Permission, Not Performance
The machine can now perform everywhere — search, service, the enterprise. What it cannot do is earn the permission to act. That permission is the new scarcity.
The demonstrations have never looked better. An answer engine composes a fluent paragraph about your category in under a second. A shopping assistant compares nine products and names a winner. An enterprise agent drafts the report, reconciles the ledger, files the ticket. Watch any of these for thirty seconds and you would conclude the work is done.
Then look at what actually happens next. The paragraph is read and no link is clicked. The assistant recommends and the human still reaches for the wallet themselves. The agent produces a beautiful draft that no one lets run unattended. The performance is flawless. The permission is missing.
This is the pattern underneath the three territories I track, and it is the same pattern in each. The machine has learned to perform. It has not been granted the authority to act. And across content, commerce, and the enterprise, that authority — not capability — is now the scarce thing.
The word we chose gives the game away
We call this generation of software “agentic”. The word comes from the Latin agere: to do, to drive, to act. We named the entire category after the act. It is worth noting how strange that is, because the act is precisely what the data shows is being withheld. We built machines that can do everything except the one thing their name promises. An agent that cannot act is not an agent. It is a very articulate advisor waiting for a signature.
Hold that irony in mind, because it explains why the demonstrations and the outcomes keep diverging.
The citation ceiling
Start with content, where the gap first became visible. In the United States, Google searches ended without any click 68% of the time in the first four months of 2026, the fastest acceleration in a decade; when an AI Overview appears, click-through rates fall by nearly 60%. The engine performs the answer using your words. It does not pass on the permission — the visit, the credit, the relationship — that used to come with being found.
So the question stops being whether you rank and becomes whether you are cited. And citation is not distributed the way visibility was. The evidence gathered this year by SparkToro and Similarweb, by Salesforce’s commerce data drawn from more than 1.5 billion shoppers, and by the citation audits now circulating among analysts points the same way: the sources answer engines quote are overwhelmingly ones brands already control — first-party pages and structured listings, not the open scramble of the old results page. Generating an answer now costs nobody anything, while the citation stays scarce and has to be earned. That is the ceiling: you can be summarised endlessly and referred to never.
The authorisation gap
Commerce shows the same shape in sharper relief, because here the withheld permission has a price tag. Agentic search as the first step of a purchase journey grew 200% year over year, on Salesforce’s global reading of the market. Shoppers are delighted to let an AI find and compare. They are not delighted to let it pay.
The distance between those two is the whole story. Surveys this year, from Salesforce and Genesys and the practitioner stress tests of live shopping agents, converge on a consumer who grants discovery and refuses the wallet: comfortable letting a machine narrow the field, unwilling to let it place the order. One test run by a vendor in the sector across 15 live commerce agents found that all 15 could answer and only 4 could actually act. The capability to complete the transaction already exists, while the authorisation to use it continues to be withheld. And no amount of demonstration closes a gap that is made of trust rather than technology.
The scaling wall
Inside organisations the pattern hardens into a wall, and the wall is expensive. Deloitte’s Tech Trends 2026, McKinsey’s global adoption data, and IDC’s forecasts on the governance risk of agents tell one story from three angles: pilots that dazzle in the room almost never earn the standing to run in the business. Two thirds of enterprises have experimented with agents; fewer than one in ten have scaled them to measurable value. Barely a fifth have a mature model for governing an agent that acts on its own. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, on rising costs, unclear business value and inadequate risk controls.
Read those figures together and the diagnosis is not that the models underperform. They perform in the pilot, which is why the pilot gets funded. They stall at the threshold where someone must decide to let them act without a human hand on every step — and that decision is a question of governance, identity, and accountability, none of which a better model supplies.
What is actually scarce
Three territories, one substance. When production cost falls to near zero, performance stops being the constraint. Everyone can generate the answer, compare the products, draft the report. What does not fall to zero — what may even be rising — is the warrant to act on it: the earned citation, the granted authorisation, the governed permission to run. That is the asset now, and it behaves nothing like the old ones. It cannot be produced in volume. It accrues slowly, to the sources and systems that have made themselves trustworthy and legible.
The Swiss and Italian markets I work in show this without the American distortion. The Politecnico di Milano’s Osservatorio put the Italian AI market at 1.82 billion euros, up 50% in a year, with most large firms holding generative AI licences — abundant performance. Yet only around one in five uses AI pervasively across functions, and roughly the same share coordinate genuinely autonomous workflows. The capacity has been bought, but the permission to let it act broadly has not yet been given: the same shape as before, in a different currency.
The strategic mistake of the next year will be to keep buying performance and to keep being surprised that it does not convert. It does not convert because performance was never the scarce input. Permission is. The organisations that pull ahead will be the ones that stop asking what their systems can do and start asking what they have earned the right to do — with audiences, with customers, with their own risk committees.
That is the work I want to take apart next: how warrant is engineered rather than wished for — the provenance, the first-party authority, the structure that turns a fluent performance into a cited, trusted, permitted act. Recall is not won at the moment of the answer. It is built long before, in the places you control.
Stay close.
— Stefano
Sources
- Danny Goodwin, “Google zero-click searches hit 68% in early 2026: Study”, Search Engine Land, 9 June 2026, reporting SparkToro / Similarweb clickstream research (US data). https://searchengineland.com/google-zero-click-searches-2026-study-479717
- “Shopping’s New First Step: Agentic Search Grows 200% as Purchase Journeys Start in AI Chats”, Salesforce State of Commerce, Fourth Edition, 28 July 2026 (global; 20 countries, 1.5B+ shoppers). https://www.salesforce.com/uk/news/stories/agentic-search-growth/
- “2026 State of Customer Experience Report: Global Insights for CX in the Agentic Era”, Genesys, 2026. https://www.genesys.com/blog/post/2026-state-of-customer-experience-global-insights-for-cx-in-the-agentic-era
- “State of Agentic CX 2026” (live-agent stress test), Alhena, 2026. https://alhena.ai/blog/state-of-agentic-cx/
- Deloitte, Tech Trends 2026 (agentic AI piloting, production and governance figures). https://www2.deloitte.com/us/en/insights/focus/tech-trends.html
- McKinsey, The State of AI 2026 (share of enterprises scaling agents to measurable value). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Politecnico di Milano, Osservatorio Artificial Intelligence 2025-2026 (Italian market size and enterprise adoption). https://www.osservatori.net/artificial-intelligence/
- “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, Gartner 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
- “IDC FutureScape 2026 Predictions Reveal the Rise of Agentic AI and a Turning Point in Enterprise Transformation”, IDC, 23 October 2025 (forecasts on agent governance and control). https://my.idc.com/getdoc.jsp?containerId=prUS53883425