AI development for a defined business task.

Rules follow defined conditions; an AI component can interpret unstructured input where fixed rules are insufficient. Both need tests, error handling and clear limits. We use rules where they fit and restrict model-driven actions according to the cost of a mistake.

We document which data is sent to which provider and for what purpose. Training use, retention, regional processing and deletion depend on the chosen service, account terms and available settings; they must be verified before sensitive data is used and recorded in the agreement. Retrieval alone is not proof of zero retention or zero training use by a provider.

We design for it, because every such system is wrong sometimes. Each workflow has a confidence threshold, an escalation path to a named person, and a log of what the system saw and why it acted. Anything expensive or irreversible — sending a quote, issuing a refund, contacting a client — sits behind a human approval step by default, and only moves out of it once the record justifies the change.

A single assistant or workflow is typically estimated at 4–8 weeks; a multi-step platform at 10–20 weeks. Discovery, access, evaluation and approval dependencies determine the final schedule. Any standalone discovery deliverable or stop point is agreed in the proposal.

A single assistant or automated workflow is typically AED 15,000–40,000. A multi-step agent platform across several systems ranges from AED 45,000 to AED 100,000+. Running it afterwards — monitoring, evaluation and tuning — starts at AED 3,000 per month. Model and API usage is billed by the provider directly to your account, so you can see exactly what it costs to run.

Whichever fits the task, the budget and your data requirements — most work uses commercial models from Anthropic, OpenAI or Google, and some cases justify an open-weight model on infrastructure you control. We build so the model is a replaceable part: a better one ships every few months, and a system that can only work with one provider has an expiry date.

Yes, and it is tested in Arabic before launch rather than translated afterwards. That means checking dialect and mixed Arabic-English messages the way customers actually write them, right-to-left output that renders correctly wherever the answer is displayed, and a set of real Arabic test cases the system has to pass on every release.

We review the systems, APIs, access permissions and data flows involved before committing to an integration. Where there is no supported API, feasibility and maintenance risks require separate assessment. Test environments and safe deployment steps are agreed before live access.

One process, one person who genuinely knows how it works today, and access to a sample of real examples — messages, documents, records. The examples matter most: they are what a system is built against and what it is evaluated on, and a project scoped on how a process is supposed to work rather than how it actually works is the one that disappoints.

We agree a baseline and success criteria before building, such as handling time, reviewed accuracy or completion rate. Monitoring records the minimum operational data needed to evaluate cost and outcomes, with agreed access and retention. Reports distinguish measured changes from estimates and identify inconclusive results.

We review the data needed for the task, access permissions, provider terms, retention and escalation requirements. Applicable legal obligations and any compliance commitments must be assessed with the client’s legal or privacy adviser. Technical controls and documentation do not by themselves certify compliance.

Often, yes. We start with a short review of what exists — prompts, integrations, evaluation, and where the costs are going — and tell you honestly whether it is worth continuing or cheaper to rebuild the parts that matter. Both answers happen; we would rather say the second one early than bill for the first.

Frequently not, and it is a cheaper conversation to have before the invoice. A well-built form, a proper database or a scheduled report solves a lot of what gets pitched as an AI problem, at a fraction of the price and with no ongoing model cost. If that is the honest answer for your process, that is the one you will get, and we will quote for that instead.