AI4 min read
AI implementation: where the budget actually goes
Where an AI budget actually goes, and why the model is the cheapest part of it. The four cost buckets, the token arithmetic that makes run cost knowable before you sign, why Arabic costs more per word, and where UAE data residency changes the design.
Danial AhmedFounder & engineering manager
On this page
Every AI project has two numbers, and buyers ask about only one of them. The build number is the one in the proposal; the run number is the one that decides whether the system is still switched on in a year. Here is how both are constructed, using the same estimating model we apply on AI implementation work.
Four cost buckets, not one
Every AI budget decomposes the same way. The proportions surprise people: the model is the cheapest part.
- Discovery and evaluation design buys a defined task, a labelled test set and an acceptance threshold.
- Data plumbing buys access, cleaning, chunking, permissions and a refresh pipeline.
- Application build buys the interface, the orchestration, the tool calls, the fallback paths and the logging.
- Model work — prompting, retrieval tuning, occasionally fine-tuning — is the smallest line of the four.
Data plumbing and the application build are where most of the money goes. If a proposal spends most of its budget on the last item instead, it is a demo. Demos take a fortnight and impress a steering committee; they fail in production because nobody defined what "correct" means.
The shapes of engagement that recur
Four keep coming back in the UAE market, in increasing order of scope: a document extraction pilot over invoices, trade licences and LPOs; a WhatsApp or web support assistant answering from your own content; an internal retrieval assistant over policies and contracts; and a production agent that takes actions in your systems. Each is quoted on its own scope, and the first is where most companies should start.
Run cost is arithmetic, so do the arithmetic
Inference cost is knowable in advance, and almost nobody calculates it before signing. The formula is: conversations per month × turns per conversation × (context tokens in + tokens out) × price per token.
Put your own figures into it before you sign anything: expected conversations a month, the turns a real conversation takes, the context you actually retrieve per turn, and the provider's published price per million tokens. Add vector storage, logging and monitoring on top. The answer is a monthly number, and it belongs in the proposal beside the build price.
Two levers move that number by an order of magnitude: how much context you retrieve per turn — most systems retrieve far more than they need — and which model tier handles which turn. Routing simple turns to a smaller model is usually a large saving with no measurable quality loss.
Arabic costs more per word, and it is not a rounding error
Arabic text consumes substantially more tokens than the equivalent English — commonly two to three times, depending on the tokenizer — because Arabic morphology fragments badly into subword units. A bilingual assistant is therefore not the same cost per conversation in both languages.
Budget Arabic conversations accordingly, test retrieval quality separately in each language, and never assume that a system evaluated in English works in Arabic. Diacritics, dialect versus Modern Standard, and mixed-script customer messages all degrade retrieval in ways an English test set will never reveal.
Data residency and PDPL are engineering constraints, not paperwork
Prompts frequently contain personal data — customer names, Emirates ID numbers, medical context, contract terms. That makes the model provider a processor, with the transfer, contracting and retention obligations that follow under the UAE's data protection regime.
Three practical positions, in increasing order of cost. Use a hosted API with a signed agreement, zero-retention configuration and personal data stripped before the call. Or run an open-weights model — the Falcon family from Abu Dhabi's Technology Innovation Institute is a credible starting point — on infrastructure in AWS me-central-1 or Azure UAE North. Or keep it on-premises entirely, which raises the run cost sharply and is warranted only where the data genuinely cannot leave the building. See the PDPL checklist for the documentation side.
What we would fund first
A single, narrow, high-volume task with a measurable current cost — invoice extraction, first-line support triage, tender document search. A fixed scope, a labelled test set, a monthly run-cost report, and a genuine decision point at the end. Projects scoped that way either prove themselves in one quarter or stop cheaply, which is the only responsible way to spend a first AI budget.