Gartner forecasts that 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineers by 2028, in a September 29 release reported the next day by The Register and Channel Dive. The firm calls the pattern “FDE washing.” Its analyst counts about 2,000 active FDEs against roughly four times that demand, and Gartner estimates that FDE consulting fees alone could top $200,000 per quarter per use case.
Gartner’s stated mechanism is dependency and exit: clients end up unable to run or leave what the vendor built. How the pod is designed is not part of its analysis. This post argues the dependency has a structural cause visible in the published FDE pods. The pods that name their roles each have a strategist, and none has a grader. No seat checks the work against criteria the client agreed to, held by someone who had no hand in building it. That link is this post’s argument. Gartner doesn’t make it.
The pod is already the standard delivery model
The lone engineer embedded at a client is mostly a recruiting image. The published delivery unit is a pod, and the pods look alike.
Palantir pairs the Delta (the FDE) with an Echo (a deployment strategist) and Devs on the product side. One alumnus describes FDEs going on site at the customer three to four days a week. Deloitte’s lead-FDE postings describe pods of two to five onshore-anchored, offshore-supported engineers. Thoughtworks’ principal-FDE posting describes a pod of one product strategist, one experience designer and two forward-deployed engineers. AWS’s billion-dollar FDE organization deploys pods of five or six people, embedded inside a single customer at a time. Anthropic’s own postings pair FDEs with a technical deployment lead who won’t write production code and owns the roadmap the engineers execute against. The shape repeats: one or two engineers, a lead who owns scope and the executive relationship, sometimes a designer, and offshore or platform support behind them. Postings change, so the specifics will drift. The shape hasn’t.
The supply side is about to scale this template. Anthropic’s Claude Frontier Academy commits $100M to train 10,000 “Frontier Deployed Engineers” by the end of 2027 at partners including Accenture, Deloitte, McKinsey and Morgan Stanley. Accenture’s statement says those FDEs will work “inside client workflows as part of multidisciplinary teams.” If the partners follow their own postings, those teams will look like the ones above.
Every seat in the pod helped build the thing
Read those pod descriptions for a quality seat and you won’t find one. None of them names a QA seat, an acceptance seat or a grading seat. The closest is a mention of “quality gates” in Deloitte’s posting.
The strategist seat is already standard. Palantir calls it the Echo. Other firms call it a deployment lead, a product strategist or a PM. Pods don’t need anyone to invent it. And every seat in every pod is a builder. Engineers write the code. The strategist sequences the roadmap. The designer shapes the workflow. Platform and offshore support keep it running. Each person present has a stake in the result passing.
Verification gets dropped because nobody is waiting on it
The earlier post on the FDE title broke the role into separate jobs. Here’s the version that matters for pods. The role has four jobs: build the system, redesign the process around it, run the client relationship, and verify the result. A pod splits the first three across seats. The fourth goes unassigned.
It gets dropped inside a pod for the same reason a lone engineer drops it. No one is waiting on the grade. Shipping has a demo date, and the executive relationship has a steering meeting. A verification report has no audience until something breaks. Everyone in the room also wants the work to pass, including the vendor’s lead, whose quarterly fee renews on the impression that it did.
Follow that to the client’s side. Without an independent grade against agreed criteria, the client can’t tell what it bought. Without a sponsor who has authority to stop the work (the one seat the kill sheet from the earlier post insists on), the client can’t leave without losing whatever value is in there. It can’t separate the working parts from the broken parts, so it keeps paying the only people who know. That is dependency, and that’s what Gartner is forecasting.
The missing seat grades and does nothing else
Call it the acceptance owner. The definition is narrow on purpose:
- Co-writes testable acceptance criteria with the client sponsor before work starts
- Grades the outcome against those criteria after every change
- Publishes the grade to the sponsor, not to the pod first
- Had no hand in building the work
The seat can sit on either side of the contract. A mature client’s own QA function can hold it. A third party can. A vendor employee from outside the delivery org can, if the client accepts that. What disqualifies someone is having built the work. Which side signs their paycheck matters less.
Here’s the cadence relative to the pod:
| Week | Pod | Acceptance owner | Sponsor |
|---|---|---|---|
| 0 | Scopes the use case | Co-writes criteria with sponsor; each one testable pass/fail | Signs criteria and stop conditions |
| 1-5 | Builds, demos weekly | Re-runs the criteria after each merged change; logs pass/fail | Reads the grade before the demo |
| 6 | Mid-sprint review | Publishes trend: criteria passing, failing, untested | Decides continue, re-scope or stop |
| 7-12 | Builds, hardens | Same loop; adds regression on previously passed criteria | Reads the grade weekly |
| 13 | Quarter close, renewal | Final grade against week-0 criteria, unchanged unless sponsor re-signed | Renews or exits on the grade |
Each seat also has one thing it must not do. The engineer doesn’t grade their own work. The strategist can write down what done means but doesn’t own the grade. The sponsor never hands the stop decision to someone else. The acceptance owner’s rule is the least obvious one: don’t suggest fixes. The moment the grader proposes the fix, the grader has helped build it.
Owning the criteria doesn’t make the grade independent
The strongest objection is that the pod already has this. Palantir’s Echo, or a deployment lead who owns the roadmap, writes down what done means.
That person owns the criteria. They also own the scope, the sequencing and the executive relationship, and they answer for whether the engagement renews. When a criterion half-passes in week 11, the person who wrote the roadmap is also the person who explains to the sponsor why it’s close enough. The objection fails on that mechanism. Authorship and grading sit in the same chair, and that chair is the one with the most riding on renewal.
The second objection: the client’s own QA team can do this. Often true. When it can, that team is the seat, provided it didn’t help build the system. The test is simple. If the client’s QA engineers pair with the FDEs during the build, they’ve crossed over, and someone else has to grade.
A fractional grader costs about 2.5% to 7.5% of a $200K quarter
The acceptance owner doesn’t need to be full time. The work is a criteria workshop up front, a re-run after each change, and a weekly report. That fits in roughly 0.1 to 0.2 of a seat.
Illustrative pricing, using Gartner’s $200,000-a-quarter estimate:
- Pod fee: $200,000 per quarter for one use case
- Pod of four seats (two engineers, strategist, designer): $50,000 per seat per quarter
- Acceptance owner at 0.1 seat: $5,000 per quarter, or 2.5% of the fee
- At 0.15 seat: $7,500 per quarter, or 3.75%
- At 0.2 seat: $10,000 per quarter, or 5%
An experienced independent grader may bill above the pod’s average seat. Double the rate and 0.15 of a seat costs $15,000 a quarter, or 7.5%.
Now price it against abandonment. A use case run for four quarters at $200,000 costs $800,000. The grader at 0.15 seat costs $30,000 for the same year. If an independent grade ends a failing engagement just one quarter early, it saves $200,000. That covers the seat at $7,500 a quarter for about 26 quarters, more than six years.
The cost of this advice goes beyond the invoice. A grader will fail work the pod considers done. Demos get awkward. Renewals get harder for the vendor. The pod loses some of its speed, because “shipped” now has to mean “passed.”
Where the seat doesn’t fit
An 8-week proof of concept often can’t carry any extra seat, fractional or not. The fix there is cheaper: write the criteria at week 0 and have the sponsor grade at week 8 personally.
Clients who won’t name a sponsor break the model entirely. A grade with no one empowered to act on it is a report nobody reads.
The onshore-anchored, offshore-supported pod is the hardest case. On paper it’s five engineers. In practice one engineer can end up alone at the client, with the rest of the pod a time zone away. That engineer is the whole on-site build team and the only witness to whether it works. A pod like that needs a grader most and is least likely to have budget for one.
Go back to the criterion that half-passes in week 11. With an acceptance owner, it doesn’t get explained in a steering meeting. It shows up in the weekly grade as a fail, with the week-0 wording next to it, and it shows up again in week 12. At week 13 the sponsor holds a sheet with every signed criterion marked pass or fail. That sheet is the seat every published pod is missing. If that line is still red, the sponsor decides whether the next $200,000 buys a fix or funds the exit.