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

Personalized AI infrastructure

Superintelligence Scenario Planning for Leadership Teams: Six Questions to Test

A six-question workshop for leadership teams to test whether current AI decisions still hold if machine capability rises sharply, with outputs to review.

Be Human typographic cover with the title: Superintelligence Scenario Planning for Leadership Teams: Six Questions to Test
Conceptual illustration created with AI · Be Human

Superintelligence scenario planning is a structured way for a leadership team to test whether its current AI decisions still hold if machine capability rises faster than expected. The method is a workshop built on six questions, and every answer must become something written: a decision, a trigger, an owner and a review date. It does not require agreement on when advanced systems will arrive, only on which choices would be costly to reverse.

Key takeaways

  • Test decisions, not predictions. Start from what you have already committed to and check whether it survives several capability outcomes.
  • Six questions cover the assumptions most likely to break: dependencies, vendors, access, decision rights, people and evidence.
  • Every answer should produce a written artifact with an owner and a trigger, or it will not change anything.
  • Reversible choices can move quickly. Data, permissions and vendor dependence need explicit checkpoints before they expand.
  • Repeat the workshop when a vendor, agent or connected process changes, and on a fixed quarterly cycle.

Why test decisions instead of forecasting

Most leadership discussions about advanced AI stall on timelines. Whether a major capability jump arrives soon or much later is a question nobody in the room can answer with confidence, and a plan built on one forecast breaks as soon as the forecast does. A more useful question is which decisions you have already made that quietly assume AI stays roughly as capable, as affordable and as available as it is today. Those are the decisions worth testing.

The NIST AI Risk Management Framework describes AI risk work as a continuing cycle of governing, mapping, measuring and managing, not a single document written once. Scenario planning fits that cycle. It gives the team a way to map where its assumptions sit, measure which ones matter most, and set rules for managing them when conditions change.

Good scenarios are chosen for their consequences, not their odds. A useful set might include one case where agents handle routine operations reliably, one where a vendor changes access or terms quickly, and one where capability grows but internal records are too inconsistent for any agent to use well. The current plan should survive each case, or the team should know exactly where it does not.

Prepare the register and the room

Gather the inputs before the session. Most of the useful findings come from comparing real records with what people believe is true, so preparation matters more than debate. Bring:

  • A register of current AI-related decisions, including tools in use, pilots underway and vendor agreements.
  • A simple map of where core data, documents and know-how live, and which systems can read them.
  • A list of every AI agent or automation with its permissions, owner and approval boundary.
  • The names of people who can approve changes to scope, budget or access.
  • Review notes from any AI pilot that was paused or stopped, including the reason.

Keep the group small, usually four to six people: an executive sponsor, an operations leader, someone from finance, a person responsible for data or security, and the person who runs the work day to day. The sponsor matters because some outputs will require decisions that a working group cannot make alone. The role is described in Executive Sponsor for an AI Project: What the Role Requires. Plan for a working session of two to three hours, and spend the most time on the access and decision-rights questions, where the findings are least obvious.

Questions 1 and 2: Which decisions and vendors assume today's AI?

Question one: Which current decisions assume AI stays as it is today? Ask each person to name three decisions from the register that would look different if the tools became much stronger or much weaker. Typical examples include a staffing plan that assumes documents will always be reviewed by hand, a process design that keeps a person in every step, or a training budget tied to one product. For each decision, write the decision in one line, the assumption it depends on, and whether that assumption can be reversed within a quarter.

Question two: What would you lose if a provider changed its terms, access or capability? Vendor dependence is the most common hidden assumption. If a model provider changes its interface, pricing or data terms, which workflows stop, and how long would they take to rebuild elsewhere? Check whether prompts, workflow logic, evaluation sets and stored memory can be exported in a usable form. The exit questions are covered in How to Avoid AI Vendor Lock-In With Business Agents.

Output for this section: a dependency list naming the three tools or vendors whose change would hurt most, with the estimated replacement effort for each.

Questions 3 and 4: Who can reach what, and who decides?

Question three: Who can reach our data and know-how, and through which agent? As agents become more capable, what they can reach matters more than what they can say. Each agent should have its own identity and only the permissions its role needs. The OWASP authorization guidance is clear that access should be checked on every request rather than assumed from an earlier session, and the same principle applies to agents acting for people. Review each agent against four lists: data it can read, actions it can take alone, actions that need approval, and actions it must never take. The steps are set out in How to Run a Permission Review for Your AI Agents, Step by Step.

Question four: Who decides, and what evidence triggers a change? Scenario planning without decision rights produces a document nobody uses. For each scenario, name who can change scope, pause an agent or alter its access, and what observation would start that step. Write the trigger before the event, so the team is not inventing its rules during a difficult week.

Illustrative scenario (not a client case): A regional distributor runs an agent that drafts replies to supplier emails. Its trigger rule says the operations lead can pause the agent at once if a draft quotes terms that do not match the contract record, while any change to the agent's access to supplier files needs approval from the finance and data owners. Because the rule was written before any incident, nobody had to decide it under pressure.

Output for this section: a trigger list with one owner and one approval path for each agent.

Questions 5 and 6: What changes for people, and what evidence shows the scenario?

Question five: How would the work of our people change, and what would they still need to verify? Capability gains change which tasks people check, not whether they remain responsible for the result. Ask which roles would hand routine steps to an agent, which would review its output, and what each reviewer needs to know to catch an error. Adoption is often the constraint before capability is, and the practical steps are covered in How to Get Your Team to Actually Use an AI Agent.

Question six: What evidence would show which scenario you are in? Choose signals before you need them. Useful examples include error rates on a fixed set of real tasks, the time and cost of completing a standard workflow, new capabilities a vendor announces that touch your processes, and any action an agent took outside its boundary. Collect them the same way each time. OpenTelemetry's observability primer explains how traces, metrics and logs let a team reconstruct what a system did, and the same discipline applies to agents. For the test set itself, the OpenAI evals guide describes how to build repeatable evaluations, which is the business equivalent of a regression suite.

Output for this section: a signal list with a definition, a data source, an owner and a threshold that starts a review.

From answers to reviewable outputs

At the end of the session, the team should hold five written outputs: the dependency list, a permission summary for each agent, the decision rights and triggers, the signal list with owners, and a review date. A simple sequence keeps the work honest:

  • Write each answer as one sentence and name the person who owns it.
  • Rank the items by how costly they would be to reverse.
  • Set a trigger and a check for each of the top five items.
  • Put the review date within the next quarter.
  • Store the outputs where the steering group can read them, and where changes can be requested and recorded.

In Be Human's managed model, the same outputs become part of the operating record for each agent. Approval boundaries, permission reviews, evaluation criteria and incident handling are documented, assigned to a named person and checked on a fixed schedule. Monitoring and recurring evaluations supply the signals, and changes go through a controlled request process rather than informal edits.

Running the workshop every quarter

Scenario planning is not a one-time exercise. Repeat it when a major vendor changes its terms, when a new agent is added, when a process it touches is redesigned, and at least once a quarter. Compare each new set of outputs with the previous one, and note what changed and why. A plan revised with evidence is more useful than one that was correct when it was first written. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Do we need to predict when superintelligent AI will arrive before we start? No. The workshop tests current decisions against a range of outcomes, so the team can agree on what to watch without agreeing on dates. Starting with the choices that are hardest to reverse is usually enough to produce useful first outputs.

How long should the workshop take? A working session of two to three hours is usually enough for a first pass, followed by a shorter review once owners have filled in the detail. Most of the time should go to the access and decision-rights questions, because that is where the least obvious gaps tend to sit.

Who should own the outputs after the workshop? Name one executive owner for the plan and one operational owner for each agent or system it covers. The executive owner decides changes to scope, and the operational owner keeps permissions, evaluations and incident notes current.

Is this the same as an AI risk assessment? It overlaps with one but is narrower. A risk assessment inventories risks across the organization, while this workshop starts from current decisions, asks which ones would break under a sharp change in capability, and assigns triggers and owners to those decisions.

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 →