AI Transformation vs. AI Theater

A Stagwell survey of 100 U.S. CEOs running 10,000-plus-employee organizations found 85% see AI as the defining transformative technology of the year and 78% are bullish on what it'll do for workplace efficiency. That's as close to unanimous as this kind of survey gets. Every one of those executives has told their board, their investors, and their town hall that the company is transforming with AI.

A few weeks later at Davos, Uber CEO Dara Khosrowshahi told investors to avoid most of them — the ones "saying the right words and kind of play-acting their way into a pretend transformation." The CEOs nodding along at his keynote mostly sit inside that 85%. Both numbers can't be right.

The Meeting You've Already Sat Through

You're in a board update. The slide says "AI Transformation — In Progress." You list the line items: marketing rolled out a writing assistant, engineering adopted Copilot, customer support launched a chatbot for tier-one FAQs. Adoption rates are north of seventy percent. Satisfaction scores are solid. Someone from the audit committee asks which business outcome has moved, and the room gets quiet for a beat before you pivot to a productivity anecdote from the pilot.

That pause is the tell. Nobody is lying. You're describing adoption, and adoption isn't transformation. Drop new tools into an unchanged process and the process still determines the outcome — the tool just makes the old way slightly cheaper to run.

Khosrowshahi was unusually specific about what this looks like inside a company. Uber tried the obvious thing first: take customer service policies that had been refined over years and train an AI agent to follow them. The results were underwhelming. The breakthrough came when they inverted the problem — gave the agent the objective ("customer feels good at the end of the interaction") and let it design policy from scratch. "You've got to break down those rules and start over," he said, "to get the full potential of AI inside of your company."

That's not a tooling insight. That's an organizational design insight wearing a tooling costume.

Why the Rules Exist in the First Place

Every policy in your customer service handbook was written because a specific thing went wrong. A refund got issued when it shouldn't have. An escalation path failed. A rep promised something legal couldn't deliver. The rule was the scar tissue. Over a decade, the handbook becomes a dense network of accumulated incidents, each reasonable at the time, each now a load-bearing constraint on how work gets done.

Your AI agent inherits all of that when you train it on the handbook. It also inherits the assumption underneath it: that the executor has limited judgment, variable effort, and a finite attention span. Many rules exist not to achieve an outcome but to contain human variance. When the executor is a model that reasons from an objective and holds no grudges, a lot of them are solving a problem that no longer exists.

The implication is that real AI deployment requires excavating decades of institutional decisions and asking which ones still bind. Most organizations won't. It's politically hard, it requires admitting that current practices are obsolete, and it means a cross-functional fight with the people who own the rules. Running ChatGPT Enterprise through procurement is a ten-person decision. Redesigning customer operations around a new executor is a three-hundred-person one.

Call It AI Theater

That's the pattern worth naming: AI theater. Not fraud, not cynicism — something closer to an organizational coping mechanism. You want the upside that Khosrowshahi and the survey respondents are both gesturing at. You don't want the three-hundred-person fight. So you deploy the tools, announce the transformation, and hope the adoption numbers carry the narrative long enough for real results to appear on their own.

They usually don't. Klarna publicly restructured around AI-first customer service, announced dramatic headcount reductions, then quietly reversed course when quality problems surfaced with real customers. That's what happens when the tool deployment moves faster than the organizational rethink underneath it. The surface looked transformed. The rules it was built on hadn't changed.

The Stagwell number starts to make sense in this light. 85% of CEOs honestly believe they're transforming, because from where they sit adoption is happening, spend is real, and vendors validate the story enthusiastically. Khosrowshahi's read from the same vantage point is that most of what he's seeing is veneer. Both are correct observations of the same phenomenon. Veneer is what transformation looks like from the executive summary.

The Reframe

Here's what's actually going on. AI didn't create a technology decision. It created an organizational design decision, and the technology decision is the part everyone can agree to. The reason your AI initiative can't produce a clean ROI number isn't because the tools don't work. It's because you deployed them against an organizational shape that was calibrated for a different constraint — human execution capacity — and now the constraint has moved. The tool is fine. The scaffolding around it is the problem.

This is why the veneer pattern is so stable. The tool purchase gets you the press release, the board update, and the competitive parity. The rule excavation gets you the actual outcome, but nobody sees it for two quarters and there's no clean narrative while it's happening. The incentive gradient runs toward theater. That's why theater keeps winning.

The companies that actually transform treat the AI decision as a forcing function for the harder question: which of our current practices exist because the old executor couldn't do better, and would we design them this way today? Uber asked that about customer service. GigSmart built its operating model — connecting workers to jobs across all 50 states — around which workflows genuinely scaled and which didn't. Brandfolder compressed asset-search times by 90% before its $155M Smartsheet acquisition by rewriting the assumption of what "searching for an asset" required. The tooling came in afterward.

The Honest Version of the Board Slide

Run the real audit and most AI transformation decks come back with the same finding: adoption is high, workflows are substantively unchanged, outcome metrics haven't moved in ways you can cleanly attribute, and the policies the AI is operating against were written for humans. That's theater with good intentions. It's also the majority of the market right now, which means the companies doing the harder work are about to separate from the ones that didn't — and the separation will become legible in outcome metrics roughly two years later, long after the rollout decisions have been celebrated and absorbed.

A simpler diagnostic than any ROI framework: can you point to a rule you used to have that you don't have anymore, because the AI made it obsolete? If yes, you're in the 15%. If the answer is that the tools are deployed and adoption is strong, you already know which side of Khosrowshahi's line you're on.

If that question lands uncomfortably, that's useful information. We do this work with leadership teams who can feel the gap between the AI narrative and the AI results, and want an outside eye on the rules underneath the deck.