The first AI Supper Club, in Riyadh
I arrived in Riyadh knowing nobody. I wanted to hear firsthand how AI is actually being used in Saudi Arabia, beyond the headlines and ambitious announcements. So I organized a dinner.
Around 50 people applied for the first AI Supper Club. Eight joined, from government, industry, finance and tech. There were no slides and no pitches. Over a long dinner high above the city, each person spoke about what they see in their own organization, and the conversation ran well past the plates.
Everyone at the table already uses AI every day. What kept coming up was the distance between that daily use and AI that an organization can rely on. That distance is the work we do at Look Beyond, so it was a good table to sit at.
What stayed with us
People stay in the loop
For specialist work, AI output still has to be checked. The harder question is accountability: when an AI-assisted decision is wrong, who answers for it?
Adopting agents is a bet
An agent may not be right every time today. Teams adopt it because they expect it to get better, faster and cheaper. How much imperfection an organization accepts at the start differs a lot, and it decides how a project has to be delivered: in phases, with a clear path to the quality bar.
Slow processes are the obvious target
Multi-step approvals that take months were the examples people reached for first. With agents, many of those steps could take hours. The value is easy to see. Getting decision-makers to trust the result is the hard part.
Trust travels through references
Buyers tend to pick the safest, best-known option. One early customer running a new product on a single use case often changes that for everyone who comes after.
Adoption is harder than the build
Several guests agreed that building the software is rarely the hardest part. Getting everyone to follow the new way of working is. That gap needs people who can translate between technical teams and the business.
Agents that watch agents
One conversation went deep on monitoring. A supervising layer watches agents in production for signals such as user sentiment across a conversation, or factual errors. It sends weak ones back for improvement, tests the fix in a sandbox and releases it again. It only works with clear criteria for what good looks like, which is the idea behind how we test and monitor agents.
From the evening
