Many experts are offering a range of opinions on why companies run many AI pilot projects but fail to confirm results on the business side.

Daniel Conum (대니얼 코넘), who leads the forward deployed engineer organisation at AI agent company Varick, drew attention with a recent post he shared on social media platform X (Twitter). He pointed out that people often blame failures on data or models, or vaguely on change management, but that these are usually not the root causes.

Varick supports companies with more than $1 billion in revenue, from AI strategy and workflow redesign to building and operating agents.

Conum cited seven reasons AI pilots fail, and model selection is not on the list.

According to him, the first reason AI pilot projects do not lead to results is that companies set scope based only on standard operating procedures, without a process map that reflects the real workflow.

He gave as an example the task of registering a new customer in a company system. The manual says it takes five days, but in reality it takes 22 days.

Conum said, "Documents come in with pieces missing, bank account information comes in different formats, and 1 out of 10 cases is a company that was already registered and comes back in with only a slightly changed name. These exceptions make up more than 40 percent of the total. That is why an AI built only from the manual cannot hold up in the field."

The second reason is building only one standout step seen in an AI demo, rather than the full workflow. Even if AI reads tax forms (W-9) and cuts eight minutes, a 22-day job still takes 22 days. Conum said, "Most of the 22 days is not working time, but time spent waiting for processing order and checks by other companies and other departments."

The third is the lack of a baseline. If processing time, per-case cost and error rate before adoption are not specified, there is no answer when the finance team asks about return on investment after four months. Conum said, "The finance team is not interested in a survey saying users feel 30 percent more productive."

The fourth is building an assistant such as a chatbot or copilot that only moves when a person calls it. According to him, such tools are used by about 10 percent of employees and lift efficiency only 10 to 20 percent. Conum said, "A background agent that works without being told and hands only exceptions to people can improve 60 to 90 percent."

The fifth is building a separate app outside the systems employees use every day. Conum said, "No one uses a tool that requires going to a new address and starting from logging in. Agents need to work inside systems already in use, such as SAP, ServiceNow and Outlook. If it runs on the company's cloud and data, approvals from security, legal and IT departments also come faster."

The sixth is having no owner in the business operation. When the IT department builds an agent and hands it off, there is no one to fix it when problems arise. Conum said, "The head of the department responsible for the performance of that work needs to own the agent, and technical staff who can fix it when problems occur should be kept nearby."

The seventh is having no plan after month seven. Pilots usually begin with a six-month innovation budget. When the budget is cut off, there is no answer to who pays the AI usage fees, or what happens if a model is discontinued or its price rises. Conum said, "From the start, costs should be covered from the operating budget of the relevant department, and it should be built so it is not locked into and dependent on a specific AI model."

He cited conflicting interests as a reason failures are repeated in AI projects. Software vendors are focused on increasing the number of users, consulting firms on writing reports, and internal teams on showing something quickly, and have less interest in actual results, he said.

Conum stressed, "After solving all seven, the partner registration period fell from 22 days to 4 to 5 days, and cases requiring human judgement were around 15 percent. Because there was a baseline, the CFO could confirm the results. That is the way to secure the budget to use for the next 10 tasks."