Self-directed concept · Business setup · UAE
A recommendation engine that admits when it isn't sure.
A decision tool for people setting up a company in the UAE. Founders are asked to choose between Mainland and Free Zone before they understand what either means, and most tools answer confidently either way.
- Engagement
- Self-directed, about two weeks
- Role
- Logic, interface and copy. Working alone.
- Grounded in
- Public documentation on UAE structures, licences and first-year costs, plus how existing formation platforms frame the choice
- Disclosure
- Arnifi is used as a worked example, not a client. This was not commissioned by or affiliated with them.
- Evidence
- No founders have used this. Question order and cost bands are reasoned from public sources, not validated.
At a glance
Founders are asked to choose between Mainland and Free Zone before they understand what those options mean or which one fits. Existing tools return one confident answer every time, which looks decisive and is sometimes wrong.
This engine separates what the law decides from what is genuinely a judgment call. Where a legal rule fires, it closes the question. Where the signals conflict, it refuses to average them into false certainty, names the single variable that decides it, and says so.
Four decisions, each one depending on the others
Setting up in the UAE means resolving four questions at once, and none of them can be answered cleanly on its own.
Mainland or Free Zone
The structural choice everything else follows from, asked first, when the founder has the least information to answer it.
Licence type
Named in jargon the founder does not recognise. Professional, Commercial, Industrial mean nothing until somebody translates them into what the business actually does.
First-year cost
Quoted as a promise when it is really a range, and the gap between the two is where people get hurt.
Timeline to operational
Rarely stated honestly up front, and it is often the constraint that matters most.
The same confusion has a business cost. Longer advisor calls spent explaining basics, and lower-quality leads from people who abandon before they understand their options. A tool that removes the confusion earlier makes the call shorter and the lead better.
Never ask a founder a question they cannot answer
The tool asks what they already know: what the business does, who it sells to, the budget, the timeline. It does the translating for them. Every recommendation carries its reasoning: why this structure, what it trades off, and what could change the answer.
One distinction shaped the whole engine. Some inputs are hard constraints that legally determine the structure, like needing to invoice UAE clients directly. Others are preferences that shape the recommendation without changing the legal outcome, like budget or timeline.
Separating the two lets the tool be decisive where the law is decisive, and honest where it is genuinely a judgment call.
Four constraints, and where each one bites
| Constraint | What it meant | The decision it forced |
|---|---|---|
| Rules vary by emirate and free zone | There is no single correct answer across jurisdictions. | The tool indicates a direction with honest ranges and never concludes for a specific jurisdiction. Every figure is labelled indicative. |
| Users arrive with very different knowledge | Some founders know the vocabulary. Most do not. | Copy works for the least-informed founder without patronising the rest. Licence types are shown as tags next to plain-language activities, not as the question itself. |
| A recommendation cannot replace legal consultation | The product has a hard ceiling on what it is allowed to settle. | The design hands off honestly rather than pretending to be final. The advisor call is a designed destination, not an upsell bolted on the end. |
| Under about four minutes | Every additional question trades directly against completion. | Three steps, two questions each. A question earns its place only if it materially changes the output. |
The user whose case has no clean answer
Ahmed, technology consultant. He plans to serve both UAE and international clients while keeping his setup budget below AED 20,000.
A meaningful share of his clients are UAE companies, which pulls toward Mainland. His budget fits a Free Zone and not much else, which pulls the other way. Neither option is clearly right for him.
His fears are the fears of everyone in this situation: choosing the wrong structure, underestimating the cost, and discovering the limits after the money is spent. That tension became the spine of the recommendation logic. If the design stays honest and calm through his case, the easy cases take care of themselves.
How the experience works
Three input steps. Each collects only what materially changes the recommendation. About you: current status and visa context, because employment affects the setup path. Your business: activity and customer geography, the two strongest signals in the structure decision. Your priorities: budget, timeline and visas, which decide feasibility and how confident the recommendation can be.
The reasoning panel is the trust mechanism, not decoration. It shows what the system can already infer, in the founder's own terms, while they are still answering. The recommendation forms in view instead of arriving from nowhere, which means by the time they reach the result they have already watched it being built.
Two details carry a lot of weight in the intake. Every option has a one-line consequence under it, so a founder learns what the answer implies as they pick it. And every question has an honest escape: Not sure yet, Other or not sure, Confirm with your advisor. Each one states what the tool will do with that answer rather than blocking the path.
The engine: constraints first, then conflict
I designed the recommendation logic before the interface. An earlier version returned one confident answer every time. It looked more decisive and read as less trustworthy. I rebuilt the results layer around hard rules, conflict and confidence states.
Constraints close the question
If a hard rule fires, invoicing UAE clients directly, manufacturing with a warehouse, the structure is decided at that moment. The remaining questions stop pretending the choice is open and refine cost and timeline within it. A settled question never reopens.
Conflict is shown, not averaged
If no rule fires, directional pulls are tallied. When they agree: one recommendation, high confidence. When they conflict, the product does not blend them into false certainty. It shows both pulls, names the single variable that decides it, marks confidence as low, and says this is exactly the case worth thirty minutes with an advisor.
Confidence is a state, not a decoration
The confidence badge reflects whether the logic actually resolved. High means a rule fired or the pulls agreed. Low means the tool is telling you it cannot settle this, which is the answer Ahmed gets.
Key product decisions
| Decision | The obvious build | Why I rejected it | Accepted cost |
|---|---|---|---|
| Recommend from business signals, not the user's guess | Ask the founder to pick a structure and work backwards from it. | The user's inability to answer that question is the entire reason the product exists. | More questions to reach the same place, and a founder who already knows must still answer them. |
| Hard constraints settle the decision early | Score everything equally and blend it into a weighted result. | Once a legal rule fires, the choice is closed. Keeping a settled question visually open invites the user to second-guess a fact. | The flow behaves differently depending on the answer path, which is more logic to build and more states to specify. |
| Budget changes the output, it is not just collected | Cut the budget question, since on its own it never forces a structure. | When the right structure costs more than the stated budget, the product says so and lays out the two honest paths. A form that collects your budget and then ignores it is theatre. | A question that does not drive the primary decision still spends one of the four minutes. |
| Uncertainty is a valid answer | Always return a confident recommendation. It looks more polished. | For a confused first-timer, a confidently wrong answer is worse than an honest judgment call. The hardest call here. | The product looks less decisive in exactly the demo where it would most like to impress. I built the low-confidence paths knowing that. |
| Show the cost of changing your mind later | Present the choice and stop there, as most tools do. | What keeps a founder up at night is being wrong and paying for it afterwards. Starting in a free zone and adding a mainland entity later costs roughly AED 15k to 25k and four to eight weeks. That is what a good advisor would say, so it belongs on the screen. | A number that can date, on a screen that already carries several. It needs owning as fees change. |
The handoff is designed, not bolted on
A recommendation cannot replace legal consultation, so the end of the flow had to hand off honestly rather than pretend to be final. The advisor call is where the open judgment calls get resolved, and the design says that in those words.
Two lines do disproportionate work here. "We'll reach out within one business day. This does not start any payment or commitment" removes the reason a founder hesitates at the last step. And on the landing page, "No payment. No account required to start" does the same job at the first one. The same anxiety, addressed at both ends.
Limitations, and what comes next
The tool recommends a structure and a licence direction, not a specific free zone or jurisdiction. It leaves out tax optimisation, banking eligibility and legal consultation, which need operational expertise beyond an early recommendation tool.
The question order and the cost bands are reasoned from public documentation, not yet validated with real founders. It is honest thinking, not a finished product, and anyone using it should hear that plainly.
Next, in order: jurisdiction-specific recommendations built on real fee data; saved scenarios so a founder can compare paths side by side; and extending the same constraints-first engine to other markets.
Recommendations alone do not build trust. Understanding does.
People rarely struggle because they have too little information. They struggle because they do not know which information actually matters.
The instinct in decision-support products is to add: more data, more comparison tables, more options. The work is almost always subtraction, finding the two or three inputs that genuinely move the answer, and having the discipline to be honest when they conflict.