AI Project Handover: Moving a Custom AI Workflow From a Consultant to Your Team
The consultant's last week is booked. The invoice workflow they built reads supplier PDFs, fills the accounts payable screen, and routes the odd ones to a clerk. It passed acceptance. Now someone on your team has to own it, and the consultant will not be there when it breaks.
Most of what you need for that handover is already well described. What the checklists mostly skip are two risks that are specific to AI workflows: the warnings that a model is being retired go to whoever owns the provider account, and every AI workflow arrives with change dates it cannot avoid. This guide covers the familiar ground quickly, then spends its time on those two gaps.
It assumes the build has already been accepted against agreed criteria and that the acceptance record names the model version and commit that passed. If you have not reached that point, settle acceptance first.
What the handover checklists already get right
Several guides cover the core of an AI handover, and they agree with each other. Sagnik Bhattacharya's checklist asks for owner-level control of every account, fresh API keys, the final prompts with the tests they passed, a one-page runbook per automation, and staff who have run a fix once. KUMO's exit checklist sets ten acceptance tests, from company-controlled repositories and billing to a rollback and an evaluation baseline that runs before and after the transfer.
The strongest point in that corpus is that documents are not proof. Kastling writes that a folder full of documents does not establish that the new operator can use them. Startup Engineering puts it more bluntly: documentation without a reader is a ritual. Marius Manolachi recommends accepting the transfer only after a fresh operator, using the packet and transferred access alone, can run a baseline, diagnose a seeded failure, apply a change, and roll it back.
Public buyers are told the same. The UK government's Guidelines for AI procurement say to make knowledge transfer and training part of the requirement and to plan against vendor lock-in.
Use those guides. What follows adds the parts they leave out.
Inventory accounts, keys, hosted objects, and notification contacts
Start with a list, not a meeting. For each item, record who owns it today, who will own it after handover, and where the evidence is.
| Item | What to record | Why it matters |
|---|---|---|
| Model provider accounts | Organization, projects, owners, billing contact, notification email | Controls usage, spend, and who hears about retirements |
| API keys and service accounts | Name, project, permissions, last use, where the value is stored | Tells you what must be recreated and what must be revoked |
| Pinned models | Exact model name or snapshot per workflow step | Each one has its own retirement schedule |
| Hosted provider objects | Files, vector stores, fine-tuned models, dashboard prompts, hosted evaluations, stored conversations | Some cannot be moved and some have their own shutdown dates |
| Connected systems | Email, ERP, document store, chat tools, and the account each connection was authorized with | A connection authorized by the consultant leaves with the consultant |
| Source and data | Repository, prompts, acceptance set, configuration, deployment scripts | Lets someone else rebuild and redeploy |
Two parts of that table are easy to miss. The first is the notification email, which most checklists never ask for. The second is the hosted objects column. A prompt that lives in a provider dashboard, or an evaluation suite that lives in a provider's hosted tool, is part of the system even though it is not in the repository.
Providers can help you build the list. Anthropic's model deprecations page describes exporting usage from the Claude Console as a CSV broken down by API key and model. A usage export is a good check on the consultant's inventory, because it shows what is actually being called rather than what the documentation says.
Move the accounts first, then rotate the keys
The order matters. If you rotate keys while the account still belongs to the consultant, you have rotated keys you still do not control.
The cleanest arrangement is for the client to open the provider accounts before the build starts and give the consultant named access. If that did not happen, you cannot simply take over the consultant's keys. OpenAI's Services Agreement prohibits buying, selling, or transferring API keys from, to, or with a third party, and the OpenAI Help Center notes that service accounts are scoped to the project they were created in. In practice that means creating keys in a project your organization owns and pointing the workflow at them.
Then rotate. Axonbuild's takeover checklist rotates every credential the previous builder held or could read, after the accounts have moved, and checks request logs before expiring the old key so nothing breaks unseen. Our guide to rotating credentials in long-running AI workflows covers the mechanics, and keeping API keys out of AI agent context covers where the new keys should live.
Rebuild provider-side objects from your repository
Moving an account does not always move what was built inside it. The OpenAI Help Center says project resources such as files, assistants, storage, and threads are scoped to the project and cannot be moved between projects, with one exception for fine-tuned models in the default project. It also says only project members and organization owners can see the fine-tuned models and files created in a project.
So if the consultant built the workflow in their own organization, your new project starts empty. The only dependable transfer is a script in your repository that recreates each object in your project: uploads the reference files, rebuilds the vector store from the source documents, and recreates any configuration. Run that script during handover, not after the consultant has gone.
Some objects are harder to recreate than they look. OpenAI's deprecations page records that since May 7, 2026, creating fine-tuning jobs is not available to organizations that have not previously run fine-tuning, and that active existing customers lose the ability to create new fine-tuning jobs on January 6, 2027. A fine-tuned model that lives in the consultant's organization may not be reproducible in yours. If the workflow depends on one, decide at handover whether to keep it where it is under a written agreement, replace it with a prompted base model that passes the acceptance set, or accept the risk with an owner and a date.
Stored conversation history is the other trap. Before OpenAI sunset the Assistants API on August 26, 2026, an ondelva post warned that thread contents would be unreachable through the API once the endpoints stopped answering, and that no tool would get them out afterward. If your workflow keeps history on the provider's side, export it to storage you own as part of the handover.
Route retirement notices to your own people
This is the gap most handover guides miss. Model providers warn you before a model goes away, but they warn the account.
OpenAI's deprecations page says impacted customers are notified by email and in the documentation, and that when a deprecation is announced it notifies customers who are actively using the model by email. Its minimum notice periods are at least six months for generally available models and at least three months for specialized variants, while preview models may be retired with much shorter notice, such as two weeks. Anthropic's deprecations page says it notifies customers with active deployments and gives at least 60 days' notice before retiring a publicly released model.
The email is not a courtesy copy. OpenAI's Services Agreement says notices to the customer may be sent to the applicable account email address and are deemed given when sent. Material reductions in functionality are also announced to that address, and the customer then has five business days to choose whether to terminate.
The email can also arrive before the public page changes. In an OpenAI Developer Community thread dated October 1, 2026, developers reported receiving a shutdown email for text-to-speech models that the deprecations page did not yet list. That is one group's report, and the page has since been updated, but it shows why the inbox matters.
Put those facts together and the risk is simple to state. If the provider account, or the email address on it, still belongs to the consultant, the formal warning that your workflow will stop goes to someone who no longer works for you. The opened sources do not specify exactly which members of a provider organization receive these emails, so treat this as a recommendation built from them rather than a documented rule: move the owner and notification email of every model provider account to your own people before the consultant leaves, because providers send model retirement and service-change notices to the account by email, and OpenAI's terms treat a notice sent to the account email address as given when sent.
What to do in practice:
- Make two people on your side owners of each provider organization, so a holiday or a resignation does not silence it.
- Use a shared team address for the account and billing contacts rather than one person's inbox.
- Ask the consultant to confirm, in writing, which addresses currently receive provider notices for each account, and change them while they are still engaged.
- If the workflow runs on a partner platform, check that platform's schedule too. Anthropic notes that Amazon Bedrock and Google Cloud set their own retirement dates, which can differ from the Claude API.
Put the next forced change on the handover calendar
A handed-over AI workflow is not finished software. It comes with dates on which something will change whether or not anyone touches it. Those dates are knowable on handover day, so record them then.
The record is short. Write the next forced change into the handover: the retirement date of every pinned model and the shutdown date of every hosted provider feature the workflow uses, each with a named owner and a planned rerun of the acceptance set on the replacement.
The provider pages show how close these dates can be. As listed on November 16, 2026:
| Dependency | Date | Source |
|---|---|---|
| OpenAI Evals platform | Read-only; shut down November 30, 2026 | OpenAI deprecations page |
OpenAI reusable prompt objects and the v1/prompts API | Scheduled shutdown November 30, 2026 | OpenAI deprecations page |
claude-sonnet-4-5-20250929 | Retirement November 30, 2026 | Anthropic deprecations page |
gpt-5-2025-08-07 and o3-2025-04-16 | Shutdown December 11, 2026 | OpenAI deprecations page |
claude-haiku-4-5-20251001 | Active, retirement not sooner than October 15, 2026 | Anthropic deprecations page |
Hosted features deserve the same attention as models. OpenAI's Assistants migration guide told teams leaving the Assistants API that its replacement, prompts, could only be created in the dashboard. The deprecations page now schedules those reusable prompt objects to shut down on November 30, 2026. A workflow that followed the official migration path a few months ago may now depend on a second hosted object with its own end date. Sagnik Bhattacharya flags the same pattern for custom GPTs, which stop working on 11 December 2026.
For each row on your own calendar, name the owner, the date by which a replacement will be tested, and the acceptance set it will be tested against. The acceptance set and recorded baseline from the build are what make that rerun cheap. When the date arrives, our model deprecation migration runbook covers the engineering work. The handover's job is only to make sure someone on your team knows the date exists.
A provider's "not sooner than" date is a floor, not a promise, and a deprecation can be announced at any time. That is why the calendar needs a monthly check of the provider pages by the named owner, in addition to the notification routing above.
Prove it with a break test, then sign off
With the accounts moved, the objects rebuilt, notices routed, and the calendar written, run the test the guides recommend. Your operator, not the consultant, uses only the runbook and the access you now own to:
- Run the acceptance set and get a result close to the recorded baseline.
- Diagnose a failure the consultant has seeded in a test environment.
- Make one approved change, deploy it through the real release path, and roll it back.
- Pause the workflow and resume it without losing or duplicating work.
Tie the final payment to that test. Sagnik Bhattacharya suggests one contract sentence that does most of the work: handover is accepted when the client's nominated staff member has completed the break test using the runbooks and every item on the attached list is marked done. Add the notification check and the dated calendar to that list.
A worked example: supplier invoice extraction
The numbers here are illustrative. A finance team of four takes over an invoice extraction workflow built by an outside consultant during a fixed-price pilot.
The inventory finds one OpenAI organization owned by the consultant, with billing passed through on their invoice, two API keys, one vector store of supplier master data, a dashboard prompt for the extraction step, and an evaluation suite in OpenAI's hosted Evals tool. The finance team's systems administrator creates an organization for the company with two owners and a shared finance-systems address as the contact. The consultant writes a rebuild script into the repository that recreates the vector store from the supplier master export in the new project, and moves the prompt text and the acceptance set into the repository too. The team runs the acceptance set from the repository with an open-source runner, so it no longer depends on the hosted Evals tool scheduled to shut down on November 30.
New keys go into the company's secret manager. The workflow switches over, and the old keys are expired once their request logs show no traffic.
The calendar gets three rows: the old dashboard prompt, no longer called once its text moved into the repository but kept on the list until someone deletes it; the pinned extraction model's shutdown date from the deprecations page, owned by the systems administrator, with a replacement test booked a month earlier; and a first-Monday check of both providers' pages. The accounts payable lead runs the break test. Only then is the last invoice paid.
What this guide does not cover
Choosing between a consultant and an in-house build is covered in AI consultant vs in-house. Exit cost as part of the original build decision is in build vs. buy custom AI software. Who owns the code when we build an internal tool is answered on our internal tools page.
Next action
Open your provider consoles today and write down, for every account the workflow uses, who the owners are and which email address receives notices. If any of those belong to someone outside your company, that is the first item on your handover list. Then look up the retirement status of each pinned model on the provider's deprecations page and put the nearest date on a calendar with a name next to it.