Why do most AI pilots never reach production?
Quick answer, Nov 3, 2026As of 3 November 2026, the published research points at one cause: pilots are scoped to be shown, not to be run. Tools get evaluated far more often than they get deployed, budgets follow the functions that are easiest to measure, and projects built with an outside partner reach production about twice as often as ones built in-house.
What the published numbers say
| Finding | Figure | Source |
|---|---|---|
| The drop-off for enterprise-grade tools | 60 percent of organizations evaluated one, 20 percent reached a pilot, and 5 percent reached production | The GenAI Divide, MIT NANDA, July 2025 |
| What that spend returned | 95 percent of organizations getting zero return, against $30 to $40 billion of enterprise GenAI investment | MIT NANDA |
| Who built the ones that shipped | External partnerships reached deployment about 67 percent of the time, against about 33 percent for tools built fully in-house | MIT NANDA |
| How far companies have actually got | 25 percent of leaders have moved 40 percent or more of their AI pilots into production, and 54 percent expect to reach that in the next three to six months | State of AI in the Enterprise, Deloitte, 21 January 2026 |
| How deep the use goes | 37 percent report using AI only at a surface level, with little change to the underlying processes; 30 percent are redesigning key processes around it | Deloitte |
| Who can say who owns an agent | Close to three quarters of companies plan to deploy agentic AI within two years; of those, 21 percent report a mature model for agent governance | Deloitte |
Both sources were read on 3 November 2026. The MIT NANDA figures come from a review of more than 300 publicly disclosed AI initiatives, 52 structured interviews and 153 survey responses, gathered January to June 2025. The Deloitte figures come from 3,235 business and IT leaders across 24 countries, surveyed in August and September 2025.
The pilot is scored on the wrong thing
A pilot is judged on whether it works in a demonstration. Production asks a different question: does it survive the real work, with the real data, every day, when the person who built it is busy with something else.
MIT NANDA puts the failures down to "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations", not to model quality, regulation or talent. The same report notes that generic chat tools like ChatGPT and Copilot show much higher pilot-to-implementation rates, around 83 percent, because they ask nothing of the process around them. The tools that would change a margin are the ones that stall.
The money goes where the result is easiest to attribute
The research describes a bias toward visible, top-line functions over back-office work that returns more. The report's own sample puts the clearest savings in the back office: eliminating business-process-outsourcing contracts worth $2 to $10 million a year in customer service and document processing, a 30 percent cut in external creative and content spend, and $1 million a year on outsourced risk management in financial services. Those are documented outcomes at specific organizations in that sample, not a standard to expect.
It also records that the gains arrived without material workforce reduction. The money came back from reduced external spend, not from headcount.
What to settle before anyone writes code
- A baseline for the process. Hours, cost, volume and errors, measured before the build. Without it there is no way to say later whether the thing worked.
- How you bill for the work. On a contract billed per user, device or month, hours taken off the delivery stay with you. On hourly work the saving goes to the customer unless the contract changes.
- Which process, ranked by margin. Pick by what the change does to the economics, not by what demonstrates well.
- Who runs it afterwards. Name the person and write the runbook as part of the build, not after it.
- Who signs off on it going live. Deloitte finds a mature agent-governance model at 21 percent of the companies planning to deploy agentic AI. Decide the review before you need it, not during an incident.
What Frontier Desk does
Our forward-deployed engineers work inside the company and automate specific processes, such as request triage, document review, scheduling, billing and reporting, starting from a measured baseline and ending with the result measured against it, and runbooks your own team runs. Forward-deployed engineers: what the work is sets out the steps and how they are measured.
If a sale, recap or loan is anywhere in view, the order matters more. Buyers and lenders in the Cybernetic Age price what is running, not what was demonstrated: what buyers and lenders check is the list to build against.
Questions
Does the 95 percent figure mean AI does not work? No. The same report says the other 5 percent are extracting millions, and separates widely adopted chat tools from the custom and vendor-sold systems being rejected. The split it describes is about approach, not technology.
So should we hire a partner rather than build it ourselves? The 67 against 33 percent gap is real but the report states its own limits plainly: the figures are self-reported, drawn from an interview sample of 52 organizations, and "the correlation between external partnerships and success does not necessarily prove causation." Firms that partner may simply be the ones that were readier.
We already ran a pilot that stalled. Is the work wasted? Usually not. What a stalled pilot gives you is the measurement nobody made at the start. Re-scope it around one process with a baseline and an owner, and the second attempt starts from evidence.