Seven in Ten Firms Will Walk Away From Vendor-Built Agentic AI, Gartner Forecasts
Gartner forecasts 70% of enterprises will abandon vendor-built agentic AI by 2028 as ownership costs and lost maintenance control drive the retreat.
Gartner expects seven in ten enterprises to abandon vendor-built agentic AI that was assembled for them through forward-deployed engineering, a call that shifts the enterprise AI argument away from model quality and onto ownership cost. The research firm's forecast, issued on 29 September 2026, puts the abandonment rate at 70% by 2028. It cites two forces: running costs that keep climbing after go-live, and internal teams that cannot maintain or extend the systems once the vendor's engineers move to the next account.
The second force carries the more uncomfortable implication. Forward-deployed engineering, or FDE, places vendor specialists inside the customer's organisation to build agents against that customer's data, workflows and legacy systems. The agents usually work. What the buyer does not acquire is the ability to keep them working.
What Vendor-Built Agentic AI Actually Transfers
FDE is a staffing model before it is a product. Vendors bill for embedded engineers, so the cost tracks headcount rather than usage or outcome. That structure does not shrink when a pilot becomes production. It grows, because production agents need monitoring, retraining and integration work that the original build team already understood and rarely documented.
Gartner's companion figure sharpens the problem. Fewer than 20% of FDE projects are expected to produce core product capabilities for the vendors themselves. If the vendor is not distilling reusable product from the engagement, the customer is funding bespoke work that nobody amortises. The buyer pays full price for a system with one supplier and no second market.
That is where lock-in becomes concrete rather than contractual. A handover dependent on undocumented prompt chains, customer-specific tool integrations and an engineer's tacit knowledge is not a handover. It is a standing order for the people who built it.
FDE Washing and the Procurement Blind Spot
Gartner also flags what it calls "FDE washing", the practice of marketing standard consulting and professional services as strategic forward-deployed engineering. Once that label stretches, it stops carrying information for procurement teams. A rebadged systems-integration contract and a genuinely embedded engineering team can look identical in a proposal, and the difference surfaces eighteen months later when the buyer tries to change a workflow.
Gartner's guidance to software engineering leaders follows directly: fix scope before signature, and treat knowledge transfer as a deliverable with its own acceptance criteria rather than a courtesy at the end of an engagement. Scope discipline is inexpensive at procurement and very expensive at renewal.
Why the Constraint Moved From Capability to Cost
For the past three years, the enterprise question about agents was whether they could do the work at all. Inside pilot scope, that question is largely settled. What remains open is whether a system can be afforded, governed and altered across a multi-year life. Gartner's framing places ownership cost, not model quality, at the binding constraint on how widely agents get deployed.
The practical effect is a change in evaluation criteria. Benchmarks and demo quality decide which vendors reach the shortlist. Total cost of ownership, handover terms and the buyer's own staffing plan decide which of them are still in the account in 2028.
The Vendor Incentive Problem
The economics point in an awkward direction. If fewer than one engagement in five yields reusable vendor product, revenue depends on billable time, and billable time ends when a handover succeeds. The mismatch is structural. A vendor's fastest route to profit and a customer's fastest route to independence are not the same route.
Vendors with a genuine product layer underneath the engagement have a different calculation, because a working customer deployment feeds their roadmap. Buyers cannot easily tell the two apart from the outside. Asking what share of a vendor's last ten engagements became reusable product is one way to find out.
The Perpetual Rewrite Nobody Budgeted For
Cost pressure is showing up well beyond FDE contracts. Discussions at Constellation Research's AI Forum this month centred on cost optimisation, model routing and governance, alongside a newer pattern: AI software that must be rewritten repeatedly because the models and agent frameworks underneath production systems keep moving.
This is the mechanism that converts a working agent into a liability. An agent built in 2026 against a specific model version and a specific orchestration framework does not stay stable when either layer changes. Model deprecations, shifts in tool-calling behaviour and framework updates to state handling all generate re-engineering work. When the vendor holds that knowledge, the buyer's only lever is the vendor's rate card.
Model routing cuts one kind of dependency and adds another. Spreading workloads across several providers reduces exposure to any single model supplier, and it widens the orchestration surface that somebody has to maintain. Governance requirements layered on top, including audit trails and permission boundaries for agents that take real actions, add a further maintenance load that lands on whoever owns the code.
The Numbers Gartner Is Putting on the Table
| Metric | Gartner projection |
|---|---|
| Enterprises abandoning vendor-built agentic AI | 70% |
| Deadline for the forecast | 2028 |
| FDE projects producing core product capability for the vendor | Under 20% |
Read together, the two percentages describe a market where the build side and the keep side are misaligned. Vendors are not converting engagements into product at scale. Customers are not converting engagements into capability. Both are absorbing the cost of a delivery model designed around speed to first demo.
The 70% figure is a forecast rather than a measurement, and abandonment forecasts tend to overstate clean exits. Most enterprises will not cancel a working agent. They will let it run unmodified until it breaks, then decline to rebuild it. That pattern produces the same outcome on a longer timeline, and it is harder to spot in a budget line.
What Abandonment Actually Looks Like
Abandonment rarely arrives as a formal cancellation. The more common sequence is a feature freeze: the agent keeps handling the workflow it was built for, the team stops requesting changes because every change requires the vendor, and the renewal proceeds at a lower tier while the integration work quietly reverts to manual handling. Gartner's cost driver and its maintenance driver describe the same endpoint from two directions.
That distinction matters for anyone tracking the forecast. A 70% abandonment rate by 2028 does not require 70% of buyers to terminate contracts. It requires a majority to conclude that the system cannot be evolved without the supplier, and to stop investing in it. The signal will appear first in renewal scope, then in headcount, and only later in cancelled agreements.
What the Enterprises Outside the 70% Do Differently
The organisations most likely to keep their agents in production treat maintenance as first-class scope. That means written handover requirements, abstractions that do not bind the system to one framework, internal engineers assigned to receive the work, and contractual terms covering who pays when an underlying model changes. It costs more in year one and considerably less in year three.
It also means accepting slower delivery. The FDE model sells speed to first working demo, and that speed is real. The trade on offer is faster proof against a higher cost of ownership. Gartner's forecast is a claim about which side of that trade most buyers will come to regret.
Why This Matters
Gartner's forecast reframes agentic AI as a lifecycle purchase rather than a build purchase, and that changes what a good contract looks like. Buyers who score FDE proposals on time to first working agent are optimising the cheapest and most visible phase of a system that will need re-engineering. The questions that decide whether an agent is still running in 2028 are the ones about who holds the knowledge, who absorbs model churn, and what the exit costs.
Sources
Related Articles
- Genpact Report: 92% of Executives Forecast Agentic AI Will Reshape Corporate Workflows
- Deloitte's 89% AI Agent Pilot Failure Rate Exposes the Enterprise Production Gap
- McKinsey: Enterprise AI Budget Overruns Tied to Coding Agent Sprawl
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.