Executives blame cost and risk for scrapped AI pilots while practitioners blame the data

How many pilots fail is the settled part of this record; why they fail is where the sources split. The column carries two answers, and they lead to different budget lines. Executives in S&P Global Market Intelligence's survey named cost, data privacy and security risk as their top reasons for walking away ^[6]. Gartner's forecast cited rising costs, unclear business value and weak risk controls ^[8]. MIT's report put the problem elsewhere, describing a gap in how organizations integrate the tools into existing workflows, separate from the quality of the underlying models ^[3].

Those diagnoses do not buy the same thing. Cost and unclear value argue for narrower scope and a harder value case before the next pilot is funded. A data blocker argues for spending a quarter on taxonomy and record cleanup before an agent is licensed at all. Jain, a senior product manager working on AI-driven CRM strategy for telecom clients via Mphasis ^[1], wrote that the work usually gets folded into data prep and scheduled for two weeks, when the honest estimate is closer to a quarter ^[11]. Counting a quarter as thirteen weeks, that is about six and a half times the schedule the roadmap carries ^[2].

His evidence is one project. At Amazon he co-built a clause extraction system covering a $40 billion annual supply chain contract portfolio, which eventually reached 95% accuracy ^[12]. He attributes most of the gain to months spent building a structured taxonomy of clause types. Before that taxonomy existed, he wrote, the same model produced results nobody could use, because "payment terms" meant six different things depending on which template a contract used ^[13].

The figure a buyer would most want to lean on is also the softest one here. Among the pilots that succeeded, purchased and integrated tools worked about 67% of the time against about 33% for tools built in house ^[4], roughly two to one ^[1]. As published, the comparison measures success rates among successes, so the direction is more usable than the level.

PwC's numbers show where the caution sits. Asked which tasks they would trust an agent with, 20% of executives said financial transactions and 38% said data analysis ^[10]. That ordering matches the quote-to-cash environment Jain describes, where pricing exceptions get approved over email and never make it back into the system that is supposed to be the record of truth ^[14].

On dates, these numbers are a year stale. The S&P survey, the Gartner forecast, the PwC poll and BCG's September 2025 Build for the Future study of more than 1,250 companies all predate the column, which Forbes ran on September 14, 2026 ^[17][18][3]. None of these surveys covers 2026.

Jain screens vendors with two questions: what the agent does without a human reviewing each step, and what happens when it meets a case it has not seen ^[15]. "Most of the difference between a real agent and a relabeled chatbot shows up in the answer to that second question," he wrote ^[16].