Personalized AI infrastructure
AI Agent Handover Deliverables: What Your Company Should Own in Writing
A written AI agent handover should include prompts, a memory export, a permission map, runbooks and evaluation records you can run without the provider.

When an outside team hands an AI agent back to your company, the handover should be a set of written files that your staff can read, run and audit without that team. At minimum, you should receive the agent's instructions, a memory export with its sources, a permission map that lists every tool and data scope, runbooks for operating and stopping the agent, and the evaluation records that show whether it works. If any of these exist only in the provider's systems or in one person's head, the engagement has not ended. It has only changed where the dependency sits.
Key takeaways
- A handover is complete only when your company can operate, pause and move the agent without the original builder.
- Prompts, instructions and tool descriptions belong to you as versioned text, not as chat history.
- A memory export should show what the agent knows, where each item came from and who can correct it.
- A permission map lists every tool, account and data scope, with the approval boundary attached to each one.
- Runbooks and evaluation records turn tacit knowledge into checks that your staff can repeat.
Why the handover is a set of files, not a meeting
Many engagements close with a walkthrough call and a recording. A recording tells you what was discussed, but it does not tell you what the agent is configured to do next Tuesday. Written deliverables can be tested. A member of your team can open them, compare them with the live system and notice when the two disagree.
A useful test is to imagine that your operations manager leaves next week. Could a capable colleague keep the agent running, pause it safely and explain its decisions using only the files you hold? If the answer depends on a phone call to the builder, the handover is incomplete. Request the files before the engagement is formally accepted, because your leverage to demand them shrinks once the work is closed.
Prompts and instructions you should own
The instructions file is the agent's job description. It states what the agent is for, what it must never do, when it must hand a case to a person and what tone it uses on each channel. Treat it as a controlled document. It should be plain text or a standard document format, stored in a location your company controls, with version history and a named owner inside your organization.
A complete instructions package usually contains: - The role statement, written as the work the agent does and the work it does not do. - Escalation rules, naming the person or queue that receives each type of exception and the expected response time. - Language and tone rules for each channel, including Portuguese or English where both are used. - Tool descriptions, including the cases in which a tool should not be called. - A version number, the date of the last change and the name of the person who approved it.
Do not accept instructions that exist only inside a vendor's dashboard or a shared chat thread. Those can be edited without a record, and they can disappear when the account closes. Also check whether the agent treats text from emails, documents or web pages as instructions. Untrusted content can try to override an agent's rules, so the boundaries between instructions and data should be written down and enforced by the system, not left to the prompt alone.
Memory exports and their sources
Memory is what the agent can recall when it answers: customer notes, product rules, policy documents, past decisions and open tasks. In most business agents, memory is not one database. It is a combination of stored facts and documents that the agent retrieves at answer time, a pattern usually called retrieval-augmented generation, which AWS describes in plain terms. That structure matters for handover, because the export must include both the content and the map of where each piece came from.
A memory export that is only a raw dump is hard to audit. Ask for: - The source documents, or stable pointers to them, in a format your company can read without special software. - The index or memory store in an exportable form, with a short document describing its structure. - A record of what was added, changed or deleted, with dates and the person or process responsible. - A rule for expiring or correcting outdated information. - Flags on any items that contain personal data, so that your privacy obligations travel with the file.
Ask the provider to demonstrate the export by choosing five items at random and tracing each one back to its source. If a trace fails, the memory is not yet owned by your company, even if it has been technically delivered.
Permission map and tool inventory
The permission map answers a question every operations leader eventually asks: what can this agent touch, and who approved it? Each tool, account and data scope the agent uses should appear in the map with the identity it runs under and the actions it may take. Standard authorization guidance favors denying access by default and granting only what each task needs. The map is where that principle becomes something you can check.
For every entry, record: - The system, account or API the agent reaches, and the identity it uses to connect. - The actions allowed, separated into read, draft and write. - The approval boundary, meaning which actions need a named person to confirm them before they run. - The owner who can grant, change or revoke the access. - The date of the last review and its outcome.
Compare the map with the live account settings before you accept it. A write permission added for a test and never removed is one of the most common findings in a first review, and it is easy to miss when the document and the system are maintained separately. For the recurring process, see how to run a permission review for your AI agents.
Runbooks, evaluations and backups
A runbook is the document a person follows when something goes wrong or when routine action is needed. It should explain how to start and stop the agent, how to switch it into a reduced mode where it only drafts responses, who to contact during an incident, how to roll back a prompt or permission change and where the logs are kept. Write it for a competent operator who has never seen the agent, because that is likely the person who will read it.
Logs make the runbook usable. The concepts of traces, metrics and logs, set out in OpenTelemetry's observability primer, apply to agents as much as to any other software service. Ask that each agent action be recorded with a timestamp, the tool called, the approval outcome and the version of the instructions in force. Without that trail, nobody can explain a poor decision after it happens.
Evaluation records show whether the agent meets its success criterion. Hand over the test cases, the expected outcomes, the results of each run and the version tested on each date. A small, fixed set of cases that your team understands is more valuable than a large benchmark that no one can read.
Finally, ask for a backup of the instructions, memory and configuration, together with a recorded restore into a test environment. A backup that has never been restored is only an assumption. The AI agent backup and restore guidance covers what to save and how to test it.
Illustrative scenario and a handover checklist
Illustrative scenario: a regional distributor hands over an agent that answers trade account questions for its sales staff. At closing, the team receives the instructions file, a memory export that links each policy summary to its source document, a permission map showing read-only access to the order system and draft-only access to the outgoing email queue, and a runbook that explains how to switch the agent into drafting mode. Later, the account manager who understood the agent leaves. Because the files are complete, the operations lead pauses the agent, checks the permission map, finds one stale write scope left from a test and removes it the same day.
Before you accept any handover, confirm the following: - Every file has a named owner inside your company. - The instructions are stored in a location you control, with version history. - A sample of memory items traces back to a source document. - The permission map matches the live account settings. - The evaluation set is saved, with results for the current version. - A restore has been performed at least once in a test environment. - Someone who did not build the agent has read the runbook and found it sufficient.
How Be Human keeps the handover current after launch
Be Human designs, implements and operates personalized AI infrastructure, and the handover is part of that work rather than a closing document. In managed operations, prompt edits, permission changes and memory updates are recorded as requests that state the reason, the approver and the expected effect, as described in how to request and approve changes to a managed AI agent. Monitoring and recurring evaluations show whether the agent still meets its success criterion, and permission reviews confirm that the map still describes reality. The company keeps ownership of the files throughout, so the handover remains a current state rather than a single delivery. If you are unsure which task should be handed to an agent first, the free Be Human AI diagnostic can help you find it.
Frequently asked questions
Do we need the provider's source code to own the agent? Not always, but you need every element that shapes the agent's behavior in a readable form. That means the instructions, tool descriptions, memory export, permission map and configuration settings. Source code matters most if you may move the agent to another platform, so raise the question while the platform choice is still open.
Who should hold the administrator accounts after handover? Named people inside your company should hold administrator access, and at least two of them should be able to act in an emergency. Any access kept by the provider should be time-limited, logged and removed when the scope of the engagement changes. Review these accounts as part of the permission map, not as a separate task.
Can we ask for memory in a format our own tools can read? Yes, and you should specify that format before the work begins. Ask for open, documented formats such as plain text, spreadsheets or standard document files, along with a short description of how the records relate to each other. An export that only the original platform can read is a sign of dependence.
How often should the handover files be updated? Update them whenever the instructions, permissions or memory change, and review the complete set on a fixed schedule. The files are useful only while they match the live agent, so a stale document should be treated as a fault to correct, not a minor detail. Assign that review to a named person so it does not depend on memory.
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.


