Does Superintelligence Create a Business Moat? What Companies Keep

A capable AI agent does not create a business moat on its own. Model intelligence is sold through the same providers to competitors, so an advantage that depends only on the model tends to be temporary. What lasts is what a company builds around the model: data it has the right to use, memory and decision records it owns, processes it has written and tested, and permissions it can review and revoke.

Key takeaways

• The model alone is rarely a moat, because rivals can usually reach the same capability through the same providers.
• Advantages that stay are the ones a company can write down, verify, export and move to another system.
• Data helps only when the company holds clear rights to it, keeps it current and can explain what each field means.
• Learning compounds only when lessons are captured in records the company controls, not in one person's chat history.
• For each claimed advantage, ask whether it would still belong to the company if the model provider changed tomorrow.

What a moat means when the tools are shared

A moat is a durable advantage that a competitor cannot copy quickly or cheaply. Companies have traditionally built moats from scale, distribution, proprietary assets and switching costs. Capable agents change how quickly work gets done, but they also make underlying capability easier to rent. If superintelligence arrives, this effect would likely intensify for whoever controls the most capable systems. That future is uncertain, and a leadership team cannot plan its whole strategy around it. It can, however, examine what it controls now and what would still be its own if the tools changed.

The useful question is not whether our model will be better than theirs. The useful question is where the advantage lives and who holds it. An advantage that sits with a model vendor belongs to the vendor first. An advantage that sits in the company's records, rules and evaluations can be applied to whichever model the company chooses later. That distinction should shape decisions about data, memory and contracts.

What tends to stay with the company

Several kinds of assets tend to survive model changes:

• Operational records: customer history, order exceptions and the reasons behind past decisions, stored in formats the company can export.
• Operating memory: the context an agent needs, such as approved terms, escalation paths and product rules, kept as documents the team can read and edit.
• Evaluation sets: examples of good and bad outcomes, written by people who do the work, used to test any new model or prompt before it reaches customers.
• Process definitions: the steps, handoffs and approval boundaries that describe how work is supposed to flow.
• Permission and audit history: a record of what each agent was allowed to touch, what it did, and who approved changes.
• Relationships and trust: customer and partner confidence, which no agent can transfer to a competitor.

The first three are the most often underestimated. An evaluation set in particular turns a vague claim of quality into something a team can run again after every change. Without one, a company cannot tell whether a newer model actually serves its customers better or only sounds more impressive in a demonstration.

What does not stay, and why

Some apparent advantages decay quickly:

• General reasoning or writing quality that major providers offer at similar levels for the same kinds of tasks.
• Prompts that worked once and were never saved or tested.
• Generic features that a vendor ships to every customer at the same time.
• Speed gained by removing approval steps, which competitors can copy and which creates exposure when something goes wrong.
• Activity dashboards that count messages or runs but do not connect to outcomes the business cares about.

The pattern is consistent. A head start in tooling decays because others can buy the same tooling. A head start in recorded knowledge compounds, because each new record makes the next improvement easier to measure and safer to make. Unrecorded knowledge, held in someone's memory or in a single conversation, is more of a risk than an asset, since it leaves when the person does.

A five-step test for any claimed advantage

Use this sequence before you treat any capability as a moat, and write down each answer:

• Name the advantage in one sentence, including who benefits and how you would know it is working.
• Locate the knowledge: in a person's head, a shared drive, a vendor's platform, or a system your company operates.
• Check rights: can you export, correct and delete the data and memory without asking a vendor for permission?
• Check evidence: is there a repeatable test, with recorded outcomes, that shows the advantage holds after changes?
• Check portability: if the model or agent platform changed, which parts would still work, and what would need to be rebuilt?

Most claimed advantages fail at the second or third step. The knowledge is real, but it sits where the company cannot export it, or the rights are unclear. Closing those gaps usually costs less than chasing a more capable model, and it makes the advantage visible to the people who must defend it. For the portability question in particular, how to avoid AI vendor lock-in covers the practical patterns.

An illustrative scenario: same model, different records

Illustrative scenario: two regional equipment distributors, both fictional, deploy the same commercial model to draft quotes for trade customers. Both connect it to their product catalogs at the start, so on the first day their agents look equally capable.

The first distributor keeps its pricing rules, discount approvals and exception decisions inside the vendor's configuration screens. Quote notes are not exported. When it considers changing providers, the team discovers that the rules must be rebuilt from memory and that the exception history cannot be retrieved at all.

The second distributor keeps its discount rules and exception history as documents it owns, links each agent action to a recorded approval, and maintains a set of past quotes with outcomes that sales managers judged correct. When a newer model appears, the team runs its evaluation set against it, sees where the new model performs better and where it performs worse, and switches only the parts that improve. Its advantage is not the model. It is the written record of how the business decides, and that record moves with the company.

How to protect and keep the advantage

Protection is partly technical and partly managerial. Authorization should be enforced by the systems an agent calls, on the server side, rather than by instructions written in a prompt, because a prompt can be influenced by content the agent reads. The OWASP authorization guidance describes this principle of checking access on every request. Each agent should hold only the permissions its role needs, and those permissions should be reviewed on a schedule, as laid out in how to run a permission review for your AI agents.

Traces of what an agent did are also part of the asset. The OpenTelemetry observability concepts describe how logs, metrics and traces let a team reconstruct behavior after the fact, which is the evidence needed to learn from mistakes and to prove that a change helped.

Be Human designs the agent's memory, tools, permissions and approval boundaries in the first engagement, so the company can see which records and rules it holds and who owns them. Managed operations then keep the evaluation set running, review permissions on a schedule, and route changes through a written request and approval process, so know-how accumulates in documentation the company keeps rather than only inside a session. Be Human does not promise a competitive outcome. It keeps the inputs visible, portable and under the company's control.

Where to start this quarter

Start with an inventory rather than a strategy document. List the records, rules, evaluation examples and permission histories your business already produces, and mark which ones you can export today. The AI stack architecture map is a useful one-page way to place each of them. Then choose one process where the knowledge is clear, the data rights are settled and a mistake would be visible and reversible. The NIST AI Risk Management Framework offers a shared vocabulary for mapping and managing those risks as the work grows. Waiting for superintelligence to settle the question is not a plan, because the assets that create advantage are built by the company, one record at a time.

If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Does a more capable model give my company an advantage on its own? Not reliably. Rivals can usually access the same capability, so the model is a starting point rather than a moat. The advantage appears when your records, evaluations and permissions let you apply any model well and check the result.

Which data should a company protect first? Start with data that is hard to recreate, that customers trust you with, or that shows how your decisions were made in the past. Confirm that you hold clear rights to it and can export it without a vendor's help. Those three checks will show quickly where the real exposure sits.

Do AI agents keep learning from our staff, and who owns what they learn? Agents do not carry lessons across sessions unless the system is built to store them. When memory is kept, it should live in documents or systems the company controls, with export and deletion rights written into the contract. Ask your provider where memory is stored before the agent goes live.

Is it worth planning for superintelligence now if the timeline is uncertain? Yes, in a practical sense. Decisions about records, data rights, permissions and evaluations pay off with today's agents as well as with more capable ones. You do not need a forecast to begin.

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