Not too long ago, one of the world’s most data-driven companies admitted that its AI spending was becoming “harder to justify”. Uber’s President and COO, Andrew Macdonald, told the Rapid Response podcast that the company had blown through its entire 2026 AI budget in roughly four months, with around 5,000 engineers leaning on Anthropic’s Claude Code.
Uber isn’t alone. Forrester research found that enterprises are deferring around 25% of planned AI spend to 2027, as CFO scrutiny over ROI intensifies. And McKinsey’s State of AI report summarized that while 62% of companies are experimenting with AI agents, only 23% have scaled them in even a single business function.
While these may look like the statistics of a technology that isn’t working, they’re actually the statistics of a technology being used in the wrong way.
If we rewind back to 1987, Nobel laureate economist, Robert Solow, observed something that many at the time really resonated with: “You can see the computer age everywhere but in the productivity statistics.” Computers were everywhere across the innovative businesses that had invested heavily in them, but the productivity numbers didn’t move.
The returns only materialized years later, once organizations stopped bolting computers onto old processes and started fundamentally redesigning how they worked.
Many executives, alongside Uber’s COO, are at exactly that inflection point with AI – people are using it, running out of budgets to maintain usage, and at the same time, not really seeing the productivity boost they were hoping for.
This is the Solow Paradox repeating its course, which begs the question: will enterprises learn from previous mistakes?
The flatline behind the hype
There’s a critical distinction that most enterprise leaders are still failing to make: AI activity is not the same as AI maturity. You can run 40 pilots, adopt six platforms and report impressive usage statistics, and still be no closer to measurable business value.
Uber found this out in painful, public fashion, and has since openly questioned whether the rising cost of AI token usage is translating into proportional productivity gains. What makes Uber’s situation instructive is not just the financial exposure, but how the organization approached adoption. Internal leaderboards were introduced to rank teams by AI tool usage, with the incentive being to use more tools.
The outcome was more usage, but the business impact slowly became harder to justify.
This is what happens when you gamify adoption without redesigning the workflows underneath it. You optimize the tool usage metric, not the business outcome. Uber has now joined several other top organizations, including Microsoft, Meta and Amazon, in capping AI usage to tackle the issue.
What’s interesting here is that when AI token usage is unconstrained, activity becomes the proxy for progress, but when it’s capped, organizations are forced to confront a harder question: what is each token actually producing?
In that sense, token spend behaves like an economic mirror. It scales immediately with adoption, while productivity only improves when workflows are redesigned. The gap between the two is where most AI ROI disappears.
The real culprit: Individual task optimization
When AI tools are deployed at the individual level, they tend to optimize the task, not the workflow.
A developer writes code faster, a marketer drafts copy in a fraction of the time, or a data analyst generates a summary report in minutes rather than hours.
All of these examples are real gains, but if the code still sits in a review queue for four days, if the draft still passes through three rounds of manual approval, or if the report still requires someone to manually transfer it into a decision-making dashboard – the time saved will pool at the next bottleneck.
Individual productivity gains that don’t translate into workflow redesign don’t compound. They stagnate, and this is the core of the maturity gap. AI maturity isn’t about how many tools you’ve adopted, or how many pilots you’ve launched. It’s about whether you’ve moved consistently from opportunity to outcome.
That shift requires a fundamentally different way of working.
Now, the term ‘production-ready AI’ gets used loosely. It’s worth being precise about what it actually means in practice, because most enterprise AI deployments fall short of the bar.
Production-ready AI has four characteristics. First, the output feeds directly into a downstream decision or action without manual transfer. It’s embedded in the workflow, not adjacent to it.
Second, the system has clearly defined failure modes, so the organization knows exactly what happens when the AI gets something wrong, and who’s accountable for remedying it. Third, there’s a named owner responsible for performance, adoption and iteration. Finally, and most critically, the surrounding process has been redesigned – not merely augmented.
The framework that closes the gap
So, how can businesses make this shift in practice? The answer is maintaining disciplined execution.
One way to build that discipline is a structured cadence we call the 3-3-3 framework: three days to prioritize, three weeks to prove value, and three months to launch a first release. The logic is deceptively simple and deliberately structured.
In the prioritization phase, the question is not “what can AI do?”, it’s “which specific opportunity, tied to a specific business outcome, has the right combination of value, feasibility, data readiness, and organizational sponsorship to pursue right now?”
That focus alone eliminates a significant proportion of AI initiatives that consume resources without clear purpose. It’s the antidote to the open-ended experimentation that left Uber burning through its budget before April was out.
In the proof phase, the focus is validation – not in technical terms, but in commercial ones. Can this solution create measurable value for users and for the business? A proof of concept that demonstrates technical possibility without demonstrating business value isn’t a proof of concept, it’s a prototype without a destination.
In the launch phase, the solution moves into a real environment. It integrates with existing systems, gets adopted by the people it was built for and gets measured against the outcome it was designed to improve. Not against token usage and not against adoption rates.
This rhythm isn’t a rigid formula, however. The shape of the work always depends on the business problem, the data environment, the technical complexity and the organization’s appetite for change. What the framework provides is momentum and the discipline to keep that momentum anchored in value.
Shifting from adoption metrics to outcome metrics
The most consequential change enterprise leaders can make right now is a measurement decision.
Enterprises that are serious about closing the gap between AI activity and business impact need to make three shifts. The first is from tool deployment to operating model redesign. Rolling out AI tools is table stakes but building a repeatable operating model – a structured path from idea to proof to scale – is the competitive differentiator.
The second shift is from adoption metrics to outcome metrics. Usage rates, logins and token volumes tell you whether people are using the tools, but they don’t tell you whether the tools are working. Define success in business terms from the outset: cost reduction, time-to-decision, revenue impact, customer satisfaction, and build your measurement framework around those outcomes.
The third shift is from centralized experimentation to distributed accountability. The AI factory model – where a central operating model connects business priorities with delivery, adoption and value measurement – works precisely because it distributes accountability across the organization. Every initiative starts with a clear owner, a clear problem and a clear definition of what success looks like.
Winning the AI day
The productivity paradox Solow identified in 1987 eventually resolved itself. Not because computers got better – though they did – but because organizations learned to reorganize work around the technology, rather than fitting the technology around old ways of working.
The same resolution is available to enterprises deploying AI today. But it requires leaders to make a deliberate choice – to stop measuring success in terms of how many tools have been adopted and start measuring it in terms of how many outcomes have been delivered.
The organizations that win the AI era will not necessarily be those with the largest number of pilots, but rather the ones that build the maturity to turn the right ideas into value and then scale that value with confidence.
That kind of maturity doesn’t happen by accident. It requires structure, discipline and the willingness to ask harder questions about what AI is actually delivering, and what it’s not.
We’ve featured the best AI chatbot for business.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
