Gartner predicts 70% will drop agents built by vendor engineers by 2028. What the contract needs

What Gartner predicted

Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineering.¹ It published the prediction on 29 September 2026 and describes those enterprises as "trapped by soaring costs and unable to evolve it on their own".¹

Forward-deployed engineering means a vendor sends its own engineers to sit with your team and build agents on your systems. It is fast, and it puts scarce skills where the problem is. Gartner's warning is about what comes after: the knowledge leaves with the engineers, the vendor stays on the invoice, and the agent cannot change without them.

Read the number for what it is. It is a forecast, not a measured abandonment rate, and the press release does not publish the model behind it. Our view: the direction is credible even though the figure cannot be tested yet, because the causes it names are ones a buyer can check in a contract today.

Why the failure is structural, not technical

The causes Gartner names are cost and ownership, not model quality. A better agent fixes neither. Both are settled, well or badly, when the contract is signed.

Cost grows when every change needs the vendor. Prompts, tool connections, evaluation sets and model choices all need adjusting as the business changes. If only the vendor's engineers know where those live, each adjustment becomes a change request with a day rate.

Ownership fails when nobody inside owns the result. Gartner recommends governance, IP ownership, co-ownership, knowledge transfer and an exit strategy from day one, and executing the exit plan rather than extending the engagement.¹ That last point is the hard one, because extending is always the easier decision in the month it is made.

Buyers already check cost at the start. In KPMG's survey of 314 US leaders at firms with $1 billion or more in revenue (24 July–25 August 2026), 74% included cost reviews in AI approvals.² An approval-stage review rarely prices the cost of being unable to leave, though, and that is the cost Gartner is pointing at.

What should the contract say?

Six clauses turn Gartner's advice into terms you can sign. They fall into three moments: before contracting, during delivery and at transition.

Illustrative. Our translation of Gartner's recommendations into clause headings; take legal advice on the wording.

Ownership has to cover the parts that are not code

Prompts, tool definitions, evaluation sets and configuration are where most of an agent's behaviour lives. A contract that assigns "source code" or "deliverables" to you may not cover them. Name each one, and require it in your repository rather than the vendor's.

Include logs and evaluation results too. Without the record of how the agent behaved, a new team cannot tell a regression from normal variation. Our pilot-to-production checklist lists what should exist before an agent leaves the pilot.

The exit needs a test, not only a date

A dated exit plan says when the vendor steps back and what your team must be able to do by then. Typical tasks: change a prompt, add a tool, rerun the evaluations and roll back a release without help. Tie the final payment to your team passing those tasks.

Write the extension terms in advance, priced and time-limited. An extension then becomes a decision someone signs, rather than the default when the exit date arrives.

Where each clause bites

1. Before contracting

   Limit scope to work that needs vendor depth, and name the internal owner.

2. During delivery

   Your repository, your staff pairing, paid knowledge-transfer deliverables.

3. At transition

   Your team passes the exit tasks, the final payment is released, and extensions are priced in advance.

How this fits Gartner's older cancellation forecast

The new prediction narrows an older one to a single delivery model. Gartner predicted in June 2025, more than a year before this note, that over 40% of agentic AI projects will be cancelled by the end of 2027.³ It named escalating costs, unclear business value and inadequate risk controls as the causes.

Both are forecasts. Read together, they say cost and control keep recurring as reasons agents fail, and the vendor-built variant adds dependence on people who leave. Running costs are the part to model early, and our guide to an agent's monthly running costs breaks them down. If you are buying a platform as well as a build, our due-diligence questions for agent platforms cover what to ask before signing.

Our interest, and what we would do

Sigzen AI builds agents for clients, so we sell a version of the work Gartner is warning about. Weigh our view with that in mind. We think the six clauses should apply to us as much as to any vendor.

Our view: if a vendor resists naming prompts and configuration as yours, or will not price an exit, that tells you how the engagement is designed to end. A short build that leaves your team able to change the agent is worth more than a long one that does not. Our AI automation page describes how we scope a build.