One delivery plan for one priority deployment.
OpenAI’s posting for Applied AI Delivery Lead asks for an “integrated delivery plan”: scope, milestones, accountable owners, dependencies, resource needs and success measures across customer, OpenAI and partner teams. This is one, for a made-up customer. I built it in one day from OpenAI’s public pages, with AI agents doing the reading.
One deployment, three teams
The deployment is a claims assistant at Fernhill Mutual, a made-up insurer. An agent built on the API drafts claim summaries and next steps inside the adjusters’ claims system, and the claims organization gets ChatGPT Enterprise. A delivery partner does the integration.
Customer
Fernhill Mutual, a made-up insurer.
OpenAI
Role names as they appear on OpenAI’s public pages.
Partner
A delivery partner from the Partner Network.
Fernhill Mutual and its people are made up. The role names in the OpenAI column are ones OpenAI’s public pages use. Which role owns which workstream is my proposal: I cannot see from outside how OpenAI assigns them. Sources are on the evidence page.
The sixteen weeks
Five phases that overlap. Sixteen weeks is my assumption, to make the plan concrete. Scoping sets the real length. OpenAI’s forward-deployed team has written of one first deployment that took two months.
Scope, owners, baseline
OpenAI’s own sequence for managed deployments starts with outcomes and success criteria. So does this.
- Interview every workstream lead before the plan is written.It is how I start a program. The answers become one script that all three teams plan from.CustomerOpenAIPartner
- Agree the outcome and how it will be measured.With the sponsor and the account team: the result to improve, the measure, and who reports it.CustomerOpenAI
- Set the baseline.From the customer’s own claims reports, and a first survey started by their admin. OpenAI’s scheduled surveys begin at day 120, which is too late to be a baseline.Customer
- Name one owner for every workstream.Seventeen workstreams, each with one name beside it. A workstream with two owners has none.CustomerOpenAIPartner
How I will know. Every workstream has a named owner and a first milestone, and the sponsor has signed the measures.
Build and review
The build and the reviews run side by side, because the review is often the longer of the two.
- Security, privacy and legal review.The reviewer gets OpenAI’s Trust Portal documents and the partner’s data-flow diagram in week 3, not week 7.CustomerPartner
- Identity before invitations.Domain, single sign-on and directory sync first. That is the order OpenAI’s quickstart gives.CustomerOpenAI
- The agent, with evals written with adjusters.Forward Deployed Engineers build it. The pass mark is agreed before the first test run.OpenAICustomer
- Integration with the claims system.The partner connects the agent. The customer supplies a test environment and a change window.PartnerCustomer
How I will know. The evals reach the agreed mark on real example claims, and the security review has no open blocker.
Launch readiness
A go or no-go is only honest if every gate shows its evidence. A gate with nothing behind it shows a dash, never green.
- Check the model against the launch date.OpenAI publishes retirement dates and gives generally available models at least six months’ notice. The plan records the date beside the launch date.OpenAI
- Size capacity from testing, then order it.Reserved capacity is bought per model through sales. The load estimate comes from acceptance testing, not from a guess.OpenAIPartner
- Acceptance testing with the pilot adjusters.On real claims, in the test environment, with the people who will use it.PartnerCustomer
- Training, floor support and the support path.Who trains, who attends, and who a stuck adjuster calls in week one.CustomerPartnerOpenAI
- The go or no-go review.Eight gates, each with its evidence, and each owner in the room.CustomerOpenAIPartner
How I will know. Every gate has evidence or a stated reason it does not, and the sponsor decides in one meeting.
Launch
A staged rollout: a pilot group first, then the rest of claims. OpenAI describes the same staging for its managed deployments.
- The pilot group goes live.With a daily check in the first week: errors, hand-offs to a person, and what adjusters rewrote.CustomerOpenAIPartner
- The wider rollout.Only after the pilot’s first-week review.CustomerPartner
- A launch report to both sponsors.What shipped, what did not, and what is still open.OpenAI
How I will know. Adjusters in the pilot use it on live claims, and no gate was waved through to get there.
First adoption, and the playbook
Launch starts the measurement. It does not end the plan.
- Adoption against the baseline.Usage from the analytics API, the outcome from the claims system, and what people report. Each number carries its source and its date. OpenAI says workspace analytics can lag by up to 48 hours, so the review says so too.OpenAICustomer
- What the result means for usage.The account team hears whether the ramp is on its planned date, and why, at the review and not at quarter end.OpenAI
- The playbook page.What I would reuse, what I would not, and feedback for product, written so the next delivery lead can start from it.OpenAI
- Close-out with both sponsors, and handover.The partner hands the run-book to the customer’s platform owner.CustomerOpenAIPartner
How I will know. The sponsor can say what changed against the week-one baseline, and the next deployment starts from this one’s plan.
The delivery review I would run
Every two weeks with the customer, every week inside. One page, in the same order each time.
The posting asks for reviews that “drive decisions and action”. So decisions come first, while the people who can make them are in the room.
- 1. Decisions needed today.Each with the options, the owner, and what waits on it.
- 2. Gates, with their evidence.Passed, at risk, late, not due, or no data. Never green over a blank.
- 3. Risks that moved.Only the ones that changed, each with one owner and a date.
- 4. Launch confidence, and what it rests on.Derived from the gates by a stated rule, so anyone can see why.
- 5. Adoption against the baseline.After launch. Before launch this part is empty, and says so.
- 6. What a moved date means for usage.If launch slips, the ramp slips with it. The account team hears it here first.
What I would cut
- A second use case.The posting says priority initiatives. One deployment taken to adoption, before a second starts.
- A weekly status deck.The board is the status. A deck is a second copy that goes stale by Friday.
- Reports built on dashboard exports.OpenAI’s own guide says not to build a lasting report on dashboard labels. The board reads the analytics API and the customer’s system.
- Fine-tuning.OpenAI’s self-serve fine-tuning stops taking new jobs on January 6, 2027. Prompts, tools and evals first.
What would change this plan
- If OpenAI already has a plan format,I use it. This one shows how I think, not what I would impose.
- If the deployment is a managed voice or chat agent,it follows OpenAI’s published sequence for those, and the partner’s share changes.
- If the customer already runs ChatGPT Enterprise,the identity and workspace work drops out and the plan is shorter.
- If the account team’s forecast hangs on one date,that date becomes a gate with an owner.
- If a workstream lead tells me something this page missed,it goes in. That is what the week-one interviews are for.
How I would measure it
Four measures, agreed in week one and reported at every review.
I can see no baseline for any of them from outside, and I have made none up. Setting them is week 1 work.
- The outcome the sponsor chose.For this customer, time from first notice of a claim to first contact. From their system, not ours.
- Share of the pilot group using it each week.Adoption, not seats handed out.
- Gates passed on their planned date.The delivery measure. A launch that arrives by waving gates through has not arrived.
- Days from launch to the first review with real numbers.How soon the customer can see value for themselves.