On 29 September Gartner put a number on something many buyers have begun to suspect. By 2028, it predicts, 70% of enterprises will abandon agentic AI built for them by a vendor's forward-deployed engineers, "trapped by soaring costs and unable to evolve it on their own."
Forward-deployed engineering, or FDE, is the model in which a software vendor sends its own engineers to sit inside the client and build the thing. Palantir made it famous. The model labs have since made it fashionable, and every agent platform now claims a version of it. The pitch is speed: the people who know the product best, working on your problem, in your systems, from week one.
Gartner does not dispute the speed. Its argument is that the speed is borrowed, and the bill arrives when the engineers leave.
The number that matters is not 70%
The headline figure will get the attention. The more useful one sits further down. Gartner expects fewer than 20% of FDE engagements to turn a client's custom requirements into standard vendor product features through 2028.
Think about what that means in practice. Everything the embedded team builds for you that does not make it into the product is yours. It is a bespoke fork, written against your data and your workflows by people who will be redeployed to the next account. When the vendor ships a new model, changes an API or retires a connector, the product moves on and your fork does not. Somebody has to keep it working. If that somebody is the vendor, you are paying premium engineering rates for maintenance. If it is your own team, they are maintaining code they did not write, on a platform they were never trained to run.
That is how a system that went live on time ends up switched off two years later. Not because it failed, but because nobody could afford to keep it alive.
Mukul Saha, the Gartner senior director analyst behind the research, put the fix in one line: "Success starts with getting the engagement structure right, from scope and incentives to governance, ownership, and exit." Note the order. Scope comes first, and exit is on the list from the start, not added at the end.
Not every embedded engineer is forward-deployed
Gartner also warns of what it calls "FDE washing": conventional consulting and implementation services rebranded as forward-deployed engineering to justify a higher day rate. As coverage in Techstrong.ai reported, Gartner's concern is that many providers use the label without the depth, programme management or maturity behind it.
The distinction is not academic. Genuine FDE is worth paying for when the engineers have a direct line into the vendor's product team, can change the platform rather than work around it, and bring patterns from dozens of similar deployments. A contractor with a vendor badge who builds custom integrations on top of the product is an implementation partner, and should be priced and managed as one.
Three questions separate the two. Can the embedded engineers raise and influence changes to the core product? How many of the features built on previous engagements are now in the standard release? And who, by name, owns your deployment inside the vendor once the engineers have gone?
A fair note of scepticism
Not everyone accepts Gartner's figure. Jack Gold, an independent technology analyst quoted by Techstrong.ai, disagreed that 70% of enterprises would walk away by 2028, arguing that many will instead move to lightly modified open-source agents. Mitch Ashley of The Futurum Group made a different point: partners and integrators have always built systems that IT later has to take over.
Both are fair. Neither changes the advice. Whether the outcome is abandonment, a rebuild on open source, or a painful handover to IT, the cause is the same: the client never acquired the knowledge and control to run what was built for it.
Why this bites harder in the Gulf
Buyers in the UAE and Saudi Arabia are under unusual pressure to move fast. Dubai's agentic government targets, sector regulators issuing their own AI guidance and board-level expectations set by national strategies all push towards the quickest route to a live system. Experienced agent engineering talent in the region remains thin, which makes a vendor offering to fly in its own team very attractive.
This week Ai Everything Global opens at ADNEC in Abu Dhabi, with enterprise adoption and AI safety on the agenda. Expect a great many stands offering embedded engineers. The ones worth talking to will be the ones who can explain, before the first sprint, how they intend to leave.
There is a second reason the exit matters more here. Data residency rules, sector audit requirements and national workforce programmes all assume that the capability ends up in-country and in-house. An agent that only a visiting vendor team understands sits awkwardly against every one of those expectations.
Plan the exit as a delivery milestone
We wrote in August about the contract clauses that agent delivery needs: ownership, data retention and accountability on exit. Gartner's research makes the case for something further. The exit should be a delivery milestone with acceptance criteria, not a clause that only gets read when the relationship sours. Gartner frames its recommendations in three phases, before signing, during the engagement and at exit, and that maps cleanly on to how we would run it.
Before signing
- Use FDE only where the problem genuinely needs deep product expertise or tight integration with the vendor's platform. Routine implementations belong with a standard partner model.
- Name an internal executive sponsor accountable for the business outcome, not for the vendor relationship.
- Agree in writing what becomes product, what remains a custom build, and who maintains each.
During the engagement
- Pair every vendor engineer with a named internal engineer and a domain expert from the business. Your people should be committing code within weeks, not reading documentation at the end.
- Validate against business measures at each iteration, including whether the operating model around the agent actually works.
- Define decision rights and autonomy levels for the agent before it goes near production.
At exit
- Rehearse the departure. Before the engagement formally ends, the vendor team steps back for a fixed period and your team runs the system alone. Whatever breaks is the real handover backlog.
- Measure success the way Gartner suggests: by your ability to run, change and scale the system without the vendor in the room.
- Resist the extension. An engagement that keeps being renewed because nobody else can operate the system has already failed its exit test.
The bottom line
Forward-deployed engineering is a sound way to get a difficult agent into production quickly. It is a poor way to own one. Gartner's prediction is less a verdict on the model than on how buyers have been using it: paying for the build and assuming the capability would follow.
If you are about to sign an FDE engagement, ask the vendor to describe the day they leave. If the answer is vague, so is your business case.