# 4 in 10 AI agent projects will be canceled > Gartner gives three reasons: rising costs, value that is never proven and risks without controls. Three real cases and how to cover each one from day one. [Platform](https://www.shimliapp.com/en/blog/category/plataforma) ## 4 in 10 AI agent projects will be canceled Gartner gives three reasons: rising costs, value that is never proven and risks without controls. Three real cases and how to cover each one from day one. **Luis Aguilar** October 12, 2026 6 min read IN SHORT - Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, due to costs, unclear value or inadequate risk controls. - Of the thousands of vendors presenting themselves as agentic, Gartner estimates about 130 offer real agents. - Air Canada had to pay C$812 in 2024 for a refund policy its chatbot made up: the company answers for what its agent says. - Commonwealth Bank reversed in 2025 the cut of 45 roles it had made over a voice bot, and admitted the decision was an error. CONTENTS - Where companies stood at the start - Reason 1: rising costs - Reason 2: value that is never proven - Reason 3: risks without controls - The project that gets canceled and the one that scales - How Shimli handles it In June 2025, Gartner published a prediction that still circulates in every committee evaluating AI agents: **over 40% of agentic AI projects will be canceled by the end of 2027.** The most useful part of that prediction is not the number, it is the reasons. Gartner names three: **rising costs, unclear business value and inadequate risk controls.** None has to do with how smart the model is. All three have to do with how the project was scoped. In [an earlier article](https://www.shimliapp.com/en/blog/from-pilot-to-production) we looked at why pilots do not reach production. This one is about what happens next: why projects that did reach production end up switched off, with three public cases and what each one teaches. ### Where companies stood at the start | Significant investment | 19% | | --- | --- | | Conservative investment | 42% | | Waiting or unsure | 31% | | No investment | 8% | Gartner poll of 3,412 webinar attendees, January 2025. Gartner, June 25, 2025 Most bet cautiously, and they were right. In the words of Gartner analyst Anushree Verma, many projects are experiments “driven by hype and often misapplied”. ### Reason 1: rising costs The first unexpected cost is paying for an agent that is not one. Gartner calls it **agent washing** : presenting a fixed-answer chatbot or an assistant that only chats as an agent. It estimates that of the thousands of vendors that call themselves agentic, **about 130 offer real agents** . A project that bought the second believing it was the first will never deliver the promised result, no matter how much it is tuned. The second is consumption with no ceiling. An agent that runs every message on the most expensive model, or that nobody measures per task, turns success into a problem: the more it is used, the more it costs, and nobody knows what each resolved case is worth. **How to cover it:** check that the agent carries out real tasks in your systems, not just that it chats well. And require the cost per task before you sign, so you budget with your volume and not the demo’s. ### Reason 2: value that is never proven **Klarna** is the most cited case. In February 2024 it announced that its AI assistant was doing the work of 700 customer service agents. In May 2025, its CEO acknowledged to Bloomberg that the approach had produced “lower quality” service, and the company started hiring people again for customer service. **Commonwealth Bank** , Australia’s largest bank, cut 45 phone service roles after launching a voice bot that, according to the bank, reduced calls by 2,000 a week. The union argued that volume was in fact rising, that overtime was being paid and that team leaders were answering calls. In August 2025 the bank reversed the decision and admitted its assessment had been an error. In both cases the technology worked. What failed was how value was measured: savings were measured and quality was not, and cuts came before the result was checked in the real operation. **How to cover it:** measure against a baseline. Before launch, record how many cases your team resolves today, in how much time and with what satisfaction. Then compare the same figures with the agent. And do not cut staff on a projection: do it, if appropriate, when production data backs it. ### Reason 3: risks without controls **Air Canada** is the textbook example. Its chatbot told a customer he could claim a bereavement refund after traveling, a policy the airline did not have. When the customer claimed it, Air Canada argued the chatbot was responsible for its own answers. The British Columbia tribunal did not accept it: on February 14, 2024 it ordered the airline to pay **C$812** and made clear that the company answers for what its agent says. The amount is small. The precedent is not: every answer from an agent is an answer from your company. **How to cover it:** the agent must answer from sources you approved, not from what the model thinks it knows. It needs explicit limits on what it can promise, a record of every conversation and a handoff to a person when the question goes beyond its scope. ### The project that gets canceled and the one that scales | Criterion | Project that gets canceled | Project that scales | | --- | --- | --- | | Initial scope | Automate customer service | One specific process, such as installment reminders | | Cost | Discovered on the invoice | Known per task before signing | | Value | Projected in a presentation | Measured against a baseline in production | | Agent answers | Whatever the model generates | From approved sources, with limits and a record | | Human team | Cut at launch | Reassigned when the data justifies it | Neither column depends on the AI model. They depend on decisions made in the project's first week. Five questions before approving an agent project - Does the agent carry out tasks in our systems or does it only chat? Ask to see it record something real, not a prepared demo. - How much does each resolved task cost, at our volume? If the vendor cannot work it out, you will not be able to budget it either. - What is our baseline today? Without it, six months from now nobody will be able to prove whether the project worked. - Which sources does the agent answer from, and what can it not promise? Get it in writing before launch. - When does a conversation go to a person? Define the cases; do not leave it to the model's judgment. [See how Shimli controls what each agent does →](https://www.shimliapp.com/en/seguridad) ### How Shimli handles it Shimli is designed so an agent project does not fall into any of the three reasons. **Known cost.** One plan at $139 a month with every feature, and AI usage at a fixed price per query, published on [pricing](https://www.shimliapp.com/en/precios) . FusionAI, our orchestrator, picks the right model for each task, so you do not pay for the most expensive model on a simple question. **Measurable value.** You start with one process, measure it on the live dashboard with open and closed conversations and response times, and scale to the rest when the numbers back it. **Risk under control.** The agent answers from the information you load, within the rules and limits you define, keeps a record of every conversation and hands over to your team when the case calls for it. You create it by chatting with Shimli’s assistant in under 10 minutes, and adjust it by hand whenever you want. ### Questions on this topic Why do AI agent projects get canceled? Gartner identifies three causes: costs that grow more than planned, business value that is never fully proven and inadequate risk controls. None of the three depends on the AI model; they depend on how the project is scoped, measured and controlled. What is agent washing? It is presenting as an AI agent something that is not one, such as a fixed-answer chatbot or an assistant that only chats. Gartner calls it agent washing and estimates that, of thousands of vendors that call themselves agentic, only about 130 offer agents able to decide and carry out tasks. How do I keep my agent project from being canceled? Start with a narrow process and measure it against your current baseline before scaling. Know the cost per task from day one. Define what the agent can decide, which sources it answers from and when it hands over to a person. And do not cut staff until the agent's production data justifies it. Useful to someone on your team? Share it with whoever decides on this process. [WhatsApp](https://wa.me/?text=4%20in%2010%20AI%20agent%20projects%20will%20be%20canceled%20https%3A%2F%2Fwww.shimliapp.com%2Fen%2Fblog%2Fwhy-agent-projects-get-canceled) [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fwww.shimliapp.com%2Fen%2Fblog%2Fwhy-agent-projects-get-canceled) [Email](mailto:?subject=4%20in%2010%20AI%20agent%20projects%20will%20be%20canceled&body=4%20in%2010%20AI%20agent%20projects%20will%20be%20canceled%20https%3A%2F%2Fwww.shimliapp.com%2Fen%2Fblog%2Fwhy-agent-projects-get-canceled) Copy link ### Keep reading [See all →](https://www.shimliapp.com/en/blog) - Platform #### The agent that sells inside the WhatsApp chat WhatsApp went from service channel to point of sale. What an AI agent needs to recommend, build the cart and get paid without leaving the chat. October 12, 2026 - Platform #### FINNOSUMMIT 2026: one in three sessions was about AI Instead of summarising impressions, I counted the sessions: 22 of the 67 had AI in the title. That says a lot about where banking in the region stands. September 25, 2026 - Platform #### The next AI advantage is: trust What the Volcano Innovation Summit 2026 left behind: when capability gets cheap, the advantage becomes trust. And trust has two sides. September 16, 2026 ### Put an agent to work on this process. Create the account and start today. Or we take 30 minutes and define which metric you'll move. [Create account](https://shimli.app/user/auth/create-account) Book a demo