelligence today? As of this writing, no US statute defines superintelligence or sets rules specific to it. Companies are affected by existing obligations in privacy, consumer protection, employment, sector regulation and contracts, and those already apply to AI systems. The practical task is to track those obligations and the policy discussion around advanced systems, without planning around a single predicted rule.

Which signal should a small company watch first? Start with the data your agents can reach and with the buyers who require security terms. Privacy obligations and customer contract terms usually create the earliest concrete duties and are the easiest to check against an existing agent. Once those items have owners and evidence, add the sector and regional rules that apply to your customers.

Do I need a lawyer to run this watch? Not for every item, but counsel should confirm how a signal applies to your business, especially for regulated decisions, sensitive data or cross-border transfers. A business owner can run the monthly routine and keep the register, while counsel interprets the obligations. Clear written questions to counsel make each review faster.

Does the watch still matter if we only use agents for internal work? Yes, because internal agents still process employee data, company records and vendor content, and those can trigger privacy, retention and access duties. Internal use also tends to expand into customer-facing tasks, so each review item should name that expansion as a trigger. A short register is enough for internal work as long as every item has an owner and evidence.

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 →