AI implementation in Qatar, from pilot to production

The uncomfortable statistic in this field is how many AI projects never reach production. The reason is rarely that the technology could not do it. It is that nobody defined what "working" meant, the data was worse than anyone admitted, or the system was built beside the business rather than inside it and quietly stopped being used.

Implementation is the discipline of avoiding those three outcomes. It is unglamorous and it is the entire difference between a pilot and a system that is still running in two years.

The phases, and how long each really takes

For a first project at a mid-sized Qatari business. Larger programmes are these phases repeated, not a different shape.

• Define one task and its measure — one to two weeks. Not "improve customer service" but "answer these forty recurring questions without a person, and here is how we will count it"
• Data assessment — one to three weeks, and the phase most often skipped. Look at the actual records, not the schema. This is where projects are honestly cancelled, which is a good outcome compared with the alternative
• Build and integrate — three to eight weeks, of which the model work is a fraction. Most of it is connecting to the ERP, the mailbox, the WhatsApp number and whatever else holds the truth
• Supervised running — three to four weeks with a human checking every output. Non-negotiable. This is how you find out what it gets wrong before it matters
• Handover — the client can see what the system did, change its instructions, and switch it off. If they cannot do all three, it is not finished

Data readiness, which decides everything else

Language models are unusually good at producing a confident answer from poor data. That makes bad data more dangerous in an AI project than in a conventional one, because the failure is silent — you get a fluent, well-formatted, wrong figure, and no error message.

So the assessment is blunt: are the records complete enough, consistent enough and recent enough to answer the question being asked? If your item list has the same product under four names, no amount of model quality fixes the stock forecast. Cleaning that first is cheaper than discovering it in production, and if the answer is no, the correct advice is to fix the system of record before spending anything on AI.

Integrating with what you already run

An AI system that lives in its own interface gets used for a fortnight. The ones that survive appear inside the tools people are already in — a suggested reply in the mailbox, a draft invoice in the ERP, an answer in the WhatsApp thread the customer already opened.

Practically this means the integration work is the project. Reading and writing to ERPNext or Odoo through their APIs, handling the WhatsApp Business account properly, respecting user permissions so the assistant cannot show a salesperson the payroll, and logging every action so there is a record of what the system did on whose behalf.

Why AI projects fail here, specifically

Five failure modes account for most of what we are asked to rescue.

• No measure — the project was never given a number to hit, so it can neither succeed nor be stopped, and it drifts until the budget ends
• Data nobody checked — the pilot ran on a clean sample and production ran on ten years of inconsistent records
• Too much autonomy too early — the system was allowed to send, post or pay before anyone had watched it work for a month
• Arabic tested only in translation — it worked on translated English test cases and failed on how customers actually write
• One person owned it — the person who understood the system left, and nobody else could change a prompt or read a log

Governance, permissions and accountability

Two questions have to be answered before anything goes live: what is this system allowed to do without a human, and who is accountable when it is wrong. Neither is a technical question, and neither should be decided by the implementer alone.

The workable pattern is a narrow permission that widens with evidence. Draft, do not send. Propose, do not post. Read the ledger, do not write to it. Every action logged with what the system saw and what it decided, so a disputed outcome can be examined rather than argued about. Widening those permissions is then a deliberate decision made on a record of behaviour, which is also what an auditor will want to see.

How the cost works

Three components, and it is worth insisting a proposal separates them. Implementation is one-off project work. Model usage is a running cost that scales with volume and belongs to you, not to us — you should hold the account. Support and change are ongoing, because the business changes and the instructions have to change with it.

Be wary of a single monthly figure covering all three, because it hides which part is growing. And be wary of any proposal that does not begin with a small paid assessment: a firm willing to scope a full AI project without looking at your data has decided what to sell you before understanding the problem.

At a glance