70% of Enterprises to Abandon AI Agents | The Shocking Gartner Prediction
機械翻訳 / Machine-translated
A
AI Friends
1日前・15分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article
On September 29, 2026, research firm Gartner released a prediction: 70% of enterprises will abandon AI agents built by vendor engineers by 2028.
This means that the majority of AI agent projects companies are currently investing tens of millions to hundreds of millions of yen in will go unused just two years from now. Why is this happening?
The core of the problem lies in a construction model known as FDE. FDE stands for Forward Deployed Engineering — a model in which vendor engineers are embedded at the customer's company and build an AI system tailored specifically for that organization.
On the surface, this sounds ideal. After all, you get an AI perfectly suited to your business operations. But Gartner points out that this model has a serious flaw.
The biggest issue Gartner identifies is escalating costs and a lack of retained knowledge. Things run smoothly while the vendor's engineers are on-site. Once they leave, the company has no know-how left to maintain or improve the system.
As a result, every minor fix or feature addition requires calling the vendor back in. Unplanned additional costs arise one after another, and the project becomes unsustainable.
On top of that, fewer than 20% of FDE contracts successfully convert customized requirements into standard product features. Features built for your company remain exclusive to your company and contribute nothing to the vendor's product improvements.
The industry has even coined the term "FDE washing" — slapping the "FDE" label onto traditional IT consulting and charging premium rates for it. These engagements offer no deep product knowledge, only surface-level deployment support.
The problem isn't limited to the FDE model. AI agents broadly face significant implementation challenges. Research shows that 88% of enterprise-developed AI agents never even make it to production.
They perform brilliantly in test environments, but the moment you try to use them in actual operations, they stop working. These situations are occurring one after another.
Real-world failure cases have been reported as well. In April 2026, at car rental software company PocketOS, a coding AI agent deleted the production database and its backup in a matter of seconds.
At one fintech company, an AI agent for transaction reconciliation was refined to 99.2% accuracy in testing. During 30 days of live operation, it generated a flood of incorrect reconciliation errors and ended up locking $2.3 million (approximately ¥330 million) in customer accounts.
Unlike traditional software, AI agents operate probabilistically. They call tools in unpredictable sequences and take unexpected actions. Conventional monitoring tools simply can't keep up with what's happening.
Gartner issued another prediction in February 2026: more than 40% of agentic AI projects will be cancelled by the end of 2027.
The reason isn't technological limitations. It's operational difficulty, inadequate governance, and the inability to demonstrate measurable business value.
Gartner also predicts that by 2027, 40% of enterprises will downgrade or withdraw autonomous AI agents due to governance deficiencies discovered after deployment.
It's not enough for AI agents to simply "work." Who bears responsibility and to what extent? What decisions are they permitted to make? What happens when something goes wrong? Without these structures in place, they ultimately become unusable.
Japanese companies are in an even more difficult position. The most common cause of AI agent deployment failure is unclear ownership — no one is clearly in charge.
An executive says, "Introduce AI," and the person in charge says, "We'll try." With no one truly accountable for driving the project forward, it ends up adrift.
Moreover, Japanese companies face a double barrier: not only a shortage of AI talent, but also an absence of a culture of using AI. Even if engineers are hired, if the frontline workforce has no concept of incorporating AI into their daily work, adoption goes nowhere.
Real failure cases have been reported. One company handed off a task to an AI agent without narrowing its scope, resulting in a single execution running for more than nine hours.
Another company allowed an AI agent to send emails to business partners without any confirmation step, and the emails went out with incorrect content. There's also a case where an AI cited a non-existent internal policy in a document, and no one caught it before it was used.
API cost control is another challenge. Some companies have experienced costs far exceeding expectations because they had no mechanism in place to stop an AI agent mid-execution.
So what does it take to successfully deploy AI agents? Gartner advises that if you use the FDE model at all, it should be limited to situations that meet the following three conditions.
First, the problem requires deep product knowledge — a challenge that only engineers with an intimate understanding of the vendor's product can solve.
Second, rapid adaptation is essential — you need to respond quickly to market shifts or competitive moves.
Third, tight integration between the vendor's technology and the customer's operational environment is required — the work demands sophisticated integration that goes well beyond a simple deployment.
If these conditions don't apply, FDE should be avoided. Instead, consider choosing standard products that your organization can maintain on its own, or explore building in-house.
AI agent failures are not caused by immature technology. They are caused by organizational unpreparedness.
Who is accountable? How far is automation permitted to go? What happens when something fails? Deploy without these structures in place, and even the most sophisticated AI will fail.
Conversely, with the right structures in place, AI agents can generate tremendous value. In fact, some companies have already used AI agents to dramatically improve operational efficiency and create entirely new business models.
What's needed right now is not the latest AI model. It's the organizational framework to harness AI effectively. Before rushing to adopt new technology, companies should first take a hard look at their own readiness.
This article is a cross-post from AI Friends.
