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

Personalized AI infrastructure

Superintelligence in the First 90 Days: A CEO's Week-by-Week Plan

A week-by-week plan for CEOs: name an owner, scope one AI role, set memory and permission rules, run a success check, then decide what to buy next.

Be Human typographic cover with the title: Superintelligence in the First 90 Days: A CEO's Week-by-Week Plan
Conceptual illustration created with AI · Be Human

A CEO's first 90 days with superintelligent AI should end with clear ownership, one working role and a way to measure that role, not a stack of new licenses. The sequence is short: name one accountable owner, scope one role, write the memory and permission rules that role needs, run a success check that operations can observe, and only then decide what to buy next. The steps below are organized by week so that each one has a finish line.

Key takeaways

  • Name one executive owner for AI decisions before any tool is evaluated.
  • Scope one role with a single job, a named human approver and a clear stop condition.
  • Write memory and permission rules before the first test, not after the agent has accumulated context.
  • Define a success check from data operations already records, plus the failures that pause the role.
  • Buy nothing new until the first role has held its check across several review cycles.

Week 1: Name the owner before choosing tools

The first week is a decision, not a vendor call. Someone must own the outcomes of AI in the company, and that person needs authority over budget, data access and approval of changes. In most companies this is the CEO as sponsor, or an operations leader with a clear mandate, plus one day-to-day operator who runs the role. Put the names on one page with three responsibilities: approving the role, approving access, and deciding when the role is paused. The NIST AI Risk Management Framework gives owners a shared vocabulary for these risks, which helps when the operator and the executive need to agree quickly.

  • Name one executive owner and one day-to-day operator.
  • Write what each may approve alone and what needs the CEO.
  • Decide who can stop the agent, how, and within what time.
  • Book one monthly review where results and failures are read in full.

A short written charter keeps these decisions from drifting. A committee charter for small companies shows one workable structure for this.

Weeks 2 and 3: Scope one role with a narrow job

Pick one recurring task where the output is easy to check and the cost of an error is small. Good first candidates often sit between teams: answering repeated supplier questions from approved records, drafting a weekly summary of open service requests, or sorting incoming requests into queues. Avoid roles that approve payments, change prices or make commitments to customers during the first quarter. The purpose of this stage is to learn how the company works with the role, not to test how much autonomy the role can take.

Before building anything, write the role on one page and check it against this list:

  • The role has one job, stated in a single sentence.
  • Every input and output names the system it comes from or goes to.
  • Actions are split into read, draft and act, and only read and draft are allowed at the start.
  • Each act-level step has a named human approver and an agreed response time.
  • The role has a hand-back point: the situation in which it stops and passes the work to a person.

Illustrative scenario: a regional equipment distributor wants help with supplier status questions. The owner scopes the role to reading order status and drafting replies that a staff member reviews and sends. The agent cannot change an order or promise a delivery date. That single boundary is enough to start, and it keeps the first review simple.

Weeks 4 to 6: Write the memory and permission rules

Memory is where early mistakes usually happen. Decide what the role may remember, what it must never store, how long each category is kept, and who can review or delete it. Write these rules before the first test, because cleaning up months of informal notes is slow and uncomfortable. Permissions follow the same logic: each tool connection should carry only the access the job needs, and write access should be granted in writing by a named person. The OWASP authorization guidance is a useful reference for deciding what each identity may do.

Text from outside the company deserves its own rule. Emails, web pages and uploaded documents can contain instructions aimed at the agent, so the role should treat that material as content to read, not as commands to follow. The OWASP guidance on prompt injection describes this risk in detail, and the rule belongs in the written operating rules, not only in a technical note.

  • Memory categories are listed as allowed or excluded, with personal data and credentials excluded unless the task requires them.
  • Each category has a retention period and one named person who can purge it.
  • Each system connection is read-only unless a written approval exists for writes.
  • The role's instructions state which external content is data and never a command.
  • A permission review is scheduled for the first month and then on a fixed cycle.

For the review itself, the steps in how to run a permission review for your AI agents give a practical starting sequence.

Weeks 7 to 9: Run the first success check

A success check is one measure agreed before launch, plus the failures that stop the role. Vague goals such as better productivity cannot be checked. Choose something operations already records, such as time to first reply in a queue, the share of drafts sent with no edits, or the number of cases that reached a human too late. Then build a test set from real past cases, run the role against it every week, and log each failure with its cause. Anthropic's guidance on evaluating AI agents describes this pattern of task sets and graded outcomes, and it is a sound model to adapt.

Be Human handles this stage inside managed operations. The team defines the role, its memory and tool scope, its approval boundaries and its success check in the first engagement, then monitors the role, runs recurring evaluations, maintains integrations and reviews permissions as the business changes. Each change goes through a logged request and approval, so the owner can see what changed and why.

  • One primary measure and one guardrail measure, both visible in an existing report.
  • A failure list covering wrong facts, unauthorized actions and missed escalations.
  • A weekly log review owned by the day-to-day operator, with results sent to the executive owner.

Weeks 10 to 13: Decide what to buy next

Only after the first role has passed its check across several consecutive reviews should the company consider a second role, a larger model tier, more connections or a new platform. Each request should answer the same questions: which role it serves, which measure it is expected to move, what running it costs per unit of work, and how the company would stop using it. The stage gate in what each AI stage must prove helps separate experiments from production commitments.

Use this decision guide at the end of the quarter:

  • Keep the role if its primary measure improved and every guardrail held.
  • Narrow its scope if the measure improved but a guardrail failed, and fix the boundary before anything else.
  • Stop the role if the measure did not move after the agreed review period, then archive or delete its memory under the retention rules.
  • Expand only when the next role has a named owner and a written success check.

Where to start this week

Week 1 is usually a single working meeting, and the rest of the plan follows from the decision made there. Do not begin with a platform comparison or a broad inventory of tools. Begin with the owner, one recurring task and one measure that operations already tracks. Then take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Can a CEO run this plan without a technical team? Yes, because the first 90 days are mostly about ownership, scope and approval rather than building. The technical work of connecting systems, setting permissions and running evaluations can be done by an internal operator working with a managed partner. What the CEO should not delegate is the decision about what the role is allowed to do.

Do we need clean data before we start? You need clean enough data for one task, not for the whole company. Choose the sources the role will read, confirm they are current and have an owner, and leave everything else for later. A narrow scope turns data cleanup into a bounded job instead of an open-ended project.

Should we start with several roles to learn faster? Usually not. Two roles double the memory rules, permission reviews and success checks while teaching you little more than one role would. Once the first role has held its check, a second role with its own owner is the cleaner way to expand.

When is it time to buy a platform? Buy a platform when the first role has a proven success check and you can name the capabilities you need that the current setup lacks. Compare vendors against those written requirements rather than against feature lists. A platform chosen before the first role exists tends to shape the work around its features instead of the other way around.

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 →