Personalized AI infrastructure
Choosing an AI Implementation Partner: Ownership Questions to Ask First
Ask these ownership questions before hiring an AI implementation partner: data, prompts, memory, permissions and exit. Keep your gains as models improve.

A good AI implementation partner leaves your company owning the data, prompts, memory and operating rules behind its AI agents, so the gains survive a change of model or vendor. Before comparing proposals, ask who holds each asset, how you can export it in a usable form, and what happens to it when the contract ends. The questions below are the ones to settle first, before any demo or price discussion.
Key takeaways
- Settle who owns data, prompts, memory and configuration in writing before any build starts.
- Ask how every asset can be exported in a usable form, then test that export before you sign.
- Keep the company's knowledge layer separate from the model so a better model can be adopted without rebuilding everything.
- Require a defined exit: handover material, deletion of vendor-held copies and a transition period with access.
- Judge a partner by the evidence it leaves you (evaluations, logs, change history), not by the demonstration it runs.
Why ownership matters more as models improve
Models change quickly. The model that powers your agent today will probably be replaced by a stronger one, and the replacement should be a configuration decision, not a rebuild. What compounds inside a company is rarely the model itself. It is the context around the model: the documents the agent can cite, the instructions it follows, the record of past decisions, the rules about what it may touch and when a person must approve.
If that context sits inside a vendor's system with no export path, every model upgrade becomes a negotiation, and your operating knowledge becomes the vendor's leverage. Ownership questions are therefore not legal housekeeping. They decide whether the next improvement in AI benefits your business or only the provider.
Data: where it lives and who can copy it
Start with the data the agent reads and writes. Ask the partner to describe each source system, whether the agent queries it in place or receives a copy, and where any copy is stored. Then ask a short set of direct questions:
- Which systems will the agent read from, and does the partner copy that data or query it live?
- Where are copies stored, in which country, and who holds administrator access to them?
- Is any of your data used to train or tune models that the partner offers to other clients?
- How long are logs and conversation records kept, and how do you request deletion in writing?
- Who at your company receives the audit trail of what the agent accessed?
Brazilian companies should review these answers with counsel against the obligations under the LGPD, particularly where personal data about customers or staff is involved. A partner that answers in general terms, without naming systems and retention periods, has not yet shown you the answer.
Prompts and configuration are operating know-how
Prompts, tool descriptions, routing rules and escalation thresholds are how your business works, written in a form a machine can follow. They deserve the same care as a process manual. Ask the partner to keep them in a versioned repository that you can access, in plain text, with a readable history of who changed what and why.
Ask also who may approve a change. A sound arrangement lets the company approve changes to behavior that affects customers, money or personal data, while the partner handles technical implementation. Instructions are not the only risk here. Content the agent reads, such as an inbound email or a web page, can contain text written to override its instructions. The OWASP guidance on prompt injection describes controls such as separating untrusted content from instructions and limiting what the agent can do after reading it. Ask the partner which of these controls it applies and where they are documented.
Ask for the prompts in a file you can open without the partner's software.
Memory and retrieval: what the agent remembers and who can edit it
Memory has two forms that matter for ownership. The first is retrieval: the agent looks up relevant passages from a store of your documents at the moment it answers. The second is behavior learned or adjusted through tuning of the model itself. Retrieval keeps your knowledge in an index you can inspect and rebuild. Tuning can bake patterns into a model you do not control. The AWS comparison of retrieval and fine-tuning is a useful reference for this distinction, and it explains why many business agents start with retrieval.
For any memory store, ask where the index lives, which source documents it was built from, whether you can export both the documents and the index, and who can delete or correct an entry. Ask how a stale or wrong memory gets corrected and how that correction is recorded.
Illustrative scenario: a regional distributor wants an agent to answer questions about supplier terms. The partner builds a private index from the terms. At renewal, the distributor asks for the source files and the rebuild steps. It receives the files, but the rebuild steps depend on a configuration only the partner holds. The distributor cannot move the knowledge without the partner's help, which is exactly the situation the ownership questions are meant to prevent.
Permissions and tool access
An agent is useful when it can act: read a calendar, update a record, send a draft, open a ticket. Each of those actions needs a credential, and the owner of that credential determines who controls the agent. Ask whether tool accounts are created in your company's name, whether you hold the administrator role, and whether access is scoped to the minimum each task needs. Permissions should be written as a list you can read, with each approval threshold named.
Review them on a schedule. A practical method is described in how to run a permission review for your AI agents, step by step. If the partner cannot show you current permissions in one place, you cannot judge whether the agent still does only what you approved.
Be Human's managed approach starts from this list. Each engagement records the owner of every asset in a shared register, keeps source documents, prompts and configuration in formats the client can export, and places the agent's tool accounts under the client's administrative control. Changes to behavior or permissions go through a written request and approval, and the client receives the audit trail. These are operating choices the client can inspect, not promises about outcomes.
Build the exit before you sign
An exit plan is easiest to write before there is anything to leave. The steps below are a practical sequence to bring to any proposal:
- Step 1: List every asset the engagement will create: source documents, prompts, memory stores, tool configurations, evaluation sets, logs and runbooks.
- Step 2: Name the owner of each asset and the format in which it will be delivered.
- Step 3: Ask for a test export of a sample, then restore it into an environment your company controls.
- Step 4: Write the termination terms: the handover date, the list of materials, continued access during a transition period, and a written confirmation of deletion of vendor-held copies.
- Step 5: Repeat the test export on a regular schedule, and after any significant change to the agent.
The contract should reference the list, not summarize it. For the sections a contract should cover, the AI agent statement of work guide walks through each one. A partner that objects to a test export before signing is telling you how the exit will go.
Evidence over promises: what to ask a partner to show
A demonstration proves that something worked once. Ownership requires evidence that the work can be checked over time. Ask what the partner will measure, how often it will run evaluations against a fixed set of real tasks, what result triggers a review, and who sees the results. The NIST AI Risk Management Framework gives a useful vocabulary for governance, mapping, measurement and management of AI risk, and you can ask a partner to show how its practice maps to those functions.
Ask for logs and traces that let you reconstruct why the agent took an action, and confirm that those records are yours. Ask what happens when an evaluation fails. A serious partner will describe a rollback, a paused permission or a fix with a written change, not a vague assurance. Choose the partner whose evidence you could keep if you left, because that is the partner whose work you will still own after a model changes. If you want to find the first task worth handing to an agent and the ownership questions that task needs, take the free Be Human AI diagnostic.
Frequently asked questions
Should we pick a partner that works only with our preferred model? Not necessarily. Your ownership of data, prompts, memory and configuration matters more than the model brand, because models change faster than your operating knowledge. A partner that keeps context in a layer you can export lets you change models later without losing the work. Ask how a model swap is handled and what it costs in time and risk.
Is it enough for the partner to say our data is never used for training? A written statement helps, but it is not enough on its own. Ask where copies live, who holds administrator access, how long logs are kept and how deletion is requested and confirmed. Those answers show whether the statement is backed by the system design.
Do we need our own technical team to hold ownership? You need a named person who understands the register of assets, approves changes and reviews permissions, but that person does not need to write the code. Many companies assign this to an operations leader with a technical adviser. What matters is that the company holds the access and the records, not that it builds everything itself.
When should we run the exit test? Run a small export test before signing, then repeat it on a schedule and after any significant change to the agent. The first test often reveals missing steps, such as a configuration only the partner can rebuild. Finding that gap early is far less costly than finding it at renewal or termination.
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.


