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

Personalized AI infrastructure

Superintelligence Timelines: Plan by Capability, Not a Calendar

Predicted dates are weak planning inputs. Tie your AI preparation to observable capability thresholds, named owners and review points instead.

Be Human typographic cover with the title: Superintelligence Timelines: Plan by Capability, Not a Calendar
Conceptual illustration created with AI · Be Human

A superintelligence timeline for companies should not be a date you plan around. Forecasts about when AI systems will exceed human ability across most tasks disagree widely, and none is reliable enough to set a budget or a hiring plan. A better approach is to name the capabilities that would change your operations, define how you would observe them, and decide in advance what you will do when the evidence arrives.

Key takeaways

  • Forecast dates are poor inputs for company decisions, so tie preparation to capabilities you can observe.
  • Write each threshold as a test with an owner, an evidence source and a response agreed in advance.
  • Separate the processes that would change from the controls that must stay under human authority, whatever a model can do.
  • Review thresholds on a fixed schedule, not in reaction to headlines about new models.
  • Start with agent tasks you can run and measure today, because those produce the evidence you will need later.

Why a calendar is the wrong planning tool

A date sounds useful because it gives a meeting something to point at. The trouble is that a date rests on assumptions about research progress, computing supply, regulation and adoption, and those assumptions change. People who quote a year often mean different things by superintelligence, so two credible sources can give dates decades apart. A company that builds its plan around one of them either waits too long or spends early on the wrong things.

Capabilities are different because they show up in work. The question for an operations leader is not when a system becomes superhuman in the abstract, but whether it can reliably complete a multi-step task in your environment, with your data, under your rules, without a person correcting it. That question has an answer you can check this month, and it has a different answer next quarter if the system improves. Planning around that answer keeps your preparation tied to what actually changes.

How do you define a capability threshold in business terms?

A threshold is a statement that someone can test. It names the task, the conditions, the evidence and the response. Vague wording such as the agent becomes very capable cannot be tested, and it cannot trigger a decision. A usable threshold reads more like a specification.

Illustrative scenario: a mid-sized equipment supplier wants to know whether an agent can handle first-line supplier invoice disputes. The threshold is written as follows: the agent reads the purchase order, the delivery record and the supplier message, drafts a resolution that follows the payment policy, and routes any amount above the approval limit to a finance approver. The threshold is met when a reviewer, sampling resolved cases each month, finds no resolution that breaks policy and no case where the agent failed to route an exception. The response is fixed in advance: if the threshold is met for three consecutive reviews, the agent's scope widens to a second supplier group; if it is missed, scope narrows and the failing cases go into the evaluation set.

Notice what the example does not do. It does not predict when the agent will handle every dispute, and it does not depend on a new model release. It sets a condition, a check and an action, which is what a leader can act on.

A step-by-step plan for capability-based preparation

  • List the operational decisions that a more capable system would change, such as staffing levels, approval speed, customer response times and access to internal records.
  • For each decision, write one threshold as a testable statement that names the task, the conditions and the required evidence.
  • Assign a single owner who can accept or reject the evidence. Shared ownership usually means nobody acts.
  • Define the response in advance: widen scope, add review, hold steady, narrow permissions or pause the workflow.
  • Set a fixed review date for each threshold, plus a trigger for early review after a material model change or an incident.
  • Record each review in one decision log that shows the evidence, the decision and the person who made it.

This plan is deliberately modest. It asks for a short list, a few tests and a log. Companies that already run regular operations reviews can attach it to that meeting rather than creating a new body.

Which controls should not depend on the model?

Some controls must hold regardless of how capable the system is. The first is the set of permissions. An agent should have access only to the systems and actions its role requires, and that access should be reviewed on a schedule. The OWASP Top 10 for LLM Applications lists excessive agency and prompt injection among the main risks in these systems, and both are reduced by narrow tools, explicit approval steps and logs that show what the agent did. A stronger model does not remove the need for these limits. It raises the cost of having none.

The second control is the approval boundary. Decide which actions the agent may take alone, which need a named person's approval, and which it may never take. Put these rules in writing, and keep them separate from the prompts that shape the agent's behavior, so a change to one does not silently change the other. The third control is the audit trail. Without a record of inputs, actions and outcomes, you cannot tell whether a threshold was met or whether a failure was an isolated event. These controls are the foundation of the decision log described above.

How do you test capability claims before you believe them?

A vendor's benchmark or a demonstration on a clean example is not evidence about your workflow. Evidence comes from a fixed set of tasks drawn from real work, run repeatedly, with outputs graded against criteria the business has agreed. The OpenAI evals guide describes this pattern as a way to measure system behavior against defined tests, and the same logic applies to any agent you operate. Include hard cases, not just typical ones, and keep failed cases in the set so that a later change cannot quietly reintroduce them.

Evaluations also protect you from the opposite error, which is dismissing an agent that has improved. When the same test set is run after each change, the trend becomes visible. For a deeper treatment of which measures predict business outcomes, see Superintelligence Benchmarks: Which Ones Predict Business Results. The practical rule is simple: no threshold is marked as met on the strength of a claim alone, only on the strength of a reproducible result in your environment.

How Be Human's managed approach handles thresholds

Be Human designs and operates personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools that role needs, its approval boundaries and an observable success criterion. The capability thresholds for that role are written at that stage, attached to the evaluation set and the review calendar. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so when a model changes, the change is tested against the workflow before it is adopted. For governance, the question of who accepts evidence and who decides on scope is handled in a charter that names the decision rights and review dates, as described in AI Governance Charter for Companies: Decision Rights and Review Dates. The NIST AI Risk Management Framework offers a useful structure for these functions, organized around governing, mapping, measuring and managing AI risk.

Where to start

The first step does not require a forecast or a new budget line. Pick one workflow where an agent could plausibly help, write its current performance as a test, and see what the test shows. Teams that want to compare options before committing often find it useful to read Draw Your AI Stack: A One-Page Architecture Map for Leaders first, so the workflow sits inside a clear picture of data, tools and permissions. When you are ready to choose that first task, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Is there a reliable superintelligence timeline for companies? No. Public predictions about advanced AI vary widely and depend on assumptions about research progress, computing, regulation and adoption. Companies get more value from preparing for specific capabilities that would affect their operations and reviewing that preparation on a fixed schedule.

What is the first capability threshold a company should track? Start with the capability that would change a process you already run, such as an agent completing a multi-step task across your own systems without correction. Write it as a test with a sample of cases, a named owner and a review date. The first threshold only needs to be measurable, not ambitious.

Should we slow AI spending until superintelligence arrives? Usually not. Spending on controlled, measurable agent work builds the permissions, logs and evaluation habits you will need under any timeline. The real risks are unmanaged scope and decisions without an owner, and modest, reviewed pilots reduce both.

How often should capability thresholds be reviewed? Review them on a fixed cadence, and also after any material model change or incident. Each review should check whether a threshold has been met, whether permissions still match the role, and whether the planned response still fits the risks you face.

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 →