10 mistakes
The mistakes follow the order in which people make them: first when choosing the task, then when building, then after launch. The first three cost more than the rest, because setup cannot fix them: the agent has to be rebuilt or switched off.
1. No metric. The agent is launched without a written number it must change: response time, requests handled, staff hours. A month later, there is nothing to prove its value with.
2. The wrong task. The agent is put on a process that rarely repeats or where an error costs a lot. Start with frequent work that has a clear right answer.
3. Extra complexity. A team of several agents is built at once where one request to a model would do. Anthropic advises starting with a simple solution and adding complexity only when needed, and sometimes no agent is needed at all.
4. No answer checks. The agent goes out to customers without a set of test questions. On 14 February 2024, a tribunal in the Canadian province of British Columbia held Air Canada liable for a wrong answer by the chatbot on its website and rejected the argument that the bot is responsible for its own words.
5. Access to the whole CRM. The agent gets administrator rights, although it only needs to read one sales funnel. OWASP, a software security community, calls this excessive agency: extra functions, extra rights, and extra autonomy of the agent.
6. Actions without a person. The agent sends emails, changes prices, or issues refunds without confirmation. OWASP advises that a person confirm important actions.
7. Data in a foreign service. On 2 May 2023, Samsung banned generative AI for its employees after engineers uploaded internal source code to ChatGPT. For customers’ personal data in Russia, 152-FZ, the Russian personal data law, applies: storage in Russia and a notice to Roskomnadzor, the data protection regulator, before any transfer abroad.
8. Nobody told the agent what it does not know. The instructions have no rule like “if the answer is not in the documents, say so”, and the model makes up an answer. The reasons are covered in the article on why an AI agent gives wrong answers.
9. No owner. A vendor launched the agent, and inside the company nobody updates the documents or reads the log. A couple of months later, the agent answers from an old price list.
10. Paying for hours, not for results. A typical agent is built from open components; if you are asked to pay for it as for unique development, ask what the price is made of. Pay for the result by the metric, not for hours.
Why projects get canceled
According to a Gartner forecast of 25 June 2025, more than 40% of agentic AI projects will be canceled by the end of 2027. It names three reasons: cost, unclear value, and weak risk control.
In the same forecast, Gartner estimated that of the thousands of “agent” vendors, only about 130 are real. The rest renamed chatbots and RPA, programs that repeat a person’s actions by a fixed script.
In August 2025, MIT NANDA found that 95% of generative AI pilots have no measurable effect on profit. Buying from specialized vendors and partnerships succeed in about 67% of cases, and in-house builds a third as often.
In all three numbers, the cause is not the code. A project is canceled when nobody decided in advance which task the agent solves, what that task costs the company, and who will accept the result.
For a company of any size, this gives a simple rule: launch each agent as a pilot with a metric and a review date. If the number has not moved by that date, switch the agent off and move the money to the next process on the list.
The causes are covered in detail on the page on why AI agent projects fail.
How to check yourself before the start
Before the start, answer seven questions in writing. If one of them has no answer, it is too early to begin.
1. Which number will the agent change, and what is its value now?
2. How many times a month does the task repeat, and what does an error in it cost?
3. Can the task be solved more simply: with a template, a rule in the CRM, or one request to a model?
4. Which data does the agent need, and where will it be stored?
5. Which rights does the agent need, and which actions does a person confirm?
6. On which questions will you test the answers before launch, and what share of correct answers do you consider enough?
7. Who in the company owns the agent after launch, and how often do they read the log?
Questions 3 and 5 repeat the advice of Anthropic and OWASP: a simple solution before a complex one, minimum rights, and a person before an important action. If you have more questions about the choice of task than answers, start not with the build but with a list of processes where the company loses money.