95% of AI Pilots Stall at the Door: The Three Hurdles of the Last Mile

“95% of enterprise AI pilots never make it to production.” The figure, cited repeatedly at a Salesforce forum during Alibaba Cloud's Apsara Conference, comes from a 2025 MIT study: after surveying 300-plus enterprise generative-AI projects, the team found only about 5% of customized AI tools actually reached production. Model capability keeps sprinting; the last mile keeps collapsing.

Hurdle One: Data in Three Silos

A salesperson wants AI to judge whether an account is worth pursuing: the spreadsheet says “interested,” the chat log just logged “budget on hold,” and a service ticket sits unresolved. If the AI reads only one of these, even strong reasoning yields the wrong answer. Fragmented data is where nearly every pilot begins to fail.

Hurdle Two: Good at Advising, Bad at Doing

An ordinary AI tool drafts a plausible quote — but without access to business systems (fetch current pricing, create a formal quote, submit discount approval, write results back to the customer record), it stalls. Yet those mundane steps are exactly the work enterprises need AI to take over. The fix: expose business capabilities to agents via APIs and MCP tools, so AI stops standing on the sidelines offering advice and rolls up its sleeves.

Hurdle Three: Permissions Can't Live in Prompts

With an execution channel, the AI knows “how”; but who may command it and how far it may go must be defined by a permissions system.

A salesperson may pursue an account yet lack access to sensitive fields; a regional lead manages a region, not global records. Such rules must be enforced by the system — not by whispering “keep it confidential” into a prompt. In low-fault-tolerance industries like life sciences, where a rep's visible data is strictly regulated, permissions-by-role lets AI prepare materials within bounds and route professional questions back to medical staff.

Methodology Beats Model Size

Connect the three hurdles and they corroborate two cases covered in this column: CCB's “AI trader” runs the same split of “LLM semantic understanding — rule-based validation — human supervision,” while China's government-LLM national standard hard-codes safety compliance. The mature path of industry AI is converging: unify data first, then tools, and put permissions and audit before everything else.

Permissions-and-audit-first is likewise the precondition Jiutian enforces in government and enterprise agent deployments — an AI may start work a little later, but never start work unfit. When 95% of pilots die at the door, the surviving 5% are not the ones with bigger models, but with cleaner engineering.

(Compiled from QbitAI, MIT's 2025 enterprise GenAI research and EY reports)