# A project can close green and still die once the FDE leaves

> Handing a system to support should be an acceptance gate, with written criteria, a named owner, a fixed hypercare window and a formal sign-off date.

Original: https://fdetimes.net/en/guides/fde-handover-to-support-acceptance-gate/

Three weeks after a project closes, the customer success lead sends a message: the client says the agent "keeps getting things wrong lately", and nobody on the team knows when the errors started, where they occur or who should fix them. Take a hypothetical case. On a Friday afternoon you closed out an agent that reads order emails for a logistics company. The demo went smoothly and the dashboard was green from end to end.

Rocketlane, which makes software for customer implementations, describes this pattern precisely: successful go-lives that quietly turn into failed deployments. Its test is an uncomfortable one. If user behavior has not changed, the implementation has failed, even when the project closed green.

For an FDE, the internal handoff to support and customer success is the point where results are most easily abandoned, because it is the moment the person who understands the system best walks away.

## Handoff is an acceptance gate

PostHog's FDE handbook draws the roles clearly. Once the technical work is done, the FDE hands the client relationship back to the account owner, along with a record of what was built. From then on, CS and the TAM (technical account manager) own the long-term relationship and track whether the client is getting value from the product.

That split only works if the receiving side is able to take the load. Educative's FDE course calls the handoff the FDE's final formal act, and warns that a system delivered without documentation, monitoring and trained people to own it creates dependency rather than a deliverable.

The course is blunt on one point: a receiving team cannot operate what it does not understand.

Google solved a similar problem for its SRE teams long ago. In the SRE book, a Production Readiness Review (PRR) is a prerequisite before an SRE team takes operational responsibility for a service.

Bring the same spirit to the handoff to support. Support does not simply get handed a system; it accepts the system after inspecting it.

## The dossier must answer the 2 a.m. question

PostHog has a rule that is easy to apply. Anything the client's team must do on its own, with no FDE in the room, has to live in a durable document. A Slack thread, a recorded call or your own memory does not count.

Go back to the order-email agent. Below is a minimal handoff dossier, and each item has to answer a question support will actually face:

| Item | Question it answers | Example for the order-email agent |
|---|---|---|
| What the system does and does not do | The client asks, "Why didn't it handle this email?" | Reads only English emails from three inboxes; skips image attachments |
| Data flow and dependencies | When something breaks, where do I look? | Inbox → queue → LLM → client's ticket-creation API |
| Signs of degradation | How do I know it is getting worse? | Share of emails landing in "unclassifiable", number of tickets corrected by hand |
| Runbook | What do I do about common failures? | Ticket API returns 401: token expired, who reissues it, where |
| Running costs | Will peak season blow the budget? | Cost per email, the ceiling agreed with the client |
| Owners | Who decides when something goes wrong? | Name of the client-side owner, the CSM, the on-call support engineer |

Cost is the item most often left for later. Educative puts it among the readiness criteria to settle before handoff, arguing that sustainable cost is a readiness condition and should not wait until after launch. If support does not know the ceiling, the first time the client receives an unusually high bill becomes a customer-relationship incident.

## Who will spot drift before the client does?

Listing the signs of degradation in the dossier is not enough. The next question is who actually watches them. The article "The FDE Leaves: Who Owns Your AI After Go-Live?" on the Master of Project blog includes a table on drift monitoring. In the column for "who detects it by default", the usual answer is nobody, until users complain.

The same article argues that an AI system needs an owner for its whole lifecycle, beyond the build phase. For the order-email agent, that advice can become a routine weekly task.

Say that every Monday the assigned support engineer pulls 50 random tickets created by the agent and compares them with the original emails. The alert threshold is 3 wrong tickets, or 6%.

With that number, "keeps getting things wrong lately" becomes data that can be checked. One error in 50 in the first week is reassuring. Four in 50 in the fourth week, and support opens an internal ticket immediately, before the client even notices.

**Key point:** If nobody is assigned to watch the system, the client will be the one who finds the errors.

## How the acceptance review runs

Let support lead the review while you sit in the answering seat. A sample script, with each question paired with the answer that counts as a pass:

Support lead: "If the unclassifiable rate doubles, where do I look?" It passes when you point to a dashboard and a line in the runbook. It fails if you answer "just ping me".

CSM: "If the client asks why image invoices aren't processed, what do I tell them?" It passes when the "out of scope" list is written down and the client has signed off on it.

Support engineer: "What if the API token expires at midnight?" It passes when there is a named person on the client side authorized to reissue the token, with contact details.

Any question that fails goes on an open-items list with a deadline. Support signs off only when that list is empty, or when both sides agree the remaining items are acceptable.

## Start on day one, not in the final week

This process works only if the first two steps start early. Readiness criteria, including cost, should be agreed with support and CS as soon as the project kicks off. The handoff dossier should be written alongside the build, while you still remember why you made each design choice.

The final stage needs concrete milestones. Following Master of Project's suggestion, fix a hypercare window, the period in which you still provide support after leaving the project, and the date on which the client-side owner formally signs off. Add a record of what was built, just as PostHog hands the relationship back to the account owner.

## The mistakes that sink a handoff

The most common mistake is treating the final demo as the handoff. A demo proves the system runs; a handoff has to prove that other people can run it. The second mistake is hypercare with no end date. You get pulled back indefinitely, and support never truly owns the system.

The third is harder to see: measuring only uptime. An agent can run 100% of the time and still get more wrong every day. The signals you monitor therefore have to measure output quality and whether the client has actually changed how they work, which is Rocketlane's test.

## This skill belongs on your CV

When reading FDE job postings, watch for words such as "handoff", "enablement" and "transition to support". If they appear, ask in the interview how long the FDE stays with the client after go-live and who takes over the system afterward. When it is your turn to tell stories, a specific handoff is worth as much as a build.

On your CV, do not stop at "deployed an agent". Write that you handed the system to support with a runbook, readiness criteria and a weekly drift-check process, then say what happened after you left.

A good handoff is measured three months later: the client no longer needs to remember your name, and the system still works correctly.

**Try this week:**

- Pick a system you are about to hand off, list everything the client must do without you, and mark which of those tasks are not yet documented.
- Fill in the six-item handoff dossier from this article for that system and send it to the support lead to read in advance.
- Book a 45-minute acceptance review in which support asks the questions and you may answer only with documentation.

## Sources

- [Working with sales and CS - Handbook (PostHog)](https://posthog.com/handbook/forward-deployed-engineering/working-with-sales-and-cs)

- [The Project Handoff (Educative, Forward Deployed Engineer for AI Systems)](https://www.educative.io/courses/forward-deployed-engineer/the-project-handoff)

- [The FDE Leaves: Who Owns Your AI After Go-Live?](https://blog.masterofproject.com/who-owns-ai-after-go-live/)

- [Why Go-Live Is the Wrong Finish Line—and What to Measure Instead](https://blog.rocketlane.com/blogs/implementation-value-realization)

- [Chapter 32 - The Evolving SRE Engagement Model (Google SRE book)](https://sre.google/sre-book/evolving-sre-engagement-model/)
