Gartner predicts that by 2028, 70% of enterprises will abandon agentic AI built by vendor forward-deployed engineers, trapped by soaring costs and unable to evolve it on their own. A preprint that mined 17,703 open-source repositories estimates that 82% of the migrations it could date back to retired AI models occurred after the model had been shut down. Owning an agent takes more than a clause that says you own it.
On September 29, 2026, Gartner published a prediction about AI services contracts. By 2028, the firm expects 70% of enterprises to abandon agentic AI built by vendor forward-deployed engineering, which it defines as "a model in which vendor engineers work directly with customers to build and deploy solutions."
Two cautions. It is a prediction from a firm that sells advisory services, with no survey cited, and it is scoped to agents built by a vendor's embedded engineers, not all agentic AI.
I should say where Golabs sits. We build AI agents for clients using embedded engineering teams, so this subject cuts both ways for us. Gartner's remark that a traditional services or partner model could potentially deliver some of this work more cost-effectively flatters firms like ours. Its warning about consulting sold under a more strategic label is a charge a services firm like ours could face. For that reason, the contract and ownership sections lean on government, court, and academic sources where they can. I am also not a lawyer, and nothing here is legal advice. Take the contract questions to counsel.
Who Owns an AI Agent After the Vendor's Engineers Leave?
Usually whoever the contract says, and only for the parts that can be owned. Owning an agent built by a vendor's engineers means four things: legal rights to the code and configuration, possession of the data and evaluation sets, knowledge of why it behaves as it does, and the ability to change it. The last decides whether the agent survives.
An IBM Institute for Business Value study with Oxford Economics, published June 17, 2026, surveyed 1,000 senior executives across 16 countries. Seventy-one percent said switching their primary AI vendor or model would be difficult, and 91% said they do not fully understand their organization's AI dependencies. IBM sells consulting and infrastructure that benefit from that finding.
Why Does Gartner Expect Most Enterprises to Abandon Agents Built by Vendor Engineers?
Because, in Gartner's words, these engagements "often fail structurally before they fail technically." Gartner says customers may see faster early progress but fail to build internal capability, and end up paying premium rates for work that a traditional services or partner model could potentially deliver more cost-effectively once the scope is clear.
Mukul Saha, a Sr Director Analyst at Gartner, described what the good ones look like: "The best-scoped FDE engagements have clear guidelines on governance, business value delivery, IP ownership, project co-ownership, knowledge transfer, and an exit strategy from day one."
This matters now because the model has spread quickly. OpenAI announced its Deployment Company on May 11, 2026, to embed forward-deployed engineers inside customer organizations. We covered the hiring side of this in why every major AI company is hiring forward deployed engineers. That piece argued the builder should stay accountable through adoption. Gartner's point is the next step: staying must not turn into dependence.
Does "You Own the IP" Mean You Own the Agent?
Not by itself. There are three gaps worth asking your counsel about. All of this is US law. If the engineers work in another country, ask about that country's rules too.
Copyright may not reach everything the agent is made of.
In January 2025, the US Copyright Office announced Part 2 of its artificial intelligence report and said that generative AI outputs can be protected by copyright only when a human author contributes sufficient expressive elements, and that providing prompts alone is not enough. The D.C. Circuit held in Thaler v. Perlmutter on March 18, 2025, that the Copyright Act requires a human author, and the Supreme Court declined to review the case on March 2, 2026. Neither the announcement nor the opinion addresses software or agent configurations directly, and the court was clear that its rule "does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence." Our reading, not a legal opinion, is narrower: where large parts of an agent's code and configuration were generated by AI tools with little human shaping, a clause assigning "all copyright" may transfer less than you assume. The practical answer is to contract for delivery, access, and confidentiality as well, and not rely on copyright alone.
Commissioned software is not automatically yours. The Copyright Office's circular on works made for hire explains that a commissioned work qualifies only if it falls within one of the nine listed categories and the parties have signed a written agreement. Software is not named among the nine. Whether a particular program fits one of them is a question for counsel, which is why contracts commonly carry an express assignment clause as well. Confirm yours has one.
The vendor's own tooling stays the vendor's. Frameworks, orchestration layers, and test harnesses that the vendor brought to the project usually remain the vendor's property. If the agent cannot run or be tested without them, you need a license that outlives the engagement. That is common contracting practice, not a finding from a source.
What Happens When the Model Under Your Agent Is Retired?
An agent pinned to that model stops working on a date someone else chose.
OpenAI's published policy is to provide at least six months' notice before retiring a generally available model, unless safety or compliance concerns require a shorter timeline; specialized and preview models receive shorter notice periods. Anthropic's policy is at least 60 days' notice for publicly released models, and its documentation is blunt about what follows: "Requests to retired models will fail." As of October 5, 2026, that page listed Claude Sonnet 4.5 as deprecated on September 30, 2026, with retirement set for November 30, 2026.
A September 2026 preprint by Hyungjin Lukas Kim looked at what application teams actually do with that notice. It mined GitHub for commits that migrate away from deprecated models at OpenAI, Anthropic, and Google, across 17,703 repositories from 2024 to 2026. The paper estimates that 82% of migrations away from retired models occurred after the shutdown date, which the author interprets as the point at which the application began failing. Model identifiers were hard-coded in 94% of the migrating applications.
These figures describe open-source applications, not enterprise agents. The paper is a preprint, posted without peer review. The pattern weakens with longer notice: the late share was 89% under Anthropic's 60 to 114-day notices and 13% under OpenAI's one-year Assistants API notice. Even with their own code and advance warning, most of these projects moved late. A team that cannot change its agent at all is in a worse position.
If the only people who can swap the model and rerun the evaluations work for the vendor, the engagement never really ends.
What Should the Contract Say Before Anyone Writes Code?
The White House Office of Management and Budget's April 2025 memorandum on federal AI acquisition spells it out for agencies: "Protections against vendor lock-in can vary, but include requirements for vendor knowledge transfers, data and model portability, providing agencies with rights to code and models produced in performance of a contract, and transparency in licensing and pricing."
The 2023 interagency guidance from US banking regulators on third-party relationships, which is non-binding, tells banks to consider source code escrow, termination terms that allow an orderly transition without prohibitive expense, the timely return or destruction of their data, and who bears the cost of transition.
The memorandum binds federal agencies, and the guidance is written for banks. Neither applies to companies outside those sectors. Together they still make a free checklist.
Everything that makes the agent work, in your repository. Code, prompts, tool definitions, configuration, evaluation sets, test harnesses, and runbooks delivered continuously throughout the engagement, not in a handover package at the end.
Rights to what is produced, and a license to what is not. Ownership or assignment of the code and models built under the contract, plus a license to the vendor's pre-existing tooling that survives termination.
Data and model portability. Fine-tuned models, embeddings, logs, and evaluation results in formats you can load somewhere else.
A knowledge transfer plan with names on your side. Gartner's advice for the delivery phase is to embed the vendor's engineers with internal domain experts, engineers, and end users so that critical knowledge is shared.
Transition help at a known price. Agree the terms and rates for exit support at signing, while you still have leverage.
An independence test. Before the engagement closes, your own team makes a real change without vendor help: swap the model version, add a tool, fix a failing evaluation. Gartner calls its final phase "Exit and Prove Independence" and advises executing the exit plan set at the start, not extending the engagement because internal teams are not ready.
How Do You Tell Forward-Deployed Engineering From Consulting With a New Label?
Gartner's second prediction is that through 2028, fewer than 20% of these engagements will turn recurring customer needs into capabilities in the vendor's core product. It calls the result "FDE washing," where consulting services are marketed as forward-deployed engineering. Saha put it this way: "Many providers now use 'forward deployed' as a label for implementation, professional services, solution engineering, or AI consulting; some thoughtfully, others because it sounds more strategic."
Gartner advises using forward-deployed engineering only for problems that require deep product expertise, rapid adaptation, or close integration between the vendor's technology and your operating environment.
That applies to us too. Golabs is a services firm. We have no core product for your requirements to flow back into, so the honest description of what we sell is services: engineers we place with your team, and tailored AI agents we build in your environment. Our own solutions page says everything we build is your intellectual property. Hold that promise to the same list: delivery into your repository, named knowledge transfer, and an independence test before we go. The questions in how to choose an AI development partner apply as well.
Common Questions About Owning a Vendor-Built AI Agent
Who owns an AI agent built by a vendor's engineers?
Usually, whoever the contract says, and ownership has several parts. Legal rights to code and configuration are one. Possession of data, prompts, and evaluation sets is another. The knowledge to operate the agent and the ability to modify it are what determine whether you can keep running it once the vendor's engineers have left.
What is FDE washing?
FDE washing is Gartner's term for consulting services marketed as forward-deployed engineering. Gartner predicts that through 2028, fewer than 20% of forward-deployed engineering engagements will turn recurring customer needs into capabilities in the vendor's core product. In our reading, that product feedback loop is what separates the model from ordinary professional services.
Can AI-generated code be copyrighted in the United States?
It depends on the human contribution. The US Copyright Office concluded in January 2025 that generative AI outputs can be protected only where a human author determined sufficient expressive elements, and that prompts alone are not enough. Neither the Copyright Office's announcement nor the D.C. Circuit's Thaler decision addresses software specifically, so ask counsel how this applies to your contract.
How much notice do AI providers give before retiring a model?
It varies by provider. As of October 5, 2026, OpenAI's stated minimum is six months for generally available models, barring safety or compliance exceptions, and Anthropic's is at least 60 days for publicly released models. Anthropic's documentation says requests to retired models will fail, so an agent built on one must be migrated and retested before that date.
If you are about to sign an AI engagement, or are partway through one and unsure what you will be left holding, talk to the Golabs team. We will start with the exit plan and work backward.