The AI Economy is Slower than Advertised! * Hype runs faster than productivity.

Every CFO I have briefed this year holds two contradictory views about the AI economy at once: that it will compress their workforce within 24 months, and that the whole thing is a trillion-dollar bubble. The disorientation is what the Peak of Inflated Expectations feels like from inside the cycle. It is the cycle doing its job, not a personal failing.

This piece sets out a calibrated picture of the AI economy in 2026. Where AI agents actually sit on the hype curve, why measured productivity is moving slowly while capital expenditure and pilot failures move fast, and what the Friction Tax and the Human Premium Stack tell you to do about it.

The Hype Cycle Lens. * Where AI agents sit in the AI economy of 2026?

Gartner's 2025 Hype Cycle for Artificial Intelligence places AI agents at the Peak of Inflated Expectations, alongside AI-ready data, as the two fastest-advancing technologies on the curve ^[1]. That is Gartner's own language for the moment when a technology's promises run well ahead of its delivered results. In the same window, Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027, and estimates that only around 130 of the thousands of vendors marketing themselves as agentic actually meet a working definition of the term ^[2]. The rest sell "agent washing."

The media incentives compound the distortion. A Reuters Institute analysis of UK coverage found that nearly 60% of AI news articles are indexed to industry announcements, and industry sources make up about 33% of unique sources cited, nearly double academic representation ^[3]. Coverage oscillates between the utopian and the apocalyptic because the middle of the distribution does not get clicks. The result is an information diet that makes calm assessment feel like contrarianism.

The disorientation in the boardroom is therefore a feature of the cycle, not a flaw in the reader. Naming the cycle is the first move toward acting outside it.

The Friction Tax explains the lag. * Why automation compounds slowly?

A single statistic settles the argument about why measured productivity is slower than projected. In a peer-reviewed study of 5.179 customer support agents, generative AI delivered a 14% average gain in issues resolved per hour, with novices gaining 34% and experts gaining almost nothing ^[4]. The number is real and replicable. It is also misleading if you treat it as a forecast for throughput at the firm level.

The Friction Tax is the cumulative cost of the friction sitting between every step of a business process, which automation reduces only when it is removed end to end. A 14% gain on one step of a 12-step workflow does not produce a 14% gain in end-to-end output. The other 11 steps might still require human handoff, scheduling, exception handling, approval, and rework. The unautomated steps absorb the saved time the way slack in a supply chain absorbs a faster truck. The Friction Tax is why a sprinkling of AI across a workflow can deliver impressive point statistics and nearly invisible aggregate impact.

The macro numbers confirm the diagnosis. McKinsey's 2025 survey finds 88% of organizations using AI in at least one business function, yet only about 6% of firms capture meaningful enterprise EBIT value, and most firms reporting any EBIT impact attribute less than 5% of EBIT to AI ^[5]. Adoption is universal. Value is concentrated. The gap between the two is the Friction Tax made visible on a P&L.

A useful historical anchor sits in the electrification record. Twenty-one years after Edison's incandescent lightbulb, only 3% of US residences had electric light and under 5% of factory mechanical drive was electrified; the productivity gains arrived in the 1920s, roughly four decades after the technology existed ^[6]. Software compounds faster than steel, but "faster than 40 years" is still a decade, not a quarter. If a pilot does not reduce the Friction Tax end to end, the productivity gain stays inside the slide deck.

The Klarna Reversal. * A hype-cycle artifact, executed in production!

In February 2024, Klarna's CEO announced that an OpenAI-powered assistant was handling 2.3 million customer service chats in its first 30 days, work equivalent to roughly 700 full-time agents. Average resolution time fell from about 11 minutes to under two. The projected profit improvement landed around $ 40 million for the year ^[7]. The deck went viral. Every customer service deck in the EMEA region cited it within a quarter. In May 2025, the same CEO conceded publicly that quality had regressed and the company had begun recruiting human agents again.

This is not a story about AI failing. It is a story about a firm mistaking automation of a step for automation of an outcome. The agent handled the volume. The Friction Tax in the workflow downstream of the agent, the edge cases, the empathy moments, the escalations, the regulator-facing complaints, did not disappear. The firm rediscovered them at the cost of customer satisfaction and a public reversal.

Two independent failure-rate studies reinforce the same pattern. MIT's NANDA initiative reports that 95% of enterprise generative AI pilots fail to deliver rapid revenue acceleration, with specialized vendor purchases succeeding around 67% of the time against 33% for internal builds ^[8]. RAND finds that 80.3% of enterprise AI projects fail to deliver promised business value, roughly twice the failure rate of non-AI IT projects ^[9]. Two methodologies and two samples converge on the same shape of result. Most of what is shipping today is the inflated-expectations phase.

The asymmetry of pace. * Why the AI economy punishes both action and inaction?

The most uncomfortable structural feature of 2026 is the asymmetry between how fast failure registers and how slowly value compounds. Capital expenditure has been moving fast: IDC forecasts enterprise AI spending of $ 307 billion in 2025 rising to $ 632 billion by 2028, with global AI spending reaching $ 1,3 trillion in 2029 ^[10]. Pilot failures have moved equally fast, well above 80%, and the Klarna reversal showed that public retractions arrive on the same clock. Measured productivity is slow, and the value capture is concentrated in a small share of firms.

A CFO sitting in that configuration faces a real strategic problem. Doing nothing is expensive because competitors are building infrastructure that will compound. Doing everything is expensive because most pilots will not return. The honest answer is that the question is mis-scoped. The TtO Dividend is the time and cost the business actually returns to the customer or the next process step across the full workflow. If the handoff downstream of the agent still requires three people and two emails, the dividend is theoretical.

Where the optimism is earned? * The credible forecast range.

Two credible forecasts bracket the honest range. Goldman Sachs estimates that generative AI lifts global GDP by 7%, around $ 7 trillion, and adds 1,5 percentage points to productivity growth over a decade ^[11]. Daron Acemoglu, the 2024 Nobel laureate in economics, estimates a cumulative GDP gain of 1,1% to 1,6% over 10 years, with an annual productivity contribution near 0,05 percentage points ^[12]. As Acemoglu put it, "I don't think we should belittle 0,5 percent in 10 years. That's better than zero, but it is disappointing relative to the promises." ^[12]

Both forecasts are useful. Treating either as the answer is poor practice. The infrastructure side is unambiguous: the cost of achieving GPT-3.5-level performance fell from $ 20 per million tokens in November 2022 to $ 0,07 per million tokens in October 2024, a deflation of more than two orders of magnitude in 23 months ^[13]. The cost curve will keep compounding even when individual deployments stall. The substrate is being built. The deployments are catching up.

Synthetic Labor is the agent-delivered execution of predictable cognitive tasks at near-zero marginal cost. Inside the floor that Synthetic Labor is repricing, the rational response is the Human Premium Stack, the three tiers of cognitive work that resist commodification. I describe the Human Premium Stack at length in Chapter 12 of my book AI Agents: They Act, You Orchestrate. High-context negotiation, moral arbitration, and zero-to-one work all gain economic premium as the predictable layer below them is absorbed by agents. The Human Premium Stack is where the next decade of cognitive work concentrates, and the firms that climb it first will set the floor for everyone else. The mechanism that prices that floor is the Economy of Intent, where outcomes, not hours, become the unit of exchange.

The Reframe. * The gap between paces is the strategic surface area.

There is a reasonable case for waiting out the cycle. Hold position, let the trough arrive, buy the survivors at distress prices, deploy in the slope of enlightenment. It just is not the case the data supports. Only 6% of firms capture meaningful enterprise value today, and those firms are not waiting ^[5]. The cycle is real, but the cycle does not absolve anyone of agency.

The fear in the room when I deliver keynotes is real. Its target is wrong. Executives feel displacement at news-cycle speed while measured impact moves at organizational-change speed. The gap between those two paces is the strategic surface area. The Friction Tax maps where the work actually sits, which is what makes it useful as a 2026 diagnostic, and the jobs question follows the same calibration logic.

The calibrated position has three moves at different speeds. Move quickly on the infrastructure layer, where compounding is real and the cost curve is collapsing. Take the workflows slowly, because end-to-end redesign is the only thing that converts a 14% point gain into throughput. Start the workforce work today regardless, since climbing the Human Premium Stack is a multi-year exercise that cannot begin the quarter you need the result.

Landing. * What you control?

The hype cycle will resolve itself on its own schedule. What you control is whether your firm spends the next two years building infrastructure, redesigning workflows around the Friction Tax, and climbing the Human Premium Stack, or whether it spends them waiting for someone else to tell you when the cycle has turned. The first posture is unglamorous and compounds. The second is comfortable and does not.

This article covers two frameworks from AI Agents: They Act, You Orchestrate by Peter van Hees. The book maps 18 chapters across the agent-first economy, from the Friction Tax as the diagnostic for stalled software gains to the Human Premium Stack and the Economy of Intent as the destination architecture for cognitive work that resists commodification. If the gap between hype-cycle pacing and productivity-cycle pacing matters to your 2026 plan, the book gives you the complete operating model. Get your copy:*

🇺🇸 Amazon.com 🇬🇧 Amazon.co.uk 🇫🇷 Amazon.fr 🇩🇪 Amazon.de 🇳🇱 Amazon.nl 🇧🇪 Amazon.com.be

References

^[1] Gartner, "Gartner Hype Cycle Identifies Top AI Innovations in 2025," Press Release, 5 August 2025. https://www.gartner.com/en/newsroom/press-releases/2025-08-05-gartner-hype-cycle-identifies-top-ai-innovations-in-2025

^[2] Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," Press Release, 25 June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

^[3] Reuters Institute for the Study of Journalism, "How news coverage, often uncritical, helps build AI hype," 20 May 2024. https://reutersinstitute.politics.ox.ac.uk/news/how-news-coverage-often-uncritical-helps-build-ai-hype

^[4] Erik Brynjolfsson, Danielle Li, Lindsey Raymond, "Generative AI at Work," NBER Working Paper 31161. https://www.nber.org/papers/w31161

^[5] McKinsey & Company, "The state of AI: How organizations are rewiring to capture value," 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

^[6] American Enterprise Institute, "The Dynamo, the Computer, and ChatGPT: Explaining Today's Productivity Paradox," summarizing Paul A. David, "Computer and Dynamo," 1989. https://www.aei.org/articles/the-dynamo-the-computer-and-chatgpt-explaining-todays-productivity-paradox/

^[7] OpenAI, "Klarna: 2.3 million conversations in the first month," Case Study, February 2024. https://openai.com/index/klarna/

^[8] MIT NANDA, "The GenAI Divide: State of AI in Business 2025," via Fortune, 18 August 2025. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

^[9] RAND Corporation, "The Root Causes of Failure for AI Projects." https://www.rand.org/pubs/research_reports/RRA2680-1.html

^[10] IDC, "Worldwide Artificial Intelligence IT Spending Forecast, 2025-2029," Press Release, 2025. https://my.idc.com/getdoc.jsp?containerId=prUS53765225

^[11] Goldman Sachs Research, "Generative AI could raise global GDP by 7%," 5 April 2023. https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent

^[12] Daron Acemoglu, "The Simple Macroeconomics of AI," Economic Policy, August 2024, summarized at MIT Economics. https://economics.mit.edu/news/daron-acemoglu-what-do-we-know-about-economics-ai

^[13] Stanford HAI, "2025 AI Index Report, Chapter 4: Economy." https://hai.stanford.edu/ai-index/2025-ai-index-report/economy