Industry surveys consistently put the share of machine learning projects that never reach production somewhere above half. Our experience with AI projects in Jordan matches that, and in almost every case the model was not the reason.
The failures cluster into four patterns. All four are visible before you start, if you look, and each one applies to AI projects in Jordan regardless of sector or budget.
Pattern one: the data was not ready
This is the most common by a wide margin. A pilot runs on a clean extract that someone prepared by hand. It works. Then production hits the actual data — inconsistent formats, missing fields, duplicate customer records, three spellings of the same company name — and accuracy collapses.
Arabic text compounds this. Names transliterate several ways, addresses are descriptive, and the same entity appears under variants that a naive join treats as different records.
What to do
Audit data quality before scoping the model, not after. If entity resolution is going to be needed, it is a project in its own right — budget for it separately rather than discovering it mid-build.
Pattern two: nobody owned the outcome
An AI project sponsored by IT but affecting operations, with no operations owner, will stall at the point where someone has to change how their team works.
The technical delivery finishes. The system sits unused because the people it was built for were never asked what would make them use it.
What to do
Name a business owner with authority before writing code. Their job is to decide what changes in the process — a decision no vendor can make for them.
Pattern three: there was no deployment path
A model in a notebook is a demonstration. A model in production needs an inference endpoint, monitoring, a retraining schedule, version control, a rollback plan and someone on call. Teams routinely budget for the first and not the rest.
What to do
Decide the deployment target in week one. If it has to run inside an existing application, that constrains the model — sometimes ruling out an approach that would have scored better in isolation. Better to know at the start.
Pattern four: success was never defined
"Improve customer experience" cannot be evaluated. Six months in, nobody can say whether the project worked, so it quietly ends.
What to do
Write down one number, its current value, and the target. "Reduce average handling time from 8.5 minutes to under 6" is a project. Anything vaguer is a research exercise, which is fine — but call it that and fund it accordingly.
The best predictor of an AI project reaching production is not model sophistication or budget. It is whether someone can state, in one sentence, what changes operationally when it ships.
Where AI genuinely pays off here
Based on what we have delivered in the Jordanian market, the reliable wins are narrower than the marketing suggests:
- Document processing. Extracting structured data from invoices, contracts and forms — including Arabic scans — has a clear before-and-after and immediate labour savings.
- Demand forecasting. Given eighteen months of clean sales history, forecasting reliably beats manual planning and the savings are measurable in inventory.
- Anomaly detection. Fraud, quality defects and equipment failure all suit models well because the cost of a miss is quantifiable.
- Customer support triage. Routing and drafting responses, with a human approving before send, cuts handling time without the risk of an unsupervised model talking to customers.
Where it usually does not, yet
Fully autonomous decisions in regulated contexts, anything requiring a training dataset your organisation has not collected, and use cases where the process changes faster than a model can be retrained. Being honest about these early saves everyone a wasted quarter.
A sensible sequence
- Pick one measurable problem with an owner who wants it solved.
- Audit the data that problem depends on. Fix it or pick a different problem.
- Establish the deployment path before building the model.
- Ship something narrow that works, then widen the scope.
Projects that follow this sequence reach production. Projects that start with the model and work backwards usually do not.
