MMOOZZGG.com

How to build an AI agent

AI agent failure rate: why projects fail and how to avoid it

5 min read4 sources

The AI agent failure rate is high, and most failures are set before the first line of code: in the choice of task and in the lack of a metric. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and MIT NANDA counted 95% of generative AI pilots with no measurable effect on profit. Below are three causes and what to do to join the successful 5%.

On this page · 5 sections

A team and a head of AI are needed

The research numbers

Gartner, 25 June 2025: more than 40% of agentic AI projects will be canceled by the end of 2027 because of cost, unclear business value and weak risk control. At the same time, by 2028 agents will be part of 33% of enterprise software programs: the technology is not dying, weak projects are being filtered out.

MIT NANDA, report “The GenAI Divide”, August 2025: 95% of generative AI pilots show no measurable effect on profit. The findings rest on 150 interviews with leaders, a survey of 350 employees and a review of 300 public deployments.

McKinsey, State of AI, November 2025: 88% of organizations use AI, 62% experiment with AI agents and 23% scale them. 39% see an effect on EBIT, which means profit before interest and taxes.

The gap between the 62% of companies that try agents and the 23% that scale them is the failure zone. The project was launched but never reached work across the whole company.

The three reports lead to one conclusion: trying AI is easy, turning it into money is hard. Most projects get lost between the experiment and scale.

Cause 1: the wrong task

MIT NANDA calls the main problem a learning gap: general tools like ChatGPT do not adapt to a company’s workflows and get stuck at the experiment stage. An agent placed on a process without its own data and rules meets the same fate.

Typical mistakes of choice: a task with few requests where the effect cannot be measured; a process that the company has never described, so everyone does it their own way; a task chosen because someone showed it at a conference.

The right task has a flow of repeated work, data for the agent and losses you can put in money. How to choose such tasks is covered in the article on why what to build matters more than how.

The test is simple: if the process owner cannot say how much the company loses on it each month, the task is not ready for an agent yet.

Cause 2: no metric

Gartner names unclear business value among the reasons for cancellation. In practice, this is a project where nobody wrote down before the start which number should change and by how much.

Without a metric, a pilot ends with an impression: “the agent seems to answer fairly well”. Six months later nobody can prove it paid off, and the project is closed at the first budget cut.

A metric for an agent is one number in your systems: response time, the share of requests handled without an operator, the share of leads entered in the CRM. It is measured before launch and written into the contract with the contractor.

Cost, another cause named by Gartner, is also tied to the metric. Without it, spending on the model and support has nothing to compare with, and the project looks expensive even when it pays off.

The third cause by Gartner, weak risk control, is closed in the same way before the start: an agent with limited rights, a log of actions and human approval of risky steps. There is more detail on the page on agent protection.

The general reasons why AI does not work in companies are collected on a separate page.

Cause 3: “agents” that are not agents

Among thousands of “agent” vendors, Gartner counted about 130 real ones. The rest renamed chatbots and RPA, which means programs that repeat a person’s actions by a fixed script.

An agent chooses the steps and tools for a task by itself; a chatbot follows a script written in advance. A company that bought a scripted bot sold as an agent gets the result of a bot and loses faith in agents in general.

What to ask the vendor: which systems the agent is connected to, which actions it takes without a person, what happens to a question outside the script, and which metric is written into the contract. If the answers are general, you are most likely looking at a bot with a new name.

The opposite mistake also happens. A standard agent is built from open blocks, and if you are asked to pay for it as a fully custom development, ask what the price consists of.

How to join the 5%

MIT NANDA writes that buying from specialized vendors and partnerships succeed in about 67% of cases, while in-house builds succeed a third as often. The experience of a team that has already brought agents to a result matters more than the choice of tool.

An order of steps that cuts off all three causes:

1. Choose the process by its losses in money and the data available, not by fashion.

2. Write down the metric and measure it before the start.

3. Launch a pilot on one process with a limited term and price.

4. Review the agent’s errors every week from the dialogue log.

5. Expand what produced a number, and close what did not.

If the project is already running or a contractor has sent an offer, it can be checked from outside. AI due diligence, a full check of the company and the project, shows whether the project has a task with money in it, a metric and a real agent.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Gartner, 25 June 2025 (gartner.com)
  2. Fortune via Yahoo Finance, 18 August 2025 (finance.yahoo.com)
  3. IT Brief, 11 November 2025 (itbrief.co.uk)
  4. Fortune, 18 August 2025 (fortune.com)

Found a mistake? Tell us

Publications

Contact us