FDE 項目結束後，誰負責 Agent 交付後首季的維護？

Gartner forecasts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering, "trapped by soaring costs and unable to evolve it on their own." Gartner publishes it as a forecast. This post argues the mechanism behind it shows up in the budget. Gartner's own fee estimate is quoted per quarter per use case, which reads as a build price that runs until handoff. Nothing in the estimate as reported covers the quarter after it, so unless the contract says otherwise, nobody is named to own the agent or funded to keep it working, and the first forced change lands on a team that never built it.

Call that period quarter two: the first quarter the agent runs without the people who wrote it.

The $200K quarter is the floor, and it's priced as a build

Gartner estimates, reported by CIO Dive, suggest FDE consulting fees before platform and integration costs could range between $200,000 and $400,000 quarterly per use case. So $200K is the bottom of the range, and it covers consulting fees alone. Platform and integration costs sit outside it.

In the engagement this post has in mind, a fee priced that way points at one milestone: the engineers ship, the acceptance demo passes, the handoff deck goes out, and then they leave.

The Register's summary of the report names the failure: engagements fail "when customers do not acquire the knowledge and control needed to maintain and develop the resulting systems after the vendor leaves." Gartner's own Phase 1 advice says to "identify an executive sponsor who is accountable for business outcomes, not simply the implementation budget." An implementation budget ends at handoff. A business outcome doesn't.

The first forced change already has a date

The vendor leaving doesn't freeze the agent's dependencies. Anthropic's model deprecations page commits to "at least 60 days' notice before model retirement for publicly released models." That's a reasonable, published commitment. It's also a calendar someone has to read.

Take an illustrative engagement: two quarters, April through September 2026, with the agent built on claude-sonnet-4-5-20250929. Per the same deprecations table, that model was deprecated on September 30, 2026, with a tentative retirement of November 30, 2026. Sixty-one days. The recommended replacement is claude-sonnet-5-5. Requests to a retired model fail.

So the deprecation notice arrives in handoff week, and the retirement falls inside quarter two. The work is ordinary: swap the model ID, review prompts tuned against the old model's behavior, re-run the acceptance evals, compare outputs, redeploy. Nothing exotic. All of it needs credentials, the eval suite and a person whose job it is. Here is what that entry looks like when someone owns it (illustrative):

model_calendar:
  - model_id: claude-sonnet-4-5-20250929
    deprecated: 2026-09-30
    retires: 2026-11-30        # tentative, per provider page
    replacement: claude-sonnet-5-5
    migration_owner: "<named person, not a team>"
    evals_rerun_by: 2026-11-09
    last_reviewed: 2026-10-01  # reviewed monthly

A schema change in a source system or an eval score that drifts follows the same pattern, usually with less warning.

In illustrative math, the ownership line costs a fraction of one engagement quarter

Keyhole Software, a consultancy that sells builds, estimates that "annual run and maintenance costs then add 15% to 25% of the build price every year." It also says the maintenance budget "belongs in the original statement of work as a percentage, not a change order filed after the first incident." That range describes Keyhole's own build prices, not Gartner's FDE consulting fees. Applying the percentage to a $400K fee-only build, as below, is this post's illustrative transfer, not a sourced figure.

Illustrative arithmetic, at the bottom of Gartner's fee range:

Build, two quarters at $200K, fees only: $400,000
Run cost at 15% to 25% a year: $60,000 to $100,000
Per quarter: $15,000 to $25,000
As a share of one $200K engagement quarter: 7.5% to 12.5%
Cost of abandoning instead: the $400,000, plus the platform and integration spend Gartner's range leaves out

On these numbers the ownership line is the cheap part. It's also a part the fee estimate, as reported, says nothing about. And it has a cost the dollar figure hides: the owner is an engineer whose hours come out of some other roadmap, and the sponsor has to defend a budget line for a system that already "shipped."

The ownership sheet gets signed before the SOW

The grader post priced checking the work during the build. The sheet below covers who owns it after the vendor goes.

Line	What gets written down	Rule
Owner	One named internal person	Not a team, and not the vendor
Budget	Run cost as a percent of build	In the SOW, not a change order
Model calendar	Every model ID the agent calls, with its published retirement date	Reviewed monthly
Evals	The acceptance suite, handed over	Runnable by the owner without the vendor
On-call	Who gets paged	Plus what they're allowed to change
Kill criteria	The condition under which the sponsor turns it off	Signed by the sponsor (the kill-sheet post has a five-field version with a named cancel authority)

The week the vendor leaves, five things should be true:

The owner can redeploy the agent without vendor credentials.
The owner has run the full eval suite once, alone.
The model calendar lists every model ID in production.
The first migration has a date on it.
The sponsor has signed the kill criteria.

If any of the five is false, the exit plan isn't finished.

Extending the vendor buys more of the same problem

The obvious objection: keep the vendor on for run and support. Gartner's Phase 3, titled "Exit and Prove Independence," rejects that in plain words: "Execute the exit plan established at the start rather than extending the engagement because internal teams are not ready." The mechanism sits in the forecast's own phrasing. An extension at FDE rates is the "soaring costs" half. And while the vendor carries the pager, the client's team has less reason to learn the system, which feeds the "unable to evolve it" half. Extension compounds the gap it was meant to cover.

The second objection: the vendor will fold the custom work into its product, so there's nothing to own. Gartner predicts that through 2028, less than 20% of FDE engagements will turn recurring customer needs into capabilities in the vendor's core product. If that forecast holds, more than 80% of engagements won't turn recurring needs into product capabilities, and whatever they built still has to be kept working by someone outside the vendor's product team. Who holds the IP depends on the contract. Either way, the system keeps running on the client's side, and someone there has to keep it working.

The third: agents are cheap to keep running. The token bill may well be small. The scheduled work isn't optional. A 60-day retirement notice, an eval re-run, an upstream API that changes its response shape: each arrives on someone else's calendar. Cheap inference with no owner still breaks on retirement day, tentatively November 30.

A pilot built to die needs only the last row

The sheet doesn't transfer everywhere. A pilot designed to be killed at handoff needs no quarter-two line. It needs kill criteria and a date, and nothing else on the sheet applies.

A client with no internal engineering can't hold the owner seat at all. Naming a project manager as owner of a system they can't redeploy is theater. The honest answer there is a priced managed service with an exit clause, chosen on purpose. What fails is the version that gets drifted into, one renewal at a time, at FDE rates.

For everyone else, the unowned quarter is the mechanism this post argues sits behind Gartner's 70% forecast, and the fix is one line signed before the engagement starts. For the illustrative summer build, the first entry on that sheet is already known: claude-sonnet-4-5-20250929, tentatively retires November 30, 2026, owner: ________.