Over the past year we have sat in a lot of rooms with business leaders in the UAE, across the GCC and further afield: board members, CIOs, heads of operations, chief risk officers and the engineers, adjusters and planners who actually do the work. The industries differ, but the questions are remarkably similar.

This is not a survey. It is a summary of the questions we hear most, and the answers we have found hold up in practice.

"Where do we start?"

The most common mistake is to start with the technology. Teams pick a model or a tool, then look for a problem it can solve.

We suggest the opposite. Start with one decision that matters and happens often: approving a claim, evaluating a tender, planning festive stock, reviewing a credit file. Ask three questions about it:

  • Who makes this decision today, and how long does it take?
  • What knowledge do they need, and where does it live?
  • What would a better decision be worth?

If you can answer those, you have a pilot. If you can't, you have a discovery workshop.

"Will it replace our people?"

In our experience, the agents that succeed are the ones that make experienced people faster and new people better, not the ones that try to remove people from the process.

The work an agent does well is the groundwork: reading long documents, gathering evidence from several systems, comparing a case with the rules and with past decisions. The work people do well is judgement: weighing the evidence, handling the exception, taking responsibility for the outcome.

When teams see an agent as a colleague who prepares the file, adoption follows. When they see it as a threat, it stalls.

"How do we keep it accountable?"

This is the question boards and risk committees ask first, and it is the right one. Our answer has three parts:

  • Every output cites its sources. If an agent can't show where a finding came from, nobody should rely on it.
  • A named person approves consequential decisions. Agents recommend; people decide. The record shows who.
  • Everything is logged. Inputs, sources, model and prompt versions, reviewer, decision and reason.

With those three in place, an AI-assisted decision is more auditable than most manual ones.

"What about data protection and where our data lives?"

Organisations in the region are rightly careful about personal data and data residency. We hear concerns about data leaving the country, being used to train models for others, and being kept longer than needed.

These are design choices, not obstacles. Decide where agents run and where their data is stored before the pilot. Retention and access rules follow your policies, your privacy notice and the data protection law that applies to you. Agree these with your security and legal teams before the pilot, not after.

"Does it understand how we work here?"

Businesses in the Gulf work in Arabic and English, follow local regulation, and run to a calendar shaped by Ramadan, the two Eids and a Friday–Saturday or Saturday–Sunday weekend. Generic tools often miss this.

Local context matters in small ways that add up: a site manager writing notes in Arabic, a procedure that mixes Arabic and English terms, a demand spike that follows the Hijri calendar rather than the Gregorian one. Build for it from the start.

"How do we know it's working?"

Agree the measures before the pilot starts. Good ones are simple and owned by the business: time to complete a review, the share of agent findings reviewers agree with, the number of issues caught earlier, and how reviewers feel about the work.

Run the agent alongside today's process first. Compare its output with what your experts decided. Only move to live use when both sides are satisfied.

What we take from this

The leaders who are getting value from AI agents are not the ones with the biggest budgets. They are the ones who picked a real decision, kept their people in charge, and insisted on evidence. That is what the Basinwright framework is built to support.

AI strategyGovernanceLeadership