AI that does the work, not the demo.

Ordinary automation follows a fixed path: if this field changes, do that. An agent handles the cases where the input is unstructured or the next step depends on judgement — reading a message and deciding whether it is a booking, a complaint or a supplier invoice, then taking the right action for each. Where a rule is enough, we use a rule: it is cheaper, faster and it cannot be wrong in a new way.

No. Your documents and conversations are used to answer questions at the moment they are asked, through retrieval, and are not used as training data by us. We configure provider accounts with training and retention disabled, and the commitment is written into the contract rather than left to a settings page nobody re-checks.

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 one automated workflow typically runs 4–8 weeks from scoping to launch. A multi-step agent platform with several integrations and role-based permissions runs 10–20 weeks. The first two weeks are discovery and specification, and you can stop at the end of them with a document you own.

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.

That is usually the majority of the project. We integrate with CRMs, ERPs, accounting tools, WhatsApp Business, email, calendars, databases and internal portals — anything with an API, and several things without one. Integration requirements are documented during scoping and tested in a staging environment before anything touches live data.

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 before building — how long the process takes today, how many cases it handles, where it fails — and report against it after launch. Every run is logged with its cost and outcome, so the question is answered from your own data rather than by impression. If the numbers do not move, that is a finding we report, not one we manage around.

We scope what data the system may see before we build, keep personal data out of prompts where the job does not need it, and log access. Federal Decree-Law 45 of 2021 applies the same way it does to any other system processing personal data — lawful basis, purpose limitation, retention, and the ability to answer a subject access request. We build to that and document it; we are engineers, not your legal advisers, and a regulated business should have counsel review the design.

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.