Paste a video URL in the inspector
On August 18, MIT restricted access to a report that had just crashed tech stocks. The headline: 95% of enterprise AI projects fail. Fortune amplified it. Markets panicked. A Wharton professor called for retraction.
But the report didn't actually say AI fails. It said companies can't measure AI success using traditional metrics. The same research - based on interviews with 52 organisations - found external AI partnerships succeeding twice as often as internal builds. It found clear ROI in back-office automation. Enterprise transformation metrics missed it. The MIT report didn’t create the confusion. It revealed it.
Meanwhile, McKinsey defines top-performing companies as those attributing at least 20 percent of their EBITDA to AI. But that figure is self-reported - it measures confidence, not maturity. McKinsey’s ‘AI high performers’ are the companies that say they’ve cracked ROI, not necessarily the ones that have. Both cannot be right - unless we are looking for value in completely the wrong place.
Leaders are being told to find ROI in lagging, enterprise-level outcomes like EBITDA. But you can’t navigate a real-time technological shift by looking in the rearview mirror. Most companies aren’t failing at AI so much as failing at accounting for it.
There's a critical distinction here. A few hundred companies globally are building AI infrastructure (OpenAI, Anthropic, etc.). They can afford to lose money for years chasing monopoly. The other 99.999+% are using that infrastructure to improve their business. They need positive ROI. But many of them are copying platform strategies when they should focus on adoption - measuring real returns from applied AI.
The solution is a fundamental reorientation. To find AI ROI, we must stop looking at P&L and start looking at work.
We need to change how we think about measurement during technological transitions - whether we're measuring impact on jobs or returns on investment.
An economic historian can look back at say, the computer revolution and calculate its impact using Total Factor Productivity - a model that measures output growth beyond what labour and capital alone would produce. But TFP takes decades of complete data to calculate. It's the economist's autopsy of productivity, not the executive's vital signs. They have the complete picture - a finished map of how technology reshaped entire industries over 50 years. But this retrospective analysis is useless for leaders making decisions today. Most executives don’t have fifty years, they have a quarter. It's like trying to navigate a flood by looking at a map of the delta that will exist decades later.
Business leaders are in that flood now. They need immediate, local data to make the next decision. Unit economics provide the real-time signals you need during the disorienting period when old systems break down before new ones deliver compound returns. The accumulation of successful small bets creates the historian's map decades later. But even if you had that map-to-be it couldn’t tell you how to steer today. Right now you don’t need a map of the future delta. You need a stick to poke the river - immediate feedback on local conditions.
The next step is decomposition: breaking the unknowable into measurable units. Google famously asks interview candidates to estimate piano tuners in Chicago - Chicago has ~700,000 households, maybe 1 in 20 owns a piano, each needs tuning twice yearly, a tuner handles 8 per day.
The critical difference from static problems like piano tuners: the units that matter for AI are time-bound. What you should measure today will be wrong next year as capabilities expand. There's no time to plan perfectly - start now, or your baseline will be outdated before you can use it. This is why traditional measurement fails and why most advice about 'what worked before' may well be useless.
Economists Brynjolfsson, Rock, and Syverson call this the 'Productivity J-Curve'. During platform transitions, early investments in reorganisation and capability-building consume resources without producing measurable output. Productivity appears to decline even as valuable assets accumulate. Accountants call it a downturn. Historians call it progress. Traditional ROI looks terrible before it looks transformative.
But this isn't just economic theory - it's the lived experience of every platform transition. Unit economics, however, makes visible the progress that aggregate metrics miss. You can show real returns within quarters, not decades.
While enterprise transformation may take years, unit economics delivers measurable wins within months. You can show real ROI on specific work even while the J-curve plays out at the enterprise level. Unit economics doesn’t replace enterprise metrics - it feeds them. P&L impact is the aggregate of many proven units, not the starting point.
Early stage realities. Most experiments fail - that's expected. Hidden costs emerge: human oversight, error correction, system integration. Some units will show positive ROI immediately. Document processing might save 20 hours weekly. Invoice matching might cut processing time 80%. The task is to find these wins, prove them with data, and expand.
Middle stage discipline. Your positive units are proven. Now compound them. Can document processing gains improve contract analysis? Can automated invoicing enable better cash flow prediction? This is where returns accelerate - not from waiting, but from connecting proven successes.
Late stage transformation. Teams start requesting different workflows based on what AI makes possible. Original units become obsolete as capabilities expand. ‘Documents processed’ becomes meaningless when AI enables continuous analysis. What started as ‘faster document review’ becomes ‘continuous compliance monitoring’ - an entirely different value proposition requiring different measurement.
This pattern of evolving units has happened with every major platform technology:
Containerisation (1950s-1980s)
I talked about Malcolm McLean's metal boxes in The Disappearing Salary. They didn't just make shipping cheaper - they reorganised global production. But the transformation took decades to unfold.
Early stage Efficiency. The initial unit was brutally simple: reduce loading costs from $5.86 to $0.16 per ton. Port operators could measure this immediately.
Middle stage: Capability. As infrastructure matured, the unit evolved to ‘inventory days eliminated.’ Just-in-time manufacturing became possible only because containers made delivery predictable.
Late stage: Transformation. Eventually the unit became ‘geographic flexibility of production.’ Companies could now manufacture anywhere and ship everywhere. The global supply chain was born.
ATMs (1970s-1990s)
Banks initially lost money on every ATM, but by 1990 they were opening 43% more branches than before ATMs existed - a complete reversal of predictions.
Early stage: Efficiency. Banks justified ATMs with one unit: reduce cost per transaction by 50%.
Middle stage: Capability. The unit evolved to 'banking hours coverage' as 24/7 access became the real differentiator.
Late stage: Transformation. The final evolution: 'products sold per branch.' Freed from cash-handling, tellers became sales staff. ATMs didn't eliminate branches; they transformed their purpose.
Unit evolution signals:
When a metric stops driving decisions
When output quality improves but the number doesn’t
When teams start combining units to do new kinds of work
When the Right Metric Destroys the Wrong Business
Every business has unique dynamics that generic metrics miss. The craft of management isn't blindly following a framework - it's understanding your specific system well enough to measure what actually matters.
A law firm's 'document review hours saved' means something different than a bank's. Or take 'Customer acquisition cost', which means fundamentally different things across business models. For FMCG, it's about shaving pennies off billions of transactions where marginal gains compound. For luxury automotive, CAC isn't about efficiency but about not losing anyone during the 18-month journey across dozens of touchpoints - from initial interest through test drives, configuration, financing, and delivery. A single lost prospect represents significant revenue impact given low unit volumes and high ticket sizes. For enterprise SaaS, high CAC is acceptable given multi-year contract values and expansion revenue potential.
The same metric, completely different strategic implications.
The challenge often isn't learning what to measure - you already know what breaks in your business. You know which efficiencies destroy quality, which metrics get gamed, which savings aren't real. Translating that institutional knowledge into units that capture what actually matters - that's the work.
Zillow: A Failure of Context
In 2021, Zillow shut down its home-flipping business, laying off 25% of its staff after a write-down of over half a billion dollars. Their 'Zestimate' algorithm was wildly successful as a media engagement unit. But they made the fatal error of believing a unit successful for consumer engagement could handle the complexity of buying, renovating, and selling thousands of homes. The Zestimate couldn't price in real-world volatility or renovation costs. They optimised a simple number and missed reality.
IBM Watson Health: A Failure of Usability
IBM invested billions positioning Watson as a revolutionary AI oncologist. But hospitals like MD Anderson scrapped multi-million dollar projects after years of pilots. Watson's unit was ‘treatment recommendation generated.’ But the unit was incomplete - the training data was hypothetical, not patient data. The AI was 'working' according to its narrow unit, but the complete unit - ‘safe and usable recommendation’ - failed completely.
Amazon's Recruiting AI: A Failure of Systemic Bias
Amazon abandoned an AI tool after discovering it was biased against women (you can read more about that here). The AI's unit was 'predict successful hire' based on historical data. The unit was fatally incomplete - it ignored systemic bias in the data. It optimised for past reality instead of desired future, creating something ethically and legally unusable.
For that 99.999+% of companies using AI rather than building AI infrastructure, the path forward is disciplined and iterative.
Start Small with Full-Cycle Units. Identify high-value units that capture the complete picture. 'Cost to resolve an issue to customer satisfaction' is complete. 'Cost per ticket' is not.
Buy, Don't Build. The MIT report is unequivocal: external partnerships succeed twice as often as internal builds. Your competitive advantage is in application, not infrastructure.
Focus on the Back Office. While half of AI investment goes to sales and marketing, clearest ROI often comes from back-office automation. The largest savings come from replacing expensive BPO contracts and agency spending with AI tools - external service costs drop while maintaining external delivery.
Make Small, Reversible Bets. Size each experiment to fail without damage. Transformation is accumulation of small, unit-guided steps.
Evolve Your Units. Revisit metrics quarterly. As AI tools mature from basic automation to enabling new capabilities, your units must evolve with them.
Stop waiting for enterprise AI ROI. It won't show up on your P&L for years.
Start measuring unit ROI. It's visible today in specific work: documents processed faster, customer issues resolved cheaper, contracts reviewed deeper.
The 5% of companies finding AI value aren't smarter or luckier. They're just measuring at the right resolution.
© FLUXUS 2026
Fluxus Ltd · Registered in England and Wales · Company no. 08806522 · VAT no. GB 177 5370 79
Registered office: Wellesley House, Duke of Wellington Avenue, London, England, SE18 6SS.