How to Talk About Superintelligence With Clients Without Overpromising
Plain language for discussing superintelligence and AI agents with clients, without hype, dates or guarantees, plus a checklist of safe claims.

To talk about superintelligence and AI agents with clients, describe the work the agent will do, the systems it can touch, the points where a named person must approve its actions, and the test that shows whether it works. Leave out timelines, percentages, guarantees and predictions about future capability that you cannot demonstrate today. Specific, bounded statements build more trust than large forecasts, because the client can check them.
Key takeaways
- Start with the job, the tools and the approval points. Future capability can wait until the client asks about it directly.
- Do not attach dates, percentages, prices or guarantees to claims about AI behavior you have not tested in the client's own process.
- Replace "it will be safe" with "we will test it on past requests and review the results with you before go-live."
- Name who answers for a mistake and how a person can pause the agent.
- Treat superintelligence as an open research and policy question, and answer direct questions with sober, sourced facts.
Why the word superintelligence causes trouble in client meetings
The word arrives with expectations from films, headlines and social media. A client may hear a promise that a system will outperform every expert in the room, or a warning that the business will soon be made obsolete. Both versions distort the decision in front of them. Your task is to bring the conversation back to something the client can operate: a task, a memory store, a set of tools, a permission level and a review routine.
Be direct about uncertainty. Researchers do not agree on when more capable systems will arrive or how they will behave in every setting, and nobody can responsibly give a client a date. Saying this plainly is accurate, and it lets you speak with confidence about what is already in front of you. For a reading list that separates research claims from operating facts, see Superintelligence Safety Reading List for Business Leaders.
A plain vocabulary for clients
Clients who hear technical terms without definitions tend to assume either the worst or the best. Define each word in the first meeting, then use the same words in every proposal.
- Agent: software that receives a request, uses named tools and takes steps toward a defined goal within limits you set.
- Memory: information the agent keeps between sessions. It should be listed, tied to a purpose and deletable on request.
- Tool: a system the agent reads from or writes to, such as a calendar, a ticket queue or a customer record.
- Permission: the exact actions allowed on each tool, usually narrower than the tool itself allows.
- Approval point: a step where a named person confirms before the agent continues.
- Superintelligence: a hypothetical class of AI systems that would exceed human ability across most intellectual tasks. Describe it as a research and policy topic, not as a product you sell today.
Anthropic's guidance on building effective agents makes a similar point: begin with simple, well-bounded patterns and add complexity only when a real task requires it. That principle is easy to explain to a client and hard to argue with.
What to say, what to avoid and why
Statements a client can hold you to:
- "The agent drafts replies in the support queue and cannot send them without approval from the account owner."
- "We will run the agent on past requests first and review the results with you before any customer sees it."
- "If the agent makes an error in this area, the named owner reviews the log and we narrow its permission."
- "We will track how often drafts need edits over several review cycles, and we will show you the record."
Statements to avoid:
- "It will never make a mistake." No honest provider can promise this.
- "It will be fully autonomous next year." This is a dated forecast you cannot support.
- "It is certified safe." Unless a specific certification applies to this exact use, do not imply one.
- "It will cut your costs by a set amount." Any savings figure should come from measurement in the client's process, not from a vendor presentation.
A step-by-step plan for the first conversation
Use this sequence in the first one or two meetings. It keeps the discussion on the work and makes the open questions visible.
- Start with the client's task in their own words, and write it down before naming any technology.
- List every system the agent would read from or write to, and name the person who owns each one.
- Mark which actions can be reversed easily and which need a named person to approve them.
- Agree on one success criterion the owner can observe, such as the share of drafts accepted with no edits in a sample the owner reviews each week.
- Write down what the agent will remember, where that memory lives and how a request to delete it will be handled.
- Set the first review date, name who attends it, and state in advance which result would lead to reducing the agent's permissions.
- Close by saying what you do not yet know and how you will find out. Honest gaps make a plan more credible, not less.
Illustrative scenario: a clinic owner asks about the future
This scenario is illustrative and does not describe a real client. The owner of a small physiotherapy practice asks whether superintelligent AI will make the clinic obsolete, and whether the team should worry now.
A weak answer offers reassurance ("nothing will change for years") or alarm ("everything will change soon"). Both are forecasts. A better answer separates two questions. The long-range question is open, and no one can responsibly predict it. The practice's near-term question is narrower: how to send appointment reminders and collect intake forms without adding work at the front desk.
The proposed first role is an agent that drafts reminder messages from the schedule and prepares intake summaries. It may not change appointments or send messages without staff approval. Its memory holds only the scheduling preferences patients gave, and the front desk can delete any entry on request. The success criterion is simple: staff review a sample each week and record how many drafts they change. The owner leaves with a defined scope, a review routine and an honest answer about the future, which is more useful than either forecast.
How to keep control visible in the engagement
Most client questions about superintelligence are really questions about control. Who can pause the agent? What can it remember? Which tools can it change? These have answers you can show in a document and in a test. The NIST AI Risk Management Framework organizes this work into govern, map, measure and manage, which gives a business owner a clear structure for discussing risk without inventing new vocabulary.
Treat each approval step as a design decision. Wherever a person must sign, the client's judgment outranks the agent's. For guidance on choosing those steps, read Where should your AI agent stop and ask?. Grant permissions narrowly at the start and widen them only when the evidence supports it, as described in AI Agent Autonomy Levels: Grant Each Step Only After Evidence.
Be Human designs the first engagement around one useful role, the memory and tools that role needs, the approval boundaries and one observable success criterion. Managed operations then cover monitoring, recurring evaluations, permission reviews and incident response, so conversations with the client stay tied to what the running system actually did rather than to forecasts about what it might do later. The client keeps control of the agent's identity, memory, channels and permissions throughout. Service is remote, in English and Portuguese.
Where to start the conversation this month
Begin with an internal review before any client call. Ask the team to collect every sentence about future AI capability that appears in proposals, slides or support scripts. Each one needs a source, a test or removal. The OECD AI Principles are a short, stable reference for transparency, robustness and accountability, and they give your team shared language that does not depend on a particular product.
Keep the first client conversation short and concrete. Bring one task, one approval point and one review date. A client who leaves with those three things has something to check, and that is what builds trust over time. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Should I mention superintelligence at all in a sales conversation? Mention it only when the client raises it, and then answer in two parts: what is unknown about the long-range question, and what you propose for the next few months. Do not make it the centerpiece of a proposal, because it pulls attention away from the task and invites forecasts you cannot support. A short, honest answer usually closes the topic faster than a long one.
What if a client asks when AI will be better than people at their job? Say that no one can give a reliable date, including you, and explain that the decision in front of them does not depend on that date. Describe which tasks the agent handles now, which stay with people, and how performance will be reviewed over time. Offer to revisit the scope at each review rather than predicting the next stage.
Can I promise that an agent will not make mistakes? No, and you should not try. Promise the controls instead: narrow permissions, approval points, logs of what the agent did, a named owner for errors and a way to pause the agent. These can be tested and written down, which makes them commitments you can actually keep. For how responsibility is assigned, read Who Is Accountable When an AI Agent Makes a Mistake?.
How do I answer a client who fears AI will replace their staff? Acknowledge the concern as reasonable, then describe the specific tasks being moved, the tasks that stay with people and who approves any change. Involve the staff who do the work in the first review, since they know where the process breaks. Do not promise that no roles will change unless you can commit to that in writing for this specific scope.
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.


