You have probably seen an AI demo that worked. A model drafted a reply to a customer, forecast next month's demand or answered a policy question, and the room was convinced. Months later, the same pilot is still a pilot. Daily work runs the way it did before.
This is not rare. In the GCC, 84% of organisations surveyed by McKinsey use AI in at least one function, but only 31% are scaling AI or have fully scaled it[1]. The pattern holds elsewhere too. In Gartner's survey, only 48% of AI projects reached production, and the move from prototype took 8 months on average[2].
31%
of GCC organisations surveyed are scaling AI or have fully scaled it, while 84% use it in at least one function
Source: McKinsey, The state of AI in GCC countries, Nov 2025
A pilot and an operation ask different questions
A pilot is built to prove that a model can do one task. It usually succeeds, because the task was chosen to succeed and the data was prepared by hand.
Operations asks something else. Can this run every day, on live data, inside the systems people already use, with someone accountable when it gets something wrong? A pilot rarely has an answer, because nobody was asked to design one. The stall tends to happen in four places.
1. The data was curated, not connected
The pilot ran on a clean export. Production means reading from the ERP, the CRM or a shared inbox, with all their gaps, duplicates and formats. That integration work is often bigger than the model work, and it was in neither the pilot plan nor the pilot budget. The first time the system meets real data is also the first time anyone learns how good it is.
2. No one accountable after the demo
The innovation team built it, IT has to run it, and the business has to use it. Each has a fair reason to wait for the others. When no single person is accountable for the outcome, the pilot waits for a decision that never comes. A name on the result, written down before the build, does more than another steering committee.
3. Compliance arrives late
Data residency, personal data and cybersecurity controls get reviewed after the design is fixed. The review finds a problem: the model is hosted abroad, the logs cannot support a breach report, or the data was never classified. The redesign takes longer than the pilot did. In the Kingdom, where PDPL, NCA controls and SDAIA's principles apply from the first workshop, a late review is the most expensive kind.
4. The work around it never changed
A model that drafts a reply saves time only if the process expects a draft. If approvals, handoffs and forms stay the same, people do the old work plus one new step. Usage drops after the first month, and the pilot is remembered as a tool nobody wanted.
None of these is a model problem
Each is a design decision that has to be made before the build starts: which systems the solution reads from and writes to, who owns the result, where the data lives, and which steps in today's process disappear. Those decisions need people who understand the operation and the technology at the same time. They are cheap to make in week one and costly to make after the demo.
A pilot is not progress. A system in operations is.
What it looks like when it reaches operations
A system that has reached operations looks different from a pilot. It runs on real data, inside the systems people already use. It has a named person accountable, a measure agreed before launch, and a clear path for the cases the model is unsure of. Its results are counted in daily work, not in a demo.
If you have a pilot that has stalled, answer the five questions honestly. The gaps they expose are your plan for getting it live.
Ranking pilots on value and getting the chosen ones into daily work is the core of two of our practices: AI strategy, investment and value realisation, and operations and service re-engineering with AI.
Sources (2)
Figures are quoted as published. Forecasts and drafts are labelled.