The year the euphoria met the cost spreadsheet
In 2023 and 2024, the dominant narrative about generative AI was one of weak demand: hyperscalers were investing hundreds of billions of dollars in infrastructure, but real revenue still wasn't keeping pace. In 2025, with the arrival of autonomous agents such as Claude Code and Codex, the story flipped — demand began to outstrip the supply of compute, and the big labs' servers were at times running at capacity.
But in 2026 a third chapter began to unfold, and it is the one that matters most to any company weighing how (and how much) to invest in AI: the "wait, is this worth what we're paying?" phase.
The most emblematic case is Uber. According to reporting by The Information, the company's CTO, Praveen Neppalli Naga, revealed that Uber burned through its annual AI budget for engineering teams in just four months, after aggressively encouraging adoption through internal leaderboards and usage targets. Microsoft, for its part, began restricting access to Claude Code and pushing engineers toward its own in-house tool (GitHub Copilot CLI), officially as a product strategy decision but, according to sources cited by The Verge, mainly because the cost had become unsustainable relative to productivity gains that did not materialize in the same proportion.
These are not isolated cases of poor management. They are the logical outcome of a technical and economic phenomenon that every company — including Brazilian SMBs (small and mid-sized businesses) — needs to understand before signing its next AI license contract: the economics of agents are fundamentally different from the economics of chatbots.
Why agents cost orders of magnitude more than chatbots
The difference between a chatbot and an agent is not just the interface. A chatbot receives a question and returns an answer: one transaction. An agent runs in a loop: it plans, calls tools, retrieves results, updates its context and plans again. Every pass through that loop consumes tokens, the unit of compute that AI platforms bill for.
According to SemiAnalysis data cited by journalist Derek Thompson, a typical agent task consumes about 96,000 tokens before producing an answer — more text than the entire novel "The Great Gatsby." Tom's Hardware reports that agentic systems can consume up to a thousand times more tokens than a standard LLM query,
depending on the number of steps needed to carry out the instruction. Goldman Sachs projects that token consumption by agentic AI could grow 24-fold by 2030.
The practical result: according to Ramp, average corporate spending on AI tokens is already 13 times higher than it was in January 2025. In some documented cases, a single company's monthly token spend exceeded the cost of an entire team of senior engineers — precisely the category of professionals AI promised to replace.
The original mistake: confusing "AI usage" with "productivity"
To justify this growing spend, several companies adopted a dangerous logic: creating internal metrics that reward the volume of AI usage rather than the outcome. At Meta, internal leaderboards appeared that measured token consumption by team. At Amazon, according to reports, employees admitted to using AI for unnecessary tasks just to inflate adoption metrics, a practice that came to be called "tokenmaxxing."
Faros AI measured the side effect: under high adoption of agentic AI, "code churn" — the ratio of lines of code deleted to lines added — rose by more than 800%. In other words, AI generated a lot of code, but much of it had to be thrown away or redone. More work, not less.
This is the central point any leadership team needs to internalize: tokens consumed are not a proxy for value delivered. A company can be spending a fortune on AI while generating more rework than it did before, a pattern that Box CEO Aaron Levie described as a kind of collective "AI psychosis" among executives: the belief that more AI is always better, that every line of code an agent generates is production-ready, and that every prototype is a billion-dollar idea.
The numbers nobody sees on the energy bill (but that come back as inflation)
There is also a macroeconomic layer that amplifies the problem, and it is worth understanding even if your company has never run an agent: the cost of AI infrastructure is being partly passed on to the entire consumption chain.
According to Goldman Sachs, US electricity prices rose 6.9% in 2025 — more than double overall inflation — driven by demand from data centers, which already account for 40% of the growth in US electricity demand. In regions with a high concentration of data centers, short-term wholesale power prices rose as much as 267% over five years, according to Bloomberg.
The same thing is happening with memory. DRAM contract prices rose between 80% and 95% in a single quarter, according to TrendForce and Counterpoint Research, as manufacturers such as Samsung, SK Hynix and Micron shift capacity toward the
high-performance memory chips used in AI servers. IDC projects price increases of 10% to 20% for laptops, tablets and smartphones by the end of 2026.
The point is not alarmism; it is context. AI operating costs do not sit isolated in an IT budget line; they ripple through the cost structure of the entire economy, including your company's, via energy, hardware and cloud services. Ignoring this dynamic when planning AI investments means planning with incomplete data.
The right question is not "will the bubble burst?"
It is tempting to read these signals — canceled licenses, blown budgets, failed pilots — as proof that AI is an overhyped fad. But that reading misses the mark.
The study "The GenAI Divide," from MIT's Project NANDA, is the most relevant data point here: despite US$30 to 40 billion in corporate investment in generative AI, 95% of enterprise pilots produced no measurable impact on the bottom line. Only 5% extracted significant value. And MIT was clear about the cause: the problem is not the models, it is organizational integration — poorly designed workflows, missing governance, and the confusion between "tool available" and "process redesigned."
Doug O'Laughlin of SemiAnalysis offers the most useful historical frame: every new technology goes through a cycle of (a) underinvestment, followed by (b) overinvestment, followed by (c) an exaggerated downward correction, repeated until organizations find the right balance between labor cost and technology cost. Many companies are clearly in phase (b) turning into (c), and that does not invalidate the technology. It invalidates the way it was adopted.
Why "replacing people with AI" is still the wrong bet
The direct-replacement narrative — "let's lay people off and automate" — is the one that shows up most in headlines, and it is also the one the data contradicts most strongly.
First, because the real cost of running AI at scale already exceeds, in several documented cases, the cost of the payroll it would replace. Bryan Catanzaro, Nvidia's vice president of applied deep learning, summed up the paradox to Axios: "For my team, the cost of compute is far beyond the costs of the employees." Companies that carried out mass layoffs betting on AI — Meta cut about 8,000 positions while committing more than US$100 billion to AI infrastructure — are discovering that the bill for "automated intelligence" (as O'Laughlin calls it) is a new, permanent category of operating cost, not a cheap substitute for people.
Second, and more important for the architecture of any business: agentic AI works best as a capacity multiplier inside a process designed by humans, not as an autonomous substitute for human judgment. When a company tries to remove human oversight entirely from a complex workflow — high-stakes customer service, credit decisions, logistics delivery — the most common result is not efficiency but expensive error. A recent, concrete example: Pizza Hut faced a US$100 million lawsuit from franchisees after an AI-operated delivery system pushed delivery times from under 30 minutes to more than 45.
The structural lesson is simple: AI that replaces humans in processes where errors cannot be corrected (the end customer) tends to shift the cost of the error onto the customer, and the customer hands that cost back in the form of reputation damage, churn or litigation. AI that augments human capacity in internal, repetitive and well-defined processes tends to generate real, measurable gains.
It is aesthetically appealing to watch an AI agent making prospecting calls, qualifying leads like an SDR or answering support tickets with natural-language fluency. The problem is that nobody puts on the same screen, next to the demo, the two questions that should precede any implementation decision: what is this agent's conversion rate compared with a properly trained human professional? And what is the real cost-benefit of the automation once you account for cost per token, error rate, rework cost and the impact on customer perception? The honest answer today is that in consultative sales, mid-to-high-ticket SDR work and exception handling in support — contexts that require reading nuance, handling objections and building relationships — the numbers rarely favor full replacement. That equation may change in two or three years, with more capable models and lower inference costs. But making organizational architecture decisions based on where the technology might get to, rather than where it is, is exactly the mistake that produced the blown budgets of 2026.
Where AI really delivers ROI — according to the data, not the hype
McKinsey's 2026 analysis identified productivity gains equivalent to 2.8% to 4.7% of revenue in banking and financial services, and 2.6% to 4.5% in pharmaceuticals and advanced manufacturing: modest numbers, but real and measurable. The pattern that recurs in the success cases of the MIT study itself is consistent: the projects that work are operational and narrow — code automation, document processing, initial triage, data reconciliation — not the broad, ambitious "visibility" projects that dominate the list of failed pilots.
For an SMB that is structuring its enterprise architecture and deciding where to invest in AI, this distinction is the starting point of any responsible roadmap. In practice, there are three adoption zones with very different risk and return profiles:
- Zone 1 — Internal operational core (high priority). Internal, repetitive processes with clear rules and existing human oversight: document triage and classification, first drafts of reports, data reconciliation across systems, drafting contracts or proposals, software development support with human code review, exploratory data analysis ahead of decision-making. Here, an AI error is cheap, because it is reviewed before it affects any external party, and the speed gain is direct.
- Zone 2 — Supervised marketing and sales activities (medium priority, with governance). Customer or supplier touchpoints where AI acts as a copilot, not a replacement: customer service reply drafts reviewed by a person before sending, lead pre-qualification that a salesperson confirms, marketing content reviewed by a brand owner. The value is there, but it requires an explicit human checkpoint in the process, exactly the kind of governance MIT found missing in the 95% of pilots that failed.
- Zone 3 — Autonomous decisions facing the end customer (low priority, high risk). Full replacement of human judgment in interactions with direct consequences for the customer: automatic credit approval, unsupervised logistics dispatch, customer service that decides refunds or cancellations without review. This is precisely the zone that produced the most expensive public failures, because the cost of an edge-case error is not contained within the company; it spreads to the customer, and the customer hands that cost back amplified.
The general rule, therefore, is the reverse of what many platforms' marketing suggests: start from the inside out, not from the outside in. The operational core is where AI delivers the highest return per dollar invested with the lowest reputational risk, and it is, paradoxically, the zone that gets the least strategic attention, because it is not "visible" to customers or investors.
The real trend from 2026 onward: working together, not replacement
The consensus emerging among analysts who follow this transition closely, which the Uber and Microsoft cases illustrate unintentionally, is not "less AI" but "better-governed AI." The companies that pulled back did not abandon the technology; they recalibrated its use, going back to treating agents as tools that operate under defined human oversight, with limited scope and outcome metrics (not consumption metrics).
For an SMB, this translates into a concrete architecture recommendation: before expanding your use of AI, map your internal processes against two criteria — (1) is the cost of an error in this process contained internally, or does it spread outside the company? and (2) is there a human review point in this process today, or would AI be operating without a safety net? Processes that combine "contained error" and "existing human review" are immediate candidates for AI automation. Processes that combine "externally propagated error" and "no review" are candidates for controlled pilots, not broad implementation, no matter how impressive the tool's demo is.
The 2026 bill is being presented to the companies that skipped this step. The opportunity, for those still designing their adoption architecture, is to avoid repeating the same original mistake: treating AI as a substitute for people instead of treating it as what it really is — a new layer of operational capacity that only pays off when it is integrated into a process designed, with governance, by humans.
The 2026 bill is being presented to the companies that skipped this step. The opportunity, for those still designing their adoption architecture, is to avoid repeating the same original mistake: treating AI as a substitute for people instead of treating it as what it really is — a new layer of operational capacity that only pays off when it is integrated into a process designed, with governance, by humans.
Sources: Derek Thompson (derekthompson.org), Tom's Hardware, Outlook Business, AI2Work, and data cited from SemiAnalysis, Goldman Sachs, MIT Project NANDA, McKinsey, Ramp, Faros AI and Bloomberg.
