Recently, I wrote about the questions regional CEOs keep asking me about AI. Where do I start? How do I stay in control? Will it replace my team? Honest questions from leaders at the start of something.
Then I saw a post from Alex Lieberman, the Morning Brew co-founder who now runs Tenex, an AI transformation firm working with US enterprises. He asked his team, who spend their days with enterprise execs, to share the questions they hear most in the field. The result was a list of sixty-odd questions. It’s worth reading in full.
Here’s what struck me. Almost none of those questions are the ones I hear across the table in the Gulf. And that’s not a criticism of anyone. It’s the most useful thing about the list.
The questions have changed shape
Notice what’s missing from their list. Nobody is asking whether AI matters. Nobody is asking if it’s too early. Nobody is asking whether it will replace their team. Where Tenex works, those questions are settled. Their clients are past the “whether” and deep into the “how”: how do we verify agent output, who is allowed to build what, what does a token cost us, who owns this thing once it’s live?
The gap between their list and mine is somewhere between twelve and eighteen months. Which means their list is not a curiosity. It’s a preview of what lands on your desk next. Sixty questions are a lot, but they collapse into seven anxieties. Here they are, with my honest read on each.
1. Can I trust what it produces?
The biggest theme by volume. How do we know the agents we build are any good? How do we test AI features alongside the software they run on? And the sharpest framing on the whole list: when an agent produces a full day’s work, how does a person verify it in minutes?
The honest answer: you don’t check everything, you build checking into the work. Clear definitions of what is done before the agent starts. Evaluations that run automatically. Spot-checks where the risk justifies them. Treat the agent’s output the way you would treat a capable new hire’s output in week one. Trust grows as verification holds.
2. Who’s allowed to build, and inside what guardrails?
Who gets access to tools like Claude Code? Should staff be creating their own skills and apps? How does the front line build useful things without touching mission-critical systems? And who owns a build once it’s deployed?
The answer isn’t to lock it down, and it isn’t to open the floodgates. It’s tiers. Everyone gets a sandbox. A governed few get a path to production, with an owner named before anything ships. The principle I outlined in the last piece about brand and data holds here, too: set the boundaries first, then move quickly within them. An ungoverned building isn’t empowerment. It’s future incident reports.
3. Is our data safe, and is it ready?
Two different questions that travel together. Safe: what can we put in, especially in regulated industries, and how do we protect proprietary data while still letting the tools do real work? Ready: What do we do about fragmented data, and how do we capture the knowledge that lives in people’s heads into something the whole business can draw on?
Safety is largely solved if you buy properly. Enterprise-grade agreements, data that is never used for training, and access controls that mirror the ones you already run. Ready is the longer game, and the trap is believing you need a two-year data overhaul before AI can start. You don’t. Start where the data is already good enough, and let the early wins fund the clean-up.
4. What does it cost, and what is it worth?
How do we control AI spend without capping productivity? How do we prevent runaway sessions? How do we attribute value to a token and capture the full ROI?
This is a budget line that didn’t exist two years ago, and most finance teams still don’t have a category for it. The mistake is measuring cost per token and never measuring value. Measure both at the level of the workflow: what did this process cost before, what does it cost now, and what happened to speed and quality. Tokens are an input. Workflows are where the money is.
5. Who owns this?
What does the operating model look like with AI in it? Who owns AI internally? How do we run the internal motion when everyone already has a day job?
Someone senior owns it, or nobody does. And the day-job problem is real. The reason so many internal AI programmes stall isn’t a lack of enthusiasm. It’s that transformation that got assigned as a side project. That’s also the honest case for external partners: not because your people can’t do it, but because somebody has to do it as their whole job while your people keep running the business.
6. Are we betting on the wrong horse?
How do we stay multi-model? How do we avoid being locked into one lab? How do we know which models are good?
Here’s a take most vendors won’t give you: lock-in fear is overweighted at this stage. Being locked into nothing is how companies ship nothing. Pick the model that’s best for the job today, build your workflows and skills so they’re portable, and accept that switching later is a cost of doing business, not a catastrophe. The companies that hedged everything for two years have beautiful vendor matrices and no results.
7. How do we bring people with us?
How do we get people excited rather than scared for their jobs? How do we sell AI internally? How do we transform a multi-thousand-person organisation without leaving anyone behind? And the quiet one underneath all of it: are we behind?
Same answer I gave the CEOs. The change work is the work. Tools are the easy part. And on “are we behind”: behind is not a function of the calendar. It’s a function of whether you’ve started operationalising. Plenty of companies bought early and have nothing to show. Plenty started later and are compounding.
What this means if you’re reading it from the Gulf
The advantage of being twelve months behind the frontier is that you know the syllabus. US enterprises are paying full price for these lessons. You can get them at a discount, but only if you treat this list as a roadmap rather than a spectator sport.
And you don’t answer sixty questions by discussing them. You answer them by putting one real workflow into production, safely and deliberately, and letting the questions that matter surface from there. Start narrow, prove it, scale the pattern. Then, when the questions on this list arrive at your desk, and they will, you’ll be answering them from experience rather than theory.
If you want to know which of these questions your business should be answering right now, our three-minute readiness assessment is the fastest way to find out. Or let’s talk.