Personalized AI infrastructure
AI Due Diligence Checklist: What a Buyer Inherits From an AI Stack
A practical AI due diligence checklist for buyers. Inspect agents, memory, prompts and permissions so AI gains stay with the business after closing.

AI due diligence means inspecting every part of a company's AI stack that could stop working or change hands at closing: the agents doing the work, the memory they rely on, the prompts that shape their behavior, the tools they can call, and the permissions that govern all of it. A buyer should confirm who controls each piece, whether the business can export and operate it without the seller, and what evidence shows it works. The checklist below turns those questions into items a deal team can answer before signing.
Key takeaways
- Inventory every agent, memory store, prompt, tool connection and permission grant before valuation discussions close.
- Confirm that the company, not the seller or a personal login, holds administrative control of accounts, data and configuration.
- Treat memory and prompts as company assets with owners, version history and an export path, the same way you treat contracts.
- Test whether your own operators could run the stack from written runbooks and a recent permission review.
- Write approval boundaries and success criteria into the handover pack so the capability keeps working after the builders leave.
Why AI gains are easy to lose at closing
Buyers usually price AI capability by its outputs: faster replies to customers, quicker quotes, cleaner monthly reporting. Those outputs depend on parts that seldom appear in a data room. An agent may run from an account opened under a founder's personal email address. Its memory may be a document set curated by one operations manager. Its instructions may have been tuned through dozens of informal edits that nobody recorded. A single API key may be shared by three tools. If any of these sits outside the company's control, the capability can stop at the first change of ownership, even though the software still runs.
The diligence question is therefore not whether the AI works today. It is whether the new owner can operate it, change it, audit it and rebuild it without the people who built it. That question applies to a small business acquisition as much as to a large one, and it is often easier to answer early than to repair after the deal closes.
What belongs in the inventory
Start with a flat list of everything that acts, remembers or decides on the company's behalf. Most companies find more than they expected, including tools adopted by a single team and never reviewed by anyone else. For each item, record what it does, who owns it, where its memory and configuration live, which systems it can touch, and what a person must approve before it acts. A simple map of how these pieces connect is often the fastest way to see gaps, and a one-page AI stack architecture map can serve as the template.
The inventory usually covers five areas. Agents are the roles that perform work, such as answering support messages or preparing supplier summaries. Memory is what they retain between sessions, including documents, past decisions and customer notes. Prompts and instructions define how they behave and what they must refuse to do. Tools are the connections they use to act, such as a customer database, a shared inbox, a billing dashboard or a file store. Permissions decide which tools each agent may use, at what level of access, and under whose authority.
Checklist: questions for every agent and its memory
Use these questions with the seller's operations lead and, where possible, with the person who maintains the system day to day. Ask for written answers, and ask for a live demonstration wherever the answer depends on configuration rather than on someone's recollection.
- Which company or account owns the platform subscription, and can administrative control transfer on a named date?
- Are all accounts tied to a company email domain and to named roles, rather than to one employee's personal login?
- Can the memory be exported in a readable format, with dates and sources attached to each entry?
- Does the memory contain personal data, customer records or contract terms, and is there a written note of who may read it?
- Has someone tested a restore of memory and configuration into a separate environment? A backup that has never been restored is an assumption, and the guide to AI agent backup and restore explains what to save and how to check it.
- Which decisions does the agent make alone, and which ones wait for a human approval?
A backup that nobody has restored is a hope, not an asset.
Checklist: prompts, tools and permissions
Prompts and tool access are where much of the hidden value and much of the hidden risk sit. A prompt that quietly grants an agent broad reach is a permission in disguise. A tool connection with write access is a capability the buyer must understand before taking ownership of it.
- Are prompts and instructions stored in a versioned document or managed system with a change history, rather than pasted into a chat thread?
- Can each prompt be traced to an owner, a stated purpose and a last review date?
- Does every tool connection use its own credential, scoped to the least access the task needs, instead of an administrator key shared across tools?
- Are actions that send messages, issue refunds, change records or spend money gated by approval rules that are written down?
- Are customer emails, web pages and uploaded files treated as untrusted input, so that text hidden inside them cannot override the agent's rules? The OWASP guidance on prompt injection describes this class of risk and the basic defenses.
- Is there a log of agent actions that a new operator can read without the original builder present?
The permission questions deserve their own review cycle after closing, and the step-by-step permission review guide shows how to collect the evidence. Authorization should follow the principle that each identity gets only the access it needs, which the OWASP authorization guidance states in more detail.
Illustrative scenario: a support agent after a sale
Illustrative scenario, not a real case: a software distributor acquires a smaller competitor whose support agent answers most first-line tickets. During diligence, the buyer's team asks for the agent's memory and learns that it lives in a workspace owned by the seller's former support lead, who still holds the only administrator login. The prompts exist in a long chat thread. The refund connection uses a key with no spending limit and no approval step. The agent works well, but the buyer does not yet own it.
Before closing, the parties agree on four changes. Administrative access transfers on a named date. The memory is exported and reviewed for customer data that should not be retained. The refund key is replaced with one scoped to a limit set by the finance lead, with a human approval above it. The prompts move into a versioned document with an owner. The purchase price discussion does not change, but the handover pack now contains the items that keep the gain in the business.
How a managed approach handles the handover
Be Human designs, implements and operates AI agents so that identity, memory, channels, tools, permissions and operating context remain under the customer's control. The first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, and it records each of these in a form the company can read without specialist help. Managed operations then keep that record current through monitoring, recurring evaluations, permission review and controlled changes. The handover therefore does not rest on one person's memory of how the system was built, and the company can see what changed, when and why.
A related risk is dependence on a single provider. Before renewing or transferring any agent platform, check whether the memory, prompts and logs can leave the platform in usable form. The early warning signs are covered in the guide to vendor lock-in for AI agents.
Turning diligence findings into deal terms
Once the inventory and checklists are complete, sort each finding into one of three groups before the deal documents are drafted. Findings in each group need a different kind of commitment, and mixing them is how important items slip through.
- Transfer at closing: administrative accounts, company-domain credentials, a current memory export, and the current prompt set with its owner named.
- Fix before closing: shared keys, personal logins, write permissions nobody reviewed, and any stored data the company has no clear right to keep.
- Accept with a plan: components that work but lack documentation, each with a named owner and a dated commitment to write the runbook.
The NIST AI Risk Management Framework gives a neutral vocabulary for recording these risks, which helps both sides agree on what was found. If you are not yet sure which of your own tasks is the right first one to hand to an agent, take the free Be Human AI diagnostic to find it.
Frequently asked questions
Who should lead AI due diligence on the buyer side? The lead is usually an operations or technology leader who can read the inventory and question the seller's team directly, with finance and legal involved for contract terms and data questions. A single owner keeps the findings in one list, which prevents items from being dropped between workstreams. The executive sponsor should approve the transfer and fix decisions.
Is AI due diligence different from ordinary software due diligence? It covers similar ground, including accounts, contracts and data, but it adds memory, prompts and agent permissions, which behave less like static code and more like working knowledge. Those components change with use, so a snapshot taken during diligence can go out of date quickly. The buyer should ask for the current state and a plan for how it will be maintained.
What if the seller will not share prompts or memory? Treat the refusal as a finding, not as a closed question. Ask what the seller will provide in an exported or readable form, and whether a supervised walkthrough is possible. If the answer remains no, the capability should be valued as something the buyer must rebuild, and the cost of that rebuild belongs in the deal discussion.
Does a buyer need to review prompts if the software vendor manages them? Yes, because a vendor managing a prompt does not mean the company owns its wording, history or purpose. The company must know what the agent is instructed to do, who approved those instructions and how they can be changed or moved. Vendor management is a service arrangement, while ownership is a property of the business.
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.


