The AI Productivity Paradox: Why cheap execution is exposing a crisis in strategy
Every corporate team, from engineering to marketing, operations, or product, has a quiet graveyard of ideas that never made it. For years, people believed these ideas were just too expensive to pursue. Generative AI was supposed to change that. The hope was that by making execution cheaper, companies could finally tackle their backlogs, meet hidden demand, and unlock new value.
It’s a neat story, but it misses the real reason most backlogs exist.
Making execution cheaper doesn’t suddenly reveal a hidden stash of great ideas. It just takes away the one thing that was quietly filtering the backlog: cost. Without that filter, organizations quickly learn whether they ever had a real way to judge quality, or if cost was just filling that role.
The Jevons Illusion: More Output, Less Value
Economists call this pattern the Jevons Paradox. In 1865, William Stanley Jevons noticed that as technology made coal use more efficient, people didn’t use less coal—they used more, because it became cheap enough for new uses. And nearly every AI specialist uses this term to explain modern AI trends in social media.
Today’s indicators point to a similar rebound in software and content production:
Infrastructure spend is surging. Gartner projects global IT spending will reach $6.31 trillion in 2026, a 13.5% increase from 2025, driven substantially by AI infrastructure and memory. (Gartner, April 2026)
Development activity is up sharply. GitHub’s 2025 Octoverse report found developers merged an average of 43.2 million pull requests per month over the year, up 23% year-over-year, alongside a 25% rise in code pushes. (GitHub, 2025 Octoverse)
Code, content, and features are now much cheaper to produce, so organisations are making more of them, faster. But these numbers don’t show whether this extra output is being reviewed as carefully as before. That’s a fair question, especially given the speed, but it’s not something spending or PR data can answer.
What the data does show clearly is that this surge in output has not translated into a surge in enterprise value. According to PwC’s 29th Global CEO Survey, which surveyed more than 4,450 executives across 95 countries, 56% of companies report seeing neither higher revenue nor lower costs from their AI investments. Only 12% — a group PwC calls the “vanguard” — report capturing both. (PwC, 29th Global CEO Survey)
AI Impact on Enterprise PerformanceShare of CompaniesNeither higher revenue nor lower cost56%Both higher revenue and lower cost (“vanguard”)12%
Organizations are generating substantially more output. Most are not converting that output into outcomes.
Pressure-Testing the Productivity Myth
To understand why this gap persists, leaders need to look past activity-based metrics and challenge the assumptions driving current AI investment. Consider three common pushbacks.
1. “Our team is shipping features faster, and support tickets are down. Isn’t that a clear gain?”
Not by itself. Shipped features and closed tickets measure volume, not value. Shipping faster just shows the team is active; it doesn’t show whether what they shipped is useful or well used. Fewer support tickets might mean the product is easier to use, or that fewer customers are using it enough to run into problems. Without tracking retention or adoption alongside volume, you can’t know which is true.
2. “Isn’t this just the typical, messy adoption curve we see with every foundational technology?”
Partly. Cloud and mobile also had bumpy adoption curves. But some fields have clear ways to measure results, which marketing and commerce often lack. For example, AI in healthtech is judged by clinical outcomes, and fraud detection is measured by how much fraud it stops. Marketing does have solid metrics like CAC, LTV, retention, and incrementality testing, but these are only useful if used properly. The real issue is how often teams switch to easier metrics, like impressions or engagement, as soon as AI-driven volume increases. These easier numbers can make it look like there’s progress, even when there isn’t.
If an AI strategy amounts to generating more content to push harder for a transaction, that isn’t a new value mechanism — it’s an old one with a faster engine. A hyper-personalized AI campaign and a generic “buy one, get one free” promotion are both designed to create urgency for a quick purchase. Whether the AI version is actually cheaper per acquisition, at comparable retention, is a testable question — and one most organizations running these campaigns haven’t yet tested.
3. “If judgment is the bottleneck, what’s the operational fix?“
It’s a resourcing decision, and the current data suggests most companies are making it the wrong way. Deloitte’s 17th annual Tech Trends report, discussed publicly by Deloitte chief technology officer Bill Briggs, found that companies are allocating roughly 93% of AI budgets to technology and tooling, and only about 7% to the people expected to use it — training, workflow redesign, and change management. (Deloitte Tech Trends 2026; Fortune interview with Bill Briggs, Dec. 2025)
Corporate AI Budget AllocationShareTechnology infrastructure & tooling~93%People, training & process redesign~7%
That ratio is the strategy, whether intended or not. Handing advanced tooling to an undertrained organization isn’t enablement. AI is a capable execution partner, but a weak strategic compass on its own — asked to validate a marketing or product strategy, it will produce a fluent, confident answer regardless of whether the underlying logic holds up.
Operationalizing Judgment
It’s worth being precise about what PwC’s data actually shows: the 12% “vanguard” aren’t defined by better taste. PwC attributes its results to what it calls AI foundations — a clear strategy, integrated technology environments, and stronger data infrastructure. Only a minority of companies report having more than a few of those foundations in place.
That difference is important, but it doesn’t weaken the point about judgment. Instead, it shows what judgment needs to work at scale. A team with good data systems and a clear goal can spot weak ideas, stop them, and reallocate resources. Without that setup, there’s no reliable way to catch bad ideas before they launch, no matter how good someone’s instincts are. Foundations are what turn judgment into action, instead of just a thought that never leaves someone’s head.
Three diagnostic questions get at whether an organization has that mechanism, or just has volume:
Can the organization actually kill an idea before it ships? If everything that clears technical feasibility also clears the backlog, cost was never the filter — there wasn’t one.
Are performance metrics tied to adoption and retention, or to output? Ship counts and content volume are easy to report and easy to inflate. Ninety-day active usage is harder to fake.
Does anyone’s role or incentive depend on saying no? A 93/7 spending split suggests that the people closest to a bad idea — the ones with the context to flag it early — often lack the time, training, or standing to do so.
Rebalancing the Investment
It’s easier to point out this imbalance than to fix it. Boston Consulting Group’s research on AI transformations at hundreds of companies gives a clear benchmark: about 10% of AI’s business value comes from the algorithms, 20% from the technology and data needed to use them, and 70% from the people and processes around them. BCG calls this the 10-20-70 rule. (BCG, “AI Transformation Is a Workforce Transformation”) Deloitte’s 93/7 spending split and BCG’s 10-20-70 value split both show the same problem from different angles—one tracks where the money goes, the other where the value comes from. Both point to the same issue: most organizations are putting the most money into the smallest source of value.
BCG’s research is specific enough to be actionable rather than aspirational. Companies it classifies as “future-built” — the ones actually capturing outsized value from AI — plan to upskill more than half their workforce, compared with roughly 20% at other companies, and are about four times more likely to run structured, ongoing AI-learning programs rather than one-off training sessions. They’re also roughly five times more likely to do deliberate strategic workforce planning: mapping which roles and skills will change before the change happens, rather than reacting to it afterward. None of that requires a bigger AI budget — it requires moving budget that’s currently going to technology and tooling toward three things: protected time for employees to actually learn the tools they’ve been given, workflow redesign so AI changes how work gets done rather than just sitting on top of the old process, and governance tied to a small number of priorities — BCG recommends three to four — instead of dozens of scattered use cases nobody owns.
Measurement: What to Track Instead of Output
If tracking ship counts and PR volume isn’t helpful, McKinsey’s way of measuring AI value is more useful. They suggest looking at five connected layers: whether the model works well technically, whether people are actually using it in real work, whether key operational numbers are improving (like speed, cost, or accuracy), whether those changes show up in the results leaders care about (like retention, satisfaction, or decision quality), and finally, whether it affects revenue, profit, or costs. (McKinsey, “From Promise to Impact”) The main idea is that most organizations stop at the first two layers—model performance and adoption—and treat that as proof of value. But that’s not enough. A tool can work well and be used a lot, but still never improve the business if it’s not tied to real outcomes.
McKinsey’s research suggests a practical approach: decide which of the five layers matters for a project before it starts, not after. Build in clear checkpoints, like “is this being used in real work,” “is there a measurable operational effect,” and “does the financial impact justify scaling.” A project should pass these checks before receiving additional investment or a broader rollout. This is a way to make sure the same discipline from the earlier diagnostic questions is built in. If an organization can’t say which layer an AI project is supposed to improve, or can’t show proof that it did, it doesn’t just have a measurement problem. It has a judgment problem that looks like a dashboard.
The Strategic Mandate
Making execution cheaper doesn’t fix a weak strategy—it just makes it bigger. Compute power and content creation are now basic utilities. But the discipline to decide what’s truly worth building is still rare.
The real question for leaders isn’t how much their organization can produce. It’s about whether this new ability to produce at scale shows a strategy worth speeding up, or just exposes a judgment problem that was once hidden by the cost of doing anything.
