Uber's COO Andrew Macdonald said something last week that landed harder than he probably meant it to. He told an audience the company's AI spend was getting "harder to justify" because nobody had been able to draw a clean line between what they were paying and any feature anyone used.
It's a small admission with a long shadow.
The Shape of the Question
For about two years, AI budgets have been treated as a strategic capability. Boards approved the line items without asking for ROI, on the assumption that this is a platform shift, and you don't measure platforms in quarterly returns. That assumption is running out of patience.
Most boards in this quarter are starting to ask the question that was awkward to ask in 2024. They're asking it since nobody wants to be the person who looks behind the AI curtain first. But the question is on the agenda now.
Where AI Spend Tends to Break Down
What I see most often, especially with mid-size companies in the GCC: the tool was bought before the underlying process was working. A language model bolted onto a broken sales workflow produces the same workflow at higher speed and with a monthly invoice. This is the version that fails most loudly.
The other common failure is quieter. Nobody captured a baseline before deployment. Without numbers from before, you can't show numbers from after. "Things feel faster" stopped being an acceptable answer for CFOs about six months ago.
There's a third pattern that doesn't get talked about as much. Procurement bought the tool. Operations never used it. The capability sits on a SaaS subscription line that nobody actively owns, and the renewal goes through because nobody wants the awkward conversation of cancelling.
And the vendor stack got expensive on its own. Snowflake, OpenAI, Anthropic, middleware, an internal MLOps line. Each made sense at the time of approval. Together they often exceed what they generate.
What I'd Ask Before the Next Renewal
If I were sitting on the buying side, three questions would matter more than another vendor demo.
What specific outcome is this tool supposed to move? One outcome, concrete enough to put on a dashboard. Reduced support handle time, or higher meeting-booked rate, or faster contract drafting. Anything more abstract gets argued away later.
What was that metric before the tool went in? If there's no number from before, there isn't an investment, there's a spend.
And who uses the output. By name, with a frequency. If the answer is "the team," the answer is no one.
If those three answers don't come back as numbers, the AI is an experiment with a monthly burn. That's fine for an enterprise that can afford to run experiments. It's expensive for a mid-size business that probably cannot.
The Pattern in the GCC
What I see in Riyadh in particular is mid-size companies copying enterprise AI playbooks from American business press. The enterprises that wrote those playbooks can afford the experiment. The mid-size companies copying them usually cannot, but they buy the same tools anyway.
A 200-person family business doesn't need an "AI strategy." It needs a working sales pipeline, a content function that produces things on a schedule, and an operating cadence the team trusts. Once those run, AI compounds them. Until those run, AI just adds a line item.
What I Think Happens Next
I have low confidence on timing and reasonably high confidence on direction.
Budgets get cut, not eliminated. Boards start asking for specific ROI per tool, and most companies rationalize their stack downward. The vendors with the weakest measurable case lose subscriptions first.
The framing also shifts. "AI strategy" becomes "operations strategy with AI components." It sounds like a small rebrand, but it matters. It forces the right framing in budget conversations: the AI serves the business, not the other way around.
The Useful Thing to Take From This
The version of AI adoption I've seen work is the one where the AI disappears into something already functioning. Tools you can point to as a separate "AI initiative" tend to be the ones that get cut first when the budget gets reviewed.
For anyone building right now in Saudi Arabia or anywhere in the region: build the operating system first. Add the AI when the operating system asks for it.
That's the only version of this I've seen outperform the alternative.