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

Human oversight

Control Points for Superintelligent Agents: Five Places to Hold Authority

Five control points for AI agents: scope, tools, approvals, memory and shutdown, with a practical test for each one a business owner can run.

Be Human typographic cover with the title: Control Points for Superintelligent Agents: Five Places to Hold Authority
Conceptual illustration created with AI · Be Human

A control point is a specific place where a company keeps authority over an AI agent. For most business agents there are five: the scope of work the agent may start, the tools it can call, the actions that need a named person to approve them, the memory it may keep, and the way it can be paused or stopped. Each one can be written down, tested with a plain question, and checked again after the agent changes.

Key takeaways

  • Authority over an agent lives in five places: scope, tools, approvals, memory and shutdown.
  • Each control point needs a named owner, a written limit and a test that produces a recorded result.
  • Tools and approvals fail most often when they are granted once and never reviewed.
  • Shutdown must halt actions while keeping logs, memory state and queued work intact.
  • Test all five before launch, then repeat the tests on a schedule and after every change.

Why authority needs five separate places

Many companies think of control as one switch: the agent is either on or off. That switch is necessary, but it is too coarse. An agent can cause trouble in several ways that a single switch does not address. It can start work nobody asked for, use a tool it does not need, send a message that should have waited for approval, keep personal details longer than the purpose requires, or keep acting while the team is still trying to understand what went wrong.

Separating authority into five places lets an operations leader answer specific questions. Who decided this job belongs to the agent? Which systems can it reach? Which actions wait for a person? What does it remember? How quickly can work stop, and what is left behind? Each answer belongs to a named owner and can be checked. The NIST AI Risk Management Framework treats this kind of work as an ongoing cycle of governing, mapping, measuring and managing risk, which is a useful reminder that these controls are not a one-time setup.

Scope: decide what the agent is for before anything else

Scope is the written description of the job. It names the outcome the agent is responsible for, the inputs it may read, the outputs it may produce, and the situations it must refuse or hand back to a person. A vague scope such as help the sales team gives the agent no edges and gives the owner no basis for saying no.

A usable scope fits on one page and answers four questions: what task starts with the agent, what task ends with it, what it must never do, and what success looks like in a form someone can observe. The last item matters most. If the team cannot describe a measurable result, the agent will drift toward whatever it can do easily rather than what the business actually needs.

Test for scope: give the written scope to someone outside the project and ask them to list three requests the agent should refuse. If they cannot name three, the scope is not yet a control point.

Tools: limit what the agent can reach

Tools are the connections that let an agent act: a calendar, a CRM, a shared drive, a payment interface, an email account. Every tool is a power, and powers should be granted one at a time with a reason attached. The OWASP authorization guidance makes a point that applies directly here: access decisions should check the specific action and resource, not assume that a connected system is trusted for everything it can do.

Practical limits include read-only access where writing is not needed, separate credentials for the agent rather than a staff member's login, narrow permissions inside each tool, and a list of tools the agent can see but not call. Entitlement mapping for agent tools is the simplest way to keep this honest, because it lists each tool, the records it touches and the actions it can take.

Test for tools: pull the agent's actual permissions from each connected system and compare them with the entitlement list. Every difference is either a missing permission that blocks the job or an extra permission that should be removed. Record both.

An agent should hold the access its job needs today, not the access that might be useful someday.

Illustrative scenario: a small distributor connects an agent to its order system so it can answer delivery questions. The first version is given write access to change order status because that seemed convenient. The review finds the agent could also cancel orders. The owner removes cancellation rights, keeps status reads, and adds a note to the entitlement list saying cancellations need a staff member. The change takes an afternoon, and the record shows why it was made.

Approvals: name the actions that wait for a person

An approval gate is a point where the agent prepares an action and a named person decides whether it happens. Gates belong on actions that spend money, change customer commitments, send messages to people outside the company, delete records or touch sensitive personal data. They do not belong on every routine step, because a gate that fires constantly will eventually be approved without anyone reading it. Teams planning these gates can look at where people sign off on agent actions for examples of where the decision should sit.

Write the gate list as a short set of conditions in plain language. For each action, state who approves, how long the approval stays valid, what the approver sees, and what happens if no one responds. Silence should default to not acting. For the highest-risk actions, consider requiring two people, and reserve that for the few cases where a single mistake would be costly.

Test for approvals: trigger each gated action in a test environment and confirm that the agent stops, the right person receives the request, and nothing is sent until the approval is recorded. Then test the timeout by leaving a request unanswered. If the action proceeds anyway, the gate is not working.

Memory: control what the agent keeps and for how long

Memory is what the agent carries from one conversation or task to the next. It can be a notes store, a search index built from documents, a summary of past exchanges, or saved preferences. Memory makes an agent useful, and it also turns the agent into a record of personal and commercial information. The OWASP prompt injection guidance treats retrieved and stored content as input that can carry instructions or errors, so memory is a security surface as well as a convenience.

Decide three things for each memory store: what it may contain, who can read it, and when it expires. Keep personal data out of memory unless the purpose requires it and the legal basis is clear. Keep memory for one client or department separate from memory for another. Confirm that removal is possible, since a deletion request is hard to honor if nobody knows where the agent has written things down. The steps for deleting agent memory on request are a good model for building this into the design from the start.

Test for memory: choose a known person or record, ask the agent what it knows about them through normal use, then search the memory stores for that name and identifier. Everything found should match the retention policy. If something is kept that should not be, remove it and confirm the removal in the store itself, not only in the agent's answer.

Shutdown: stop the agent without losing the record

Shutdown is the control point most often tested too late. A company needs to pause an agent quickly, stop specific actions while leaving others running, and preserve what happened so the team can explain it afterward. A pause that deletes logs or drops a queue of pending work creates a second problem on top of the first.

Design shutdown in layers. The first layer halts new actions across all tools. The second revokes or disables specific credentials. The third preserves logs, memory state and queued work for review. Name the person who can trigger each layer and write down the order in which to act. If records may be needed later for a dispute or audit, include a check for legal holds before anything is cleared. A fuller treatment of kill switch design for business agents can help you decide which layers your business needs.

Test for shutdown: in a test environment, trigger each layer during a live workflow. Confirm that new actions stop, that in-flight work is either completed or clearly marked as held, and that logs remain readable afterward. Time the whole exercise and write down the result. The first run usually takes longer than expected, which is exactly why it should be practiced before it is needed.

Testing all five before launch and after every change

Testing turns these five control points from policy into evidence. A practical sequence for a first deployment looks like this:

  • Write the one-page scope and have someone outside the project try to find gaps in it.
  • Pull the agent's live permissions and reconcile them with the entitlement list.
  • Map every gated action and run each one through approval and timeout tests.
  • Search memory for a known record, confirm the retention rule and test removal.
  • Trigger each shutdown layer in a test run and confirm that logs and queued work survive.
  • Record the results, the owner of each control point and the date of the next review.

Repeat the same sequence after any change to tools, prompts, data sources or permissions, and on a schedule the business chooses. Monitoring can show what an agent did, which is the kind of telemetry the OpenTelemetry documentation describes for observability, but someone still has to read the record and decide whether authority is still in the right place.

Be Human handles this as part of a first engagement and ongoing managed operations. The first engagement defines one role, the memory and tools it needs, its approval boundaries and an observable success criterion, so each control point has a named owner before anything goes live. Managed operations then keep those points current through monitoring, recurring evaluations, permission review and incident response, and any controlled improvement goes through the same tests. The business keeps decision rights over every control point throughout.

If you are unsure which task is worth handing to an agent first, take the free Be Human AI diagnostic to find it.

Frequently asked questions

What is an AI agent control point? It is a defined place where a company holds authority over what an agent can do, reach, keep and stop. The five common control points are scope, tools, approvals, memory and shutdown. Each one should have a named owner, a written limit and a test with a recorded result.

How many control points does a small company need? A small company does not need more machinery than it can review, but it should still address all five areas. For a single agent with a narrow job, each control point can be a short page or a few lines in one document. The important part is that every area has an owner and a test, even when the answers are simple.

Who should own each control point? The business owner should own the scope and approval rules, since those reflect commercial decisions. Whoever manages the technology should own tool permissions, memory stores and shutdown procedures, with the owner of each area signing off on changes. Ownership should be written down so that no one assumes another person is watching a control.

Can a control point change after launch? Yes, and it should change when the business changes. Each adjustment to tools, approvals, memory or shutdown should be tested again and recorded, so the team can show what authority the agent held at any given time. Changes made without a test are the most common source of surprises later.

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 →