Personalized AI infrastructure
Superintelligence Scenarios: Three Capability Levels Without Date Guesses
A leader's guide to planning for three AI capability levels without predicting dates, using moves that stay useful whichever level arrives first.

Scenario planning for superintelligence means preparing a company for three capability levels instead of one predicted arrival date. Each level changes what an AI agent can do inside your operations, so the useful question is which decisions, records and permissions would need to change first. The strongest moves are the ones that still pay off if a level arrives later than expected, or sooner.
Key takeaways
- Plan for three capability levels, not for a calendar date that no one can verify.
- Choose moves that pay off at every level: clean records, written decision rights, a memory layer you own and a regular review rhythm.
- Treat each new capability as a trigger for review, not as an automatic expansion of what an agent may do.
- Keep human approval boundaries explicit, and raise them only on evidence from observed performance.
- Give every agent one observable success criterion so progress can be measured in your own terms.
Why plan for capability levels instead of dates?
A date invites overcommitment. Leaders hear a forecast, approve a large budget or delay a decision on the strength of it, and then discover that the forecast was a guess. A capability level is different because it describes what a system can do in work like yours. That makes it observable, and it lets you ask concrete questions about your own operations at each stage.
For planning purposes, it helps to use three broad levels. The first level is supervised execution: agents reliably complete bounded tasks such as drafting, sorting, summarizing or checking, and a person reviews the output before it leaves the team. The second level is sustained work: agents carry multi-step tasks across several tools and over longer periods, with less supervision and more dependence on memory and permissions. The third level is broad capability: systems perform most knowledge work at a level that exceeds your team's current methods and can propose or revise their own procedures.
These labels are planning categories for your company, not a standard the industry has agreed on. Their value is that each one implies different risks, different records and different approvals. You do not need to agree on when each level will arrive. You need to agree on what would change for you if it did.
What stays useful at every level
Some preparation is worth doing whichever level arrives first, because it improves the business today and makes any later step safer. These are the moves that hold their value across all three scenarios:
- Clean operational records, with one source of truth for customers, products, prices, inventory and cases.
- Written decision rights that say who approves what, and at what size of commitment.
- A memory layer with a named owner, so the company controls what the agent remembers and can correct or delete it.
- Tool access scoped by role, with read and write permissions kept separate.
- An evaluation set built from real past cases, tied to a business outcome rather than to a general impression of quality.
- A documented exit path for each vendor and each integration, so no single provider holds the company's knowledge hostage.
None of these depends on predicting the future. Each one is a tidier way to run the business. The test is simple: if the most optimistic scenario never arrived, would this move still help? If the answer is yes, it belongs in the first wave of work.
How to map your exposure, step by step
Most companies make scenario planning too abstract. The following sequence keeps it tied to the work people actually do. It can be completed in a few working sessions with the operations lead, the owner of the most important process and someone who understands the systems involved.
- Step one: list the five recurring decisions or tasks in your operations that consume the most judgment, such as quote approval, refund decisions, supplier selection or scheduling exceptions.
- Step two: for each task, write down what an agent does today, what it could do at the first level, and what it would be allowed to do at the second and third levels.
- Step three: note the permissions each change would require, including read access to new records, write access to systems of record and authority to contact customers or suppliers.
- Step four: identify the record that would have to be trustworthy for that change to be safe. If the record is inconsistent, the fix comes before the agent.
- Step five: name one owner for the decision, one owner for the memory and one owner for the permission. Shared ownership usually means no ownership.
- Step six: define one success criterion for the current stage, written so that a reviewer can check it without asking the agent how it feels.
The output is a short map rather than a forecast. It shows where the company is exposed, which moves are urgent and which can wait for evidence.
Which controls should change first?
When a capability level rises, the temptation is to widen what the agent can do. The more disciplined response is to revisit controls in a fixed order. Start with the permissions, because they determine what an agent can actually touch. The OWASP authorization guidance describes the principle of granting only the access a function needs, and that principle is the right default for agents as well. Then review approval thresholds, because they determine when a person must step in. Raise thresholds only after the current threshold has produced enough reviewed cases to show the agent handles them correctly.
The NIST AI Risk Management Framework offers a useful structure for this work. It organizes AI risk around governing the system, mapping its context, measuring its behavior and managing what the measurements reveal. Translated into a leader's language, that means someone owns the decision, someone understands where the agent operates, someone checks its results against defined criteria, and someone acts when the results change. If one of those four roles is empty, the control is incomplete.
For agents that read external content, such as customer emails or supplier documents, add a further check. Instructions hidden in that content can try to redirect an agent's behavior. The OWASP guidance on prompt injection is a practical reference for the defenses that matter, including separating untrusted content from instructions and limiting what a compromised response can trigger.
Illustrative scenario: a distributor's quoting agent
The following scenario is illustrative and does not describe a real company. A regional distributor uses an agent to draft quotes from its price list and stock records. At the first level, a salesperson reviews every draft before it goes to a customer. Over time the operations lead notices that the agent could also confirm delivery dates, which would be a second-level task that touches the warehouse schedule.
Because the company had already separated pricing from availability in its records, the change did not require a rebuild. The team ran a permission review, narrowed the agent's write access to the quote itself, and kept delivery confirmations behind a person until a month of reviewed quotes showed no date errors. The plan had been written before anyone needed it, so the decision was a controlled step rather than an argument about the future.
How a managed approach handles the planning cycle
Be Human designs and operates personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under the company's control. In managed operations, that scope is documented and reviewed on a schedule. Recurring evaluations check the agent against its success criterion, permission reviews test whether access still matches the role, and changes to scope or thresholds pass through a defined request and approval step. Capability changes are treated as reasons to review the map, not as permission to expand the agent by default.
What to decide before the next planning cycle
Before the next planning cycle, agree on three things: the three capability levels as your company will use them, the five decisions you will map first, and the person who approves any change to an agent's permissions. These decisions cost little and clarify almost everything that follows. If you want to find the first task worth handing to an agent, the free Be Human AI diagnostic below is a practical place to begin.
Frequently asked questions
Do we need to decide which capability level is most likely? No. The purpose of the three levels is to describe what would change in your operations, not to rank which level is probable. Plan the moves that pay off at every level first, then review the map when observed performance gives you a reason to.
What if a higher level arrives much sooner than we expect? The preparation still helps because the most valuable steps, such as clean records, scoped permissions and a documented owner for each decision, shorten the time needed to respond safely. Your review trigger should be an observed change in what the agent can do, tested against your criteria, rather than a headline.
Is this planning only relevant to large companies? No. Smaller companies often benefit more, because a small team can define decision rights and records quickly and avoid rebuilding them later. The map can be a single page reviewed by the owner and one operator.
How often should the plan be revisited? Review it on a fixed rhythm, such as each quarter, and also whenever an agent's scope, tools or approval thresholds change. A plan that is revisited only when something goes wrong tends to be out of date when it matters most.
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.


