How to Write a Company Position on Superintelligent AI

A company position on superintelligent AI is a one-page internal paper in which the leadership team states what it believes, which principles it follows, which limits it will not cross and which decisions it has agreed to make. Write it in four short parts, have every leader sign the same page, and review it on a fixed schedule. The aim is not to forecast the future but to stop each leader from improvising a different answer when a vendor, a board member or an employee asks.

Key takeaways

• Keep the position to one page with four parts: beliefs, principles, limits and decisions. If it does not fit, it is not yet agreed.
• Separate what you believe from what you know. Label uncertain beliefs as uncertain, and say what would change your mind.
• Write limits as actions the company will not take, not as hopes. A limit that names no action cannot be checked.
• Give each decision a named owner and a trigger, so the paper changes behavior instead of sitting in a folder.
• Review the page on a schedule and after any real event, and keep the revision history.

Why does a company need a written position at all?

Superintelligent AI, meaning systems that would outperform people across most knowledge work, is a topic on which informed people disagree about timing, form and consequences. Your company does not need to settle that debate. It does need to decide how it behaves while the debate continues.

Without a written position, three things tend to happen. Different leaders give different answers to the same question, so customers, staff and partners hear mixed messages. Vendors set the frame, because a company with no stated view ends up accepting the view in the sales presentation. And decisions about data, permissions and autonomy get made one project at a time, with no shared reasoning to check them against.

A position paper fixes this at low cost. It is different from a policy. A policy lists rules for staff. A position explains what leadership thinks and why, then derives a few commitments from it. Policies, such as the ones described in AI governance for mid-sized companies, can then be written as consequences of the position instead of as free-standing rules that nobody can trace back to a reason.

The one-page limit is deliberate. A long document gets agreed by nobody in particular, because each reader finds the paragraph they like and skips the rest. A single page forces trade-offs, and trade-offs are exactly what leadership needs to have settled before the pressure arrives.

What belongs in the beliefs section?

Beliefs are the honest foundation. Keep them to a handful of sentences, and mark each one with how sure you are. A useful format is to write the belief, then a short note on your confidence and what evidence would revise it.

Here are example belief statements, written as prompts for your own team to accept, rewrite or reject:

• Model capability will keep improving, but the pace and the shape of that improvement are uncertain for us.
• Our value sits in our context: our data, processes, judgment and relationships, not in any single model.
• Highly capable systems will still make errors that look confident, so human oversight stays necessary for consequential actions.
• Some of our current work will change materially, and we would rather plan for that openly than react to it.

Notice what is absent: no date predictions, no claim that superintelligence is near or far, no promise about jobs. Leadership teams that try to agree on a timeline usually stall, and the argument rarely changes what they would do on Monday. Agree instead on what you would do under several possible timelines. If you want a way to structure that, the three horizons in superintelligence strategy for mid-sized companies are a practical companion.

Apply one test to every belief: could a thoughtful colleague disagree with it? If not, it is probably too vague to guide anything. A statement such as AI is important for our future passes no test, because nobody would argue and nothing follows from it.

How do principles and limits turn beliefs into rules?

Principles say how you will behave. Limits say what you will not do. Keep both short and tie each to a belief above. If a principle has no belief behind it, either the belief is missing or the principle is decoration.

Principles should be few enough to remember. Candidates, to adapt to your business:

• Humans stay accountable. Every AI agent has a named human owner, and consequential actions require approval.
• Context stays ours. Company knowledge, memory and configuration live in systems we control and can export.
• Capability is earned in stages. Autonomy grows only after evidence gathered in a narrower setting.
• We can stop it. Every deployed agent can be paused, and we know how, who decides and what happens next.
• We tell the truth about AI. Staff, customers and partners are told when they interact with an AI system and what it can and cannot do.

Limits are sharper. Write each as an action you will refuse to take, so that someone can check compliance. Examples:

• We will not give an agent permission to move money, change legal terms or delete records without a human approval step.
• We will not let an agent act in a channel or system that has not been reviewed and listed.
• We will not adopt a tool whose vendor cannot explain where our data is stored and how we can leave.
• We will not deploy any system we do not know how to switch off.

These limits are deliberately about operating AI today, not about hypothetical machines. That is a feature. A position that only addresses distant scenarios gives nobody anything to do. A position that addresses today's agents, with the future in mind, is easier to apply and easier to test.

External frameworks can sharpen your wording. The NIST AI Risk Management Framework and the OECD AI Principles both offer vocabulary on accountability, transparency and risk that you can borrow without adopting either wholesale. If you operate in or sell into the European Union, the EU AI Act overview is worth reading as input to your limits, and your legal advisers should confirm which obligations apply to you.

Which decisions should the position settle in advance?

Decisions are where the paper earns its place. List the choices you expect to face, who owns each, and what triggers it. Then you are not debating from scratch under pressure, with a deadline and a persuasive vendor in the room.

Use this decision guide as a starting list and edit it to fit your company:

• Capability trigger: what observed change in tools or models makes leadership reconvene, and who calls the meeting?
• Autonomy increase: who may approve a wider scope for an agent, and what evidence is required first?
• Vendor commitment: who signs off on any long agreement, and which exit terms are mandatory?
• Data exposure: who decides which classes of data may reach an external model, and in which region?
• Incident response: who can pause an agent immediately, and who informs customers or regulators if needed?
• Workforce change: who explains to staff how roles may shift, and when does that conversation begin?
• Review cadence: who calls the next review of this paper, and what happens if it is skipped?

The last item matters more than it looks. If nobody owns the paper, nobody revises it, and a stale position is worse than none because people keep citing it. The owner is usually an executive sponsor, a role described in executive sponsor for an AI project. Ask each named owner whether they accept the role before the page is signed. A decision assigned to someone who never agreed to it is a gap with a name on it.

How do we write it, step by step?

This plan fits into a few working sessions with a small group. It assumes the leadership team can meet together at least a couple of times.

• Step one: choose a drafter and a sponsor. One person holds the pen, and one executive owns the outcome.
• Step two: collect real questions. Ask leaders, managers and a few frontline staff what they are asked or worried about regarding advanced AI. Real questions make a better skeleton than abstract themes.
• Step three: draft beliefs separately, then compare. Each leader writes their own beliefs alone. The differences show where the discussion is needed.
• Step four: derive principles and limits. For each agreed belief, write the commitment it implies. Cut any commitment that no belief supports.
• Step five: assign decisions. Name owners and triggers, and confirm each owner accepts.
• Step six: stress test the page against three situations and edit wherever it fails.
• Step seven: sign, share and schedule. Circulate a plain-language version to staff, set the next review date and file the revision history.

Illustrative scenario: a regional services company drafts its position. In step three, the founder writes that the firm should adopt the most capable model as soon as it appears. The operations lead writes that nothing should change until a pilot has proven it. Neither is wrong, but one page cannot say both. The team resolves it with a principle that capability is earned in stages, and a decision that the sponsor may approve a trial of a new model in a narrow setting, with a review before any wider use. The disagreement becomes a rule both leaders can follow.

For step six, three stress tests work well. First, a vendor claims its system is superintelligent and asks for broad access to company systems. Second, an employee asks whether their role is at risk. Third, a customer asks whether AI is making decisions about them. If the page helps you answer each one consistently, it works. For the first test, superintelligence vendor evaluation lists the evidence and questions to ask. Where the page is silent or contradictory, edit it now, not during the real event.

How does Be Human support this work?

Be Human designs, implements and operates personalized AI infrastructure for companies. A first engagement defines one useful role for an agent, the memory and tools it needs, the approval boundaries and an observable success criterion. A position paper feeds directly into those boundaries: limits become permission settings, and decisions become approval paths. Managed operations can include monitoring, recurring evaluations, permission review and controlled improvements, so the paper has a working counterpart in how agents actually run. Be Human does not write your beliefs for you, because those stay with your leadership. Service is remote, in English and Portuguese.

Frequently asked questions

Do we need a position if we do not use AI agents yet? Yes, and it is easier to write before you start. Employees and vendors are already bringing AI into your company, and a written position gives you a shared answer before the first agent arrives. It also makes your first deployment decisions simpler, because limits and approval owners are already agreed.

Should the paper predict when superintelligent AI will arrive? No. Forecasts are the part leadership teams argue about longest and the part that changes your actions least. State your uncertainty plainly, then write commitments that hold whether progress is fast or slow.

How often should we review the position? Set a fixed interval and also review after real events, such as a major change in the models you rely on, an incident, a new regulation or a significant vendor change. Many teams align the review with their regular permission review so that limits on the page match the permissions in the systems. Record each revision with a date and a short reason.

Should we publish the position outside the company? Start with the internal version and a plain-language summary for staff. Publish externally only after legal review, and only include commitments you can keep and demonstrate. A public statement that outruns your practice creates more risk than silence.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.