Superintelligence and Intellectual Property: What to Protect

Superintelligence and Intellectual Property: What to Protect

Superintelligence is still a hypothesis, but increasingly capable AI agents already change which of your company's intellectual assets matter most: documented processes, decision criteria, knowledge bases, proprietary data, evaluations and code. Protecting them comes down to three fronts: vendor contracts, access control and records that prove origin and ownership. The more an agent can do with what it knows about your company, the more it matters to decide what it can read, store and reuse.

Key takeaways

• Value shifts from the model, which any competitor can access, to the context you give it: processes, criteria, data and examples.
• Ideas and methods are generally not protected by copyright. Your know-how is protected in practice through confidentiality, contracts and access control.
• Contracts with AI vendors and consultants must say who owns prompts, knowledge bases, tuning and outputs, and whether your data trains third-party models.
• The agent should have the least access possible, with its own credentials, knowledge bases separated by department and a log of everything it reads and sends.
• An inventory, version control and dated records are the proof that the asset exists, is yours and was under reasonable care.

Why does intellectual property matter more as agents get more capable?

A limited agent produces generic text. A capable agent carries out the way your company works: it answers according to your discount policy, builds the proposal with the right exceptions and escalates the case that needs a person. That only happens because someone gave it the context. The models themselves are sold to everyone. What your competitor does not have is your way of deciding.

It helps to separate three layers of claims. In the vendors' official documentation, you find what each plan says about data use and retention. That changes by plan and contract, so read the version currently in force. In implementation experience, you find the observation that the hardest asset to rebuild is usually the decision criterion nobody wrote down. In hypothesis, you find the idea that if agents become far more capable, companies with well-recorded processes will gain more from them. That is a hypothesis, not a fact.

To separate what your company operates today from speculation, it helps to read the difference between AGI, ASI and AI agents. The decision to protect your assets does not depend on superintelligence arriving: it already makes sense with today's agents.

Which intellectual assets gain value?

Not everything your company knows becomes an asset. The ones that gain value with agents are those a system can read, apply and repeat:

• Processes and playbooks: the step-by-step for support, quoting, after-sales and collections, exceptions included.
• Decision criteria: when to grant a special condition, when to decline, when to call a manager.
• Curated knowledge base: manuals, validated answers and resolution history reviewed by someone who understands the work.
• Proprietary operating data: conversations, tickets, service orders and purchase orders, always respecting LGPD (Brazil's data protection law).
• Instructions, tools and test cases: what defines the agent's behavior and how you measure whether it got it right.
• Code and integrations: connectors, automation workflows and scripts that link the agent to your systems.
• Brand, tone of voice and creative materials the agent uses when speaking on behalf of the company.

Test cases are the most underrated asset. A set of real situations with the expected answer is, in practice, the written definition of good work in your company. Whoever has it can switch models or vendors without losing the standard. The company superintelligence layer shows where each of these items fits in the architecture.

What does the law protect, and what is up to you?

In Brazil, computer programs follow their own regime, similar to copyright, and texts, images and creative materials are covered by the copyright law. Trademarks are protected through registration with INPI (the Brazilian industrial property office), under the industrial property law. So far, this is formal protection.

The point to watch is that ideas, methods and procedures as such are generally not protected by copyright. Your sales playbook, as a method, depends on another kind of protection: treating it as confidential information. Industrial property legislation punishes the misuse of confidential information as unfair competition, but that defense is stronger when the company can show it took reasonable confidentiality measures: contracts, restricted access, a record of who read what.

There is also an open area: the ownership of AI-generated content is still under debate, and you should not assume the agent's output is automatically yours or protected. For decisions that depend on it, such as publishing a creative piece or selling generated code, get an opinion from an intellectual property lawyer. To follow the regulation under discussion, see the article on Brazil's AI legal framework.

What should you review in contracts?

The contract is where protection starts, because it defines what the vendor can do with what you send. Use this list when reviewing an AI vendor, an integrator or a consultant:

• Training: can your data and conversations be used to train the vendor's models? This changes by plan, so check the contract and the terms of the plan you bought.
• Retention and deletion: how long prompts, files and logs are kept, and how you request deletion, including from backups.
• Ownership: who owns the instructions, vector databases, fine-tuning and outputs generated for you.
• Subprocessors and location: who else touches the data and in which country. See what LGPD requires in international data transfer when using AI providers.
• Confidentiality: who inside the vendor can access the content and under what conditions.
• Liability: what happens if an output infringes third-party rights.
• Portability: how you take your prompts, knowledge bases, evaluations and logs with you when the contract ends.
• Assignment in development services: custom code, documentation and workflows should have clear ownership, separating reusable generic components from what is specific to your business.

The same care applies internally. A company policy for employee ChatGPT use and confidentiality clauses with employees keep strategic content from being pasted into a personal tool.

How do you control what the agent can access?

An agent with many permissions is a potential leak channel, whether through error or through malicious instructions hidden in an email or document (prompt injection). The list of top risks for applications built on language models published by the OWASP project includes disclosure of sensitive information. OWASP's authorization best practices also apply directly: least privilege and deny by default.

In practice, this becomes simple rules:

• Give the agent only the sources needed for that task. A support agent does not need to read the margins spreadsheet.
• Use the agent's own revocable credentials, not a person's password.
• Separate knowledge bases by department, so commercial knowledge does not mix with financial data.
• Log what the agent read, what it sent and to whom, including attached files.
• Require human approval before any strategic material leaves the company, whether by email or WhatsApp.
• For highly sensitive data, consider running a local model, knowing it requires your own operation and maintenance.

The NIST AI Risk Management Framework helps organize this conversation in terms of governance, mapping and measuring risk, without prescribing a specific tool.

How do you prove the asset is yours?

Protection also means being able to show origin, date and owner. Without records, a dispute becomes one person's word against another's. The minimum viable approach is a living inventory, with these fields for each asset:

• Name and description of the asset.
• Internal owner and owning department.
• Where it is stored and in which version.
• Who has access, people and agents.
• Which vendors touch it and under which contract.
• Classification: public, internal or confidential.

Add version control for instructions, workflows and code, with a history of who changed what and when. Keep the human authorship on record: who wrote the criterion, who reviewed it, who approved it. Where it makes sense, use formal registrations: a trademark with INPI and, for significant software, the computer program registration, which is optional but helps date the work and prove authorship. Finally, keep the agent's decision log, because it shows the knowledge was applied within defined limits.

A step-by-step plan and an illustrative scenario

A lean sequence to start, without waiting for any superintelligence:

• List the assets the agent would use or generate and classify them by sensitivity.
• Decide what never leaves the company, what can go to a vendor under a suitable contract and what is public.
• Review vendor contracts using the list from the previous section and record the gaps.
• Set up minimum access, dedicated credentials and logs before connecting the agent to real data.
• Run a pilot on a single workflow, with an agreed success metric and a review of the logs.
• At the end, update the inventory and decide: expand, adjust or stop.

Illustrative scenario: a parts distributor keeps, in the head of one sales manager, the rules for when to offer a special condition to each type of customer. When deploying a support agent, the company writes those rules in a versioned document, keeps the cost table out of what the agent can query and adds to the vendor contract a ban on training with its data. Months later, the manager leaves. The criterion is still on record, with author, date and revisions, and the agent keeps applying the standard. The asset no longer depends on one person or one vendor.

This is how Juan Carlo's method fits: map the process and what counts as an asset, limit the agent's data and actions, pilot one real workflow with logs and an agreed metric, and hand over the operation with documentation and a decision on the next step. None of these steps replaces a lawyer's opinion when the question is about ownership or contracts.

Frequently asked questions

Does superintelligence change intellectual property rules? Not by itself. Current rules apply to today's agents, and superintelligence remains a hypothesis. What changes is the practical importance of protecting processes, data and criteria, because a more capable agent extracts more value from them.

Does the content the agent generates belong to my company? It depends on the vendor contract and on a topic still under debate: authorship of AI-generated content. So record the human contribution, read the terms of your plan and consult a lawyer before building a business on a specific output.

Does using AI in my company expose my trade secrets? It can, if sensitive data is sent without a contract, without access control and without an internal policy. With clear clauses on training and retention, minimum access and human approval for external sends, the risk is manageable.

Where do I start if nothing is documented? Start with the inventory of a single workflow, such as support or quoting. Write down the criteria, define who owns it, classify its sensitivity and only then choose the vendor and the contract.

Sources

• OWASP: Top 10 for LLM Applications
• OWASP: Authorization Cheat Sheet
• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados