Personalized AI infrastructure
Superintelligence Due Diligence: Agent Assets to Check Before You Buy
A buyer's checklist for agent memory, prompts, tool access and permission records in a target company, with what to verify before valuing them as assets.

Superintelligence due diligence for an acquisition means checking whether a target's AI agents are controllable assets or a set of undocumented habits that will stop working after closing. An acquirer should verify four things: what the agent remembers and where that memory came from, which prompts and instructions shape its behavior, which tools it can call and under what credentials, and whether permission records show who approved what. Treat each item as an asset only after it can be exported, explained and tested without the original builder in the room.
Key takeaways
- Agent memory is an asset only if you can name its sources, its owners and its deletion rules.
- Prompts and instructions are operating documents, so you need their version history, not only the current text.
- Tool access should be inventoried by credential, scope and allowed action, not by the name of the tool.
- Permission records must prove that approvals happened, which means matching them to actual write actions.
- Ask for an export and a restore test before you place any value on the agent layer.
Why agent assets differ from ordinary software assets
Standard diligence reviews code, contracts, data rooms and payroll. An AI agent combines a model, instructions, memory, tools and permissions, and its behavior can shift when any one of those changes. A prompt edited last month or a memory store that grew through a year of customer conversations can change what the agent does without any change to the software repository. That makes the agent partly a configuration and partly a record of how it was shaped.
The NIST AI Risk Management Framework is a useful reference for this review because it treats AI risk as something managed across the life of a system, with governance, mapping, measurement and management as separate activities. A buyer does not need to adopt the framework to borrow its questions. Ask who governs the system, what the system is known to do, how its behavior is measured, and who acts when it fails. For the layers involved, the superintelligence stack explained by layer is a helpful reference when you need to name each component in the target's documentation.
What to check in agent memory
Memory is where most hidden value and most hidden risk sit. It may contain customer preferences, supplier terms, internal decisions, staff notes and summaries of documents that were never meant to leave the company. Before you treat memory as an asset, establish what it contains, where each category came from and who is allowed to write to it.
Use this checklist for the memory review:
- Source of each memory category, such as staff notes, customer conversations, uploaded documents or model-generated summaries.
- Whether personal data is present, and what the target relied on to keep it, including consent or contract terms.
- A named owner for each memory store and a list of who or what can write to it.
- A deletion and correction path that has been tested on a sample record, not only described in a policy.
- Whether the memory holds the target's own know-how in a form that would transfer to you, and whether the seller has the right to transfer it. The guide on protecting intellectual property in superintelligence systems covers how to think about this boundary.
If the seller cannot tell you where a memory entry came from, mark it as unverified. Unverified memory can still be useful to the business, but it should not be valued as if it were a documented asset.
How to review prompts and instructions
Prompts, system instructions, tool descriptions and templates decide how an agent reasons and what it is willing to do. Many targets keep them in a single text box, edited by whoever was available that week. Your task is to recover the full set, understand the history of changes and confirm that the instructions do not weaken controls.
One risk deserves particular attention. Agents often read outside content such as emails, web pages, documents and customer messages, and that content can contain instructions meant to change the agent's behavior. The OWASP prompt injection prevention guidance describes this class of risk and the defensive habits that reduce it, such as separating trusted instructions from untrusted input and limiting what the agent can do with content it has read. Ask the target how it handles this, and look for evidence rather than assurance.
A practical prompt review includes these steps:
1. Export every system instruction, prompt template and tool description as files, including any kept in dashboards or notes. 2. Request the version history, with the author and reason for each significant change. 3. Search for instructions that tell the agent to skip approvals, hide actions or treat certain requests as exempt from logging. 4. Identify which inputs the agent reads from outside the company and whether those inputs can alter instructions or memory. 5. Ask for the evaluation set used to test the current prompts, and whether it has been run after the latest change.
Tool access: inventory by action, not by name
A tool called "CRM" or "email" tells you very little. What matters is the action it allows, the credential it uses and the scope of that credential. Read access to a folder is a different asset from the ability to send messages to every customer or to delete records. Least privilege is the standard principle here, and the OWASP authorization guidance describes how access should be granted narrowly and checked on each request rather than assumed.
For each tool, record the following: what it can read, what it can write or delete, which account or key it uses, who created that credential, whether it expires, and whether it can be revoked without stopping the business. Many acquisition problems start with shared credentials, service accounts created by former employees, or keys stored in configuration files that nobody reviews.
A tool inventory is complete only when every credential the agent uses appears in it, with an owner who can still be reached.
Illustrative scenario: a target's support agent reads a helpdesk and can also issue refunds through a payment key created by a former employee and shared with two other systems. The buyer's first question is not whether the agent answers support tickets well. It is who can revoke that key, what breaks if it is revoked, and whether the refund actions appear in any approval record. Until those answers are documented, the refund capability is a liability with a useful feature attached.
Permission records: evidence that approvals really happened
Permission records are the evidence that the agent operated inside boundaries the company chose. They include the approval thresholds, the people who granted access, the exceptions and the logs of actions taken. A common finding is that a policy exists, but the logs do not match it. The agent performed actions that the policy would have required someone to approve, and no approval exists.
To test this, sample recent write actions and trace each one back to the approval that authorized it. Observability practice helps here. The OpenTelemetry documentation on observability concepts describes how traces, metrics and logs connect a system's actions to the events that produced them. If the target records agent actions in that kind of structured form, you can reconstruct its history. If it does not, you are relying on memory and screenshots.
When reviewing permission records, look for these conditions:
- Write actions with no matching approval record.
- Approvals granted by people who have left the company or who did not hold authority over the system.
- Standing permissions with no review date or no named reviewer.
- Exceptions recorded informally in chat messages rather than in the permission register.
- Logs that can be edited or deleted by the same people whose actions they record.
For a deeper look at how a permission review is run once a system is in your hands, the step-by-step guide to permission reviews for AI agents lays out the evidence each step should produce, and that same evidence is what you should ask the seller to show you during diligence.
Valuing the agent layer and deciding what to put in the deal
After the four reviews, the agent layer should be recorded in a register with one line per asset: the component, its owner, the evidence supporting it, the known risks and the effort needed to rebuild it if it cannot be transferred. This register gives the deal team a concrete object to negotiate around, instead of a general statement that the target "has AI."
Use a simple classification to guide the discussion:
- Transferable asset: memory with documented sources, an export that has been tested, a named owner, prompts with version history, and permission logs that reconcile with actions taken.
- Conditional asset: the system works, but depends on one builder, has partial logs, or contains memory whose rights are unclear. Make resolution of the gap a condition of closing or a post-closing task with an owner.
- Liability: unknown credentials, personal data without a clear basis, approvals that cannot be traced, or instructions that weaken controls. Price the cost of remediation and decide whether the component should be shut down rather than transferred.
The guide on what an AI due diligence checklist should cover for a buyer and the ownership questions to ask an implementation partner are useful companions when the same questions come up after closing.
How Be Human handles this: Be Human designs, implements and operates personalized AI infrastructure in which an agent's memory, prompts, tools and permission records are kept as named, exportable components under the client's control. A managed engagement documents each of these from the first defined role, keeps a register of tools and credentials with owners, and runs scheduled permission reviews, so an acquirer's review starts from evidence rather than reconstruction. The same structure makes a handover to a new owner a matter of transferring documented parts.
When you are ready to decide which agent in the target is worth keeping, the free Be Human AI diagnostic can help you identify the first task worth handing to an agent.
Frequently asked questions
Can an AI agent's memory be treated as a company asset in a deal? It can, once its sources, owners, rights and deletion path are documented and the export has been tested. Memory that cannot be explained or transferred should be treated as conditional or as a liability until the gap is closed. The deal documents should name the memory stores explicitly, not refer to the agent in general terms.
Should I ask for the prompts or only for the outputs the agent produces? Ask for both, but start with the prompts and their version history. Outputs show what the agent did in a few cases, while prompts and instructions show what it is designed to do and what it may be pushed to do by outside content. Without the instructions, you cannot judge whether a behavior is a stable feature or an accident.
Does an AI review replace security and privacy diligence? No. The AI review covers questions that standard security and privacy work often misses, such as how memory accumulates, how prompts shape actions and whether approvals match what the agent did. It should run alongside those reviews, with the same data room, the same people and a shared list of open findings.
What if the target cannot produce its permission logs? Treat the missing logs as a finding with financial and operational consequences, not as a formality to waive. Estimate the cost of rebuilding the evidence or restricting the agent's write permissions, and decide whether that cost belongs in the price, in escrow terms or in a condition to closing. A target that cannot show what its agent did is not yet offering a transferable asset.
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.


