--- title: "Token Fatigue" slug: "token-fatigue" summary: "Enterprises bolted AI onto broken processes and the bills came due without the returns. Token Fatigue is the reckoning. The fix is a business rebuilt around the intelligence, not a cheaper token." thesis: "The first wave of enterprise AI bolted models onto broken processes, and the bills came due without the returns. Token Fatigue is the reckoning. The fix is not a cheaper token but a business rebuilt around the intelligence, spending almost none of them by design." date: "2026-06-04" author: "Martijn van der Does" reading_minutes: 8 tags: - "Token Fatigue" - "enterprise AI" - "AI ROI" - "agentic transformation" - "SME" claims: - text: "In 2026 enterprises hit the AI ROI wall. Bain found only 4% achieved savings above 30%. S&P watched AI project abandonment jump from 17% to 42% in a single year." - text: "Bolting AI onto a broken process makes the broken process faster, not better. The fix is rebuilding the system around the intelligence, not layering intelligence on the surface." - text: "The cost of a given level of intelligence falls roughly tenfold a year, yet bills explode, because usage grows faster than price falls. Goldman projects token use rising twenty-four fold by 2030." - text: "The win is a specialized system: cheap models for the bulk of the work, frontier models only on the steps that change an outcome." - text: "The shake-out favors SMEs. Agility beats budget. A focused company rebuilds a process in weeks that a giant cannot touch in a year." definitions: - term: "Token Fatigue" definition: "The 2026 reckoning when enterprises realised that buying more tokens was never the same as building more value. The bills from bolted-on AI grew with use while the returns did not, and budgets began to be cut." - term: "Bolt-on" definition: "AI layered onto an existing, unchanged process: a copilot, a smarter search, a chatbot on the side. It speeds up steps that should not exist instead of removing them." - term: "Agentic-first operation" definition: "A business rebuilt around a live model of the operation, with agents owning whole processes end to end and people sitting on top of the system rather than inside it." - term: "The specialized machine" definition: "A system built for one business and nothing else. Cheap models for the bulk of the work, frontier models only on the steps that genuinely need them, the expensive call reserved for when it changes an outcome." sections: - id: "mission" title: "Mission" - id: "the-hangover" title: "The Hangover" - id: "nowhere-to-hide" title: "Nowhere to Hide" - id: "valuable-unknown-truths" title: "Valuable Unknown Truths" - id: "the-paradox" title: "The Paradox" - id: "the-specialized-machine" title: "The Specialized Machine" - id: "the-inevitable-destination" title: "The Inevitable Destination" - id: "the-honest-part" title: "The Honest Part" - id: "where-the-shake-out-lands" title: "Where the Shake-Out Lands" - id: "what-we-do" title: "What We Do" - id: "the-thing-is-i-never-left" title: "The Thing Is, I Never Left" faqs: - question: "What is Token Fatigue?" answer: "The hangover after the first wave of enterprise AI. Companies spent two years bolting generative AI onto existing processes, the bills grew with every use, and the returns did not follow. By 2026 the budgets are being cut. Token Fatigue is the realisation that buying more tokens was never the same as building more value." - question: "Why did bolting AI onto our processes not work?" answer: "Because a copilot over a broken process makes the broken process faster, not better. It speeds up the steps and automates the workarounds instead of removing the reasons they exist. You pay real money to make the wrong thing more efficient. The fix is to rebuild the process around the intelligence, not to layer intelligence on the surface." - question: "Tokens keep getting cheaper. Why is our bill going up?" answer: "Because usage grows faster than price falls. Cheaper intelligence invites careless use, and agents that loop and re-read multiply consumption. Goldman projects token use rising twenty-four fold by 2030. The answer is not a cheaper token. It is a system that uses almost none of them, by design, spending only where they pay." - question: "Are agents reliable enough to own a whole process?" answer: "Not if you point them at everything and hope. The best computer-use agents only recently matched a human, so failures still happen and compound across long workflows. The discipline is to bound the agent, redesign around its limits, put hard rules in code, and measure it on the outcome it moves. Done carelessly, agentic AI is the next hangover. Done with discipline, it is the rebuild." - question: "Why is this a better moment for SMEs than for large enterprises?" answer: "Because the advantage is agility, not budget. A focused company without decades of tangled legacy systems can rebuild a process in weeks that a giant cannot touch in a year. The shake-out lands hardest on the companies with the most to unwind and the most committees to convince." - question: "How is Parallel Partners different from an AI vendor?" answer: "We do not sell tokens, prompts, or pilots, and we do not take a fixed fee. We rebuild the business with our own capital and equity at risk, paid only when the rebuild pays. That forces us to lower the token bill and raise the output at the same time, because our upside depends on the P&L moving, not on how much you spend." --- **The real agentic transformation begins.** > "The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use." > [Sam Altman](https://blog.samaltman.com/three-observations) ## Mission 01 Rebuild SMEs into agentic-first operations. Based on conviction, built for outcomes. For the company drowning in token bills, that means one thing. Kill the bolt-on. Rebuild the business underneath it, with a live model of the operation at the center and people sitting on top of it rather than inside it. The first wave of enterprise AI was a layer. A copilot here, a smarter search there, a chatbot bolted to the side of a process nobody touched. The layer is now being ripped off, and the budgets with it. What comes next is not a smaller layer. It is a different building. ## The Hangover 02 In early 2026 the champagne went flat. The largest companies on earth had spent two years pouring money into generative AI. The bill came due and the returns did not. A Bain survey of nearly a thousand large companies in 2026 found [only four percent](https://www.insurancejournal.com/news/national/2026/06/01/871951.htm) had achieved AI savings above thirty percent, while the largest group saw cost reductions of ten percent or less. McKinsey found [only thirty-nine percent of companies](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) seeing any profit impact at all, and almost none of it large. BCG counted [five percent of firms](https://www.indexbox.io/blog/only-5-of-firms-see-meaningful-ai-returns-bcg-reports/) capturing real value out of more than a thousand. S&P watched the share of companies abandoning most of their AI projects [jump from seventeen percent to forty-two](https://www.lootzysoft.com/blog/the-state-of-ai-in-2025-closing-the-gap-between-adoption-and-impact) in a single year. This is the morning after. We call it Token Fatigue. The realisation, arriving all at once across boardrooms, that buying more tokens was never the same thing as building more value. By mid-2026 the surveys caught up to the mood. [Nearly seventy percent of executives](https://www.fairplaytalks.com/2026/05/20/seven-in-10-companies-could-slash-ai-budgets-as-roi-disappoints-report-finds/) say they are ready to cut their AI budgets if the results do not come this year. The champagne was not wasted. The first wave produced real value. Faster code, content on demand, support that scales, analysis that saves hours. What disappointed was never the usefulness. It was the measurable line from the usefulness to the profit. This is not a failure. It is a technology growing up, and growing up starts with honest accounting. ## Nowhere to Hide 03 The names tell the story better than the surveys. [Uber](https://fortune.com/2026/05/28/tokenmaxxing-is-dead-companies-didnt-get-the-roi-from-ai-they-wanted-to-see/), a company that runs on engineering, handed its engineers the best AI coding tools money can buy. Adoption went from a third of engineers to eighty-four percent in four months. The budget for the entire year was gone in those same four months. Their own technology chief said the budget he thought he would need was already blown away. Their president, asked whether all that spend was producing more useful product, admitted the line could not be drawn. The work happened. The value did not show up where it was supposed to. [Microsoft](https://fortune.com/2026/05/28/tokenmaxxing-is-dead-companies-didnt-get-the-roi-from-ai-they-wanted-to-see/), which sells AI for a living, began cutting its own engineers off from the most expensive coding agent because the bill grew faster than the benefit. When the company that sells the shovels starts rationing them internally, the gold rush has reached its honest phase. There is nowhere to hide from a bill that arrives every month and grows with use. ## Valuable Unknown Truths 04 Here is the truth the hangover reveals, the one worth knowing because so few have absorbed it yet. The bolt-on was always going to fail, and not because the models were weak. The models were extraordinary. It failed because a copilot laid over a broken process makes the broken process faster, not better. It speeds up the steps that should not exist. It automates the workaround instead of removing the reason for the workaround. You spend real money making the wrong thing more efficient. A business is not a stack of tasks waiting for assistance. It is a system. Bolt intelligence onto the surface of a system and you get a more expensive surface. Rebuild the system around the intelligence and you get a different company. That is the whole lesson of Token Fatigue, and the people who learn it first will own the decade. There is a deeper reason this matters, and it is not only about cost. When everyone rents the same intelligence, intelligence stops being the advantage. The CEO of Box, a company whose entire business is putting AI to work on enterprise data, [made the point plainly in early 2026](https://www.businessinsider.com/ai-agents-expertise-box-ceo-context-gives-companies-competitive-advantage-2026-1). The models are becoming a commodity. Expert knowledge is becoming a commodity. So the question every company will face is how you win when your competitor holds the same models you do. The answer is the thing the model cannot supply. Your data, your workflows, your institutional knowledge, and the freedom to point any model at them. The advantage was never the intelligence. It was the operation the intelligence runs on. ## The Paradox 05 The strange part, the part that breaks intuition, is that tokens got cheaper the entire time. The cost of a given level of intelligence has fallen roughly tenfold a year. Stanford measured one capability level [dropping more than two hundred and eighty times](https://hai.stanford.edu/ai-index/2025-ai-index-report) in eighteen months. Every quarter the labs hand you more capability for less money. By every reasonable expectation the AI bill should be shrinking. It is exploding instead. Goldman Sachs projects token consumption [rising twenty-four fold by 2030](https://www.goldmansachs.com/insights/articles/ai-agents-forecast-to-boost-tech-cash-flow-as-usage-soars), driven by agents that loop and re-read and call themselves again. This is the trap in one sentence. When intelligence gets cheap, careless use grows faster than the price falls. Cheaper tokens do not save a bolted-on enterprise. They invite it to waste more, faster, with a clearer conscience. The answer was never a cheaper token. The answer is using almost none of them, on purpose, inside a system designed to spend them only where they pay. ## The Specialized Machine 06 The first wave reached for the generalist. One tool, pointed at everything, mastering nothing, billed by the token for the privilege. Real work does not happen in a demo. It happens on the line, the same operation a thousand times a day, where a tenth of a cent per call multiplied across millions of calls becomes the difference between a margin and a loss. There the generalist is the wrong machine and the invoice proves it. What wins is the system built for one business and nothing else. Cheap models for the bulk of the work. Frontier models only on the steps that genuinely need them. The expensive call reserved for the moment it changes an outcome. Most of the work should never touch a premium token at all, because most of the work was redesigned so it does not have to. ## The Inevitable Destination 07 The destination is not in doubt. Only the timing is. The first wave was open-ended experimentation. Try everything, wire AI into anything that would hold still, and worry about the bill later. The second wave, the one starting now, is realism. It is the reckoning with what the first wave cost and the cold sort of what is actually worth keeping. As [one observer put it](https://x.com/Altimor/status/2062389885437366342), chapter one was broad experimentation, chapter two is rationalising its costs and finding what is sustainable. The easy phase is ending. The phase where you bought a subscription, told the board you were doing AI, and hoped the savings appeared. That phase is dead and the budgets are confirming it. What replaces it is harder and worth more. Companies rebuilt from the workflow up, with agents owning whole processes end to end, measured in business outcomes instead of tokens burned. Every business that bolted AI onto a broken process will either rebuild the process or be beaten by someone who did. There is no third path where the bolt-on quietly starts working. The question facing every owner is not whether to spend on AI. It is whether to keep renting a layer or to finally build the machine. ## The Honest Part 08 We should say what cuts against us. Agents are not magic, and the failure rate is real. Gartner expects [more than forty percent of agentic projects to be cancelled](https://www.beri.net/article/ai-agent-adoption-enterprise-2026-gartner-idc) by 2027. The honest benchmarks show the best computer-use agents only recently matching a human, which means roughly one task in four still fails, and failures compound across a long workflow. Anyone selling you autonomous everything is selling you the next hangover. This is exactly why the work is disciplined and not heroic. You do not throw an agent at a process and pray. You bound it, you redesign around its limits, you put the regulation in code and the judgment where it belongs, and you measure it on the outcome it moves. This applies to us before anyone else. Bounds, human oversight, and redesign are not a brake on the ambition. They are the condition for it. The failure rate is not an argument against agentic systems. It is an argument against doing them carelessly, which is the only way the first wave knew how. ## Where the Shake-Out Lands 09 Token Fatigue is not bad news. It is a cleansing. The hype vendors who sold prompts as products are losing their budgets. The pilots that were never going to scale are being cancelled. The money is moving from AI everywhere to AI where it matters, from spray to surgery. That is healthy, and it is overdue. The shake-out lands hardest on the giants, because they have the most legacy to unwind and the most committees to convince. It lands easiest on the small and the mid-market. A focused company without decades of tangled systems can rebuild a process in weeks that a Fortune 20 cannot touch in a year. The advantage is not budget. The advantage is the freedom to change how the work happens without asking permission from the past. ## What We Do 10 We are Parallel Partners. We do not sell tokens, prompts, or pilots. We rebuild small and mid-sized businesses into agentic-first operations with our own capital and equity at risk, and we are paid only when the rebuild pays. That alignment forces the discipline this moment demands. We cannot bolt a copilot on the side and bill you for the tokens, because a bolt-on does not move the P&L and our upside depends on the P&L moving. We have to redesign the system until it spends almost nothing per unit of work and produces more than it did before. The token bill going down while the output goes up is not a side effect of our model. It is the contract. ## The Thing Is, I Never Left 11 Twenty years in this work, a company built and sold, and the same conviction underneath all of it. The technology was never the point. The business was always the point. The first wave forgot that. It fell in love with the model and bolted it onto everything in reach, and the bill came for the infatuation. Token Fatigue is just the market remembering what the work was always about. Not more intelligence. Better businesses. The easy phase is over. The real one starts now. Back to building.