Personalized AI infrastructure
Due Diligence for AI Agents: Memory, Prompts and Permissions to Check
A buyer's checklist for agent ownership, memory exports, prompts and permissions before acquiring a company whose operations depend on its AI layer.

Buying a company whose daily operations run through AI agents means buying its memory, prompts, tool connections and permissions along with its contracts and staff. Before signing, a buyer should confirm in writing who owns each of those assets, whether the memory can be exported in a form another team can read and load, and which credentials or approvals would stop working when control changes hands. The questions below are written for owners and operations leaders who must decide, not only for engineers.
Key takeaways
- Agent configuration, memory and prompts need their own ownership terms, separate from software licenses and vendor agreements.
- A memory export is only valuable if it includes the source documents and can be read without the original vendor's platform.
- Permissions tied to a former employee or a personal account are a closing-day risk, not a minor cleanup task.
- Ask for a tested restore, not a statement that backups exist.
- Treat the AI layer as operating infrastructure in the deal model, with a named owner and a review schedule after closing.
Why the AI layer belongs in acquisition diligence
Standard diligence reviews contracts, financial records, customer data, source code and cloud accounts. Agent configurations often sit outside that list. A company may run an intake agent that answers customer email, a quoting assistant that reads the price book, and a scheduling agent with access to the calendar. Over time these systems accumulate what the business knows: how exceptions are handled, which customers have special terms, which suppliers respond slowly. That knowledge rarely lives in a document anyone has read.
The practical consequence is simple. If the agent layer is treated as a software subscription, the buyer may inherit a working company that cannot explain its own operating decisions or change its agents without asking the seller for help. Diligence should answer three things: whether the business can keep operating on day one, whether it can explain what its agents do and why, and whether it can change those agents after the transaction under its own authority.
Who owns the agents, the memory and the prompts?
Ownership questions come first because every other answer depends on them. Owning a software license is not the same as owning the configuration built on top of it. Request written answers to these questions:
- Which legal entity holds the agent accounts, workspaces and API keys, and are they registered to the company or to an individual?
- Do the vendor terms let the provider use company content to train or improve shared models, and can that be switched off in writing?
- Are prompts, instructions and tool definitions stored in company-controlled files, or only inside a vendor's interface?
- Who may change an agent's instructions today, and is there a record of who changed them and why?
- What happens to the agent, its memory and its logs if the vendor relationship ends for any reason?
A clear answer usually takes the form of a short register: each agent, its owner, the account it runs on, and the location of its instructions. If the seller cannot produce that register within a few days, the delay is itself a finding. For the governance side of this, the permission review process is a useful reference for who signs off on access.
Can the company memory actually be exported?
The word memory covers several different things: raw conversation history, a set of curated knowledge documents, a retrieval index built from those documents, and structured records the agent writes to a database. Each has different portability. A retrieval index is built with specific embedding methods and usually cannot be read by another system without rebuilding it from source. Retrieval-augmented architectures separate the knowledge store, the retrieval step and the model, so the buyer should ask which of those parts it receives and in what form.
Ask for the source material, not only the index. The source documents, the update procedure and the schedule for refreshing them are what allow a new team to rebuild the layer. Then ask for a sample export, opened by someone outside the seller's team, and check whether it contains what the agent actually uses or only what was uploaded at the start. Corrections made during daily use often hold the most valuable knowledge and are the easiest part to miss. Testing a rebuild is covered in more detail in the guide to AI agent backup and restore.
Which prompts and tools would a buyer inherit?
Prompts are operating procedures written in natural language. They decide tone, escalation rules and which cases go to a person. Treat them like contract templates: versioned, reviewed and tied to the people accountable for the outcome. Request the version history, the last few substantive changes with their stated reasons, and any test cases the company used to check the agent's behavior.
Tools are the actions an agent can take: sending a message, creating an order, updating a record or issuing a refund. For each tool, ask which system it touches, which account it uses, and whether a person must approve actions above a set threshold. Also ask whether the agent reads content from outside the company, such as customer emails or web pages, because instructions hidden in that content can try to redirect the agent. The OWASP Top 10 for LLM Applications describes this class of risk, and a technical adviser should be able to walk the buyer through how each tool is protected against it.
Which permissions and credentials break on closing day?
Many operational failures during an ownership change come from credentials, not data. An agent may run on the access of an employee who leaves after the deal. A service may depend on a personal OAuth grant tied to an individual mailbox. A scheduled job may fail when a token expires on a date no one wrote down. Ask for a permission inventory before signing that covers:
- Every account, service identity and OAuth grant the agents use, with its owner and its scope.
- Which credentials are tied to a named individual, and what replaces each one at closing.
- The approval thresholds for each agent action, and who set them.
- The date and result of the most recent permission review, with the evidence kept.
- Every external system the agents write to, not only the ones they read from.
The guiding principle is that each agent should hold only the access its role requires, which is the logic behind standard authorization guidance from OWASP. A buyer that finds broad, shared administrator access across several agents should require a remediation plan as a condition of closing, not as a later improvement.
A step-by-step plan, with an illustrative scenario
Illustrative scenario: a regional maintenance company, used here only as an example, runs an intake agent that answers service requests, a scheduling agent that books technicians, and a quoting assistant that reads its own price list. The seller describes the agents as part of the business but cannot say which accounts they run on. The buyer's operations lead works through the sequence below before the deal committee meets.
- Step 1: Build the inventory. List every agent, channel, tool, data source and credential, with a named owner for each line.
- Step 2: Confirm ownership in writing. Match each account and configuration to the legal entity that will own the company after closing.
- Step 3: Request a memory export and open it outside the seller's team. Confirm that the source documents exist and that the index can be rebuilt.
- Step 4: Pull the prompt history and tool definitions. Note which actions need approval, and whether those limits are enforced in the tool itself or only written into the prompt.
- Step 5: Run a restore test in a separate environment. Load the export and check whether the agent handles representative cases the same way.
- Step 6: Map credentials to people. Mark every personal account, plan its replacement, and schedule the change before closing day.
- Step 7: Agree on the first review. Assign an owner for the agent layer and set a date for the first permission review after closing.
In this example, the restore test is the step that usually changes the conversation. The scheduling agent may depend on a spreadsheet that exists only on one employee's laptop, which no one listed as a dependency. Finding that before closing takes a few days of work. Finding it afterward can mean a service backlog that nobody on the new owner's team knows how to clear.
Be Human handles this problem as part of the first engagement rather than as a later repair. The work defines each agent's role, the memory and tools it needs, its approval boundaries and an observable success criterion, and it records ownership, export paths and credentials in documents the company controls. Managed operations then keep that record current through monitoring, recurring evaluations, permission reviews and controlled changes, so a new owner inherits a layer it can read and run, not a configuration that only its former operator understands.
What evidence should you require before you sign?
Ask for evidence rather than descriptions. Useful evidence includes logs of recent agent actions, evaluation results with the cases that failed, incident notes showing how problems were resolved, and the most recent restore test if one exists. Observability guidance such as the OpenTelemetry concepts for logs and traces describes what a usable record of agent activity looks like. If the seller cannot show what the agents did last week, the buyer cannot judge whether they did the right things. The broader risk vocabulary in the NIST AI Risk Management Framework is a useful shared language for these reviews. For a fuller list of deal questions, see the AI due diligence checklist for AI stack buyers.
Close the diligence with three decisions written down: who owns the agent layer on day one, which credentials are replaced or re-issued before closing, and which agent actions stay under human approval until the buyer's team has its own evidence. Keep the list short enough to fit on one page. If you want to find the first task in your own operations worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Does buying a company that uses AI agents need different diligence from buying one that uses ordinary software? Yes, in one important way. Ordinary software is usually described by its license and its data, while an agent layer also carries instructions, accumulated memory and decision limits that do not appear in a standard data room. Those assets need their own inventory, ownership confirmation and testing before the deal is final.
Can a buyer simply ask the seller for a memory export? A buyer should ask, but a request alone proves little. The export should include the source documents used to build the knowledge, and someone outside the seller's team should open it and confirm that it can be read and loaded elsewhere. A sandbox restore is the clearest test of whether the memory is portable.
Who should lead AI agent diligence on the buyer side? The operations leader who will run the business after closing should own the questions, with a technical adviser who can check accounts, permissions and logs. Legal review is needed for vendor terms and data use rights. Ownership of the agent layer should be assigned to a named person before the transaction closes, not after.
What is the minimum evidence to request before signing? At minimum, request an inventory of agents, accounts and credentials, the prompt and tool history, a memory export with its source documents, and recent logs or evaluation results. Also request the most recent permission review and any incident records. If these cannot be produced quickly, the buyer should treat the gap as a risk to price into the decision.
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.


