Gartner predicts AI will cost more per resolution than offshore human agents by 2030. LLM vendors are subsidizing up to ninety percent of costs. Companies made permanent restructuring decisions on temporary pricing. The invoice has not arrived yet.
On January 26, Gartner published a prediction that inverts the central assumption behind every AI restructuring announcement of the past year: by 2030, the cost per resolution for generative AI in customer service will exceed three dollars — higher than what many companies pay offshore human agents to resolve the same issues.
Three dollars does not sound like much. But for a company that just fired its customer service department because AI was supposed to be cheaper, three dollars per resolution is the difference between a strategic transformation and a write-down.
The Ninety Percent
The current price of AI is not the real price of AI.
LLM vendors are subsidizing their services by up to ninety percent, according to industry estimates cited in the Gartner analysis. This is a market-share strategy, not a pricing strategy. The playbook is familiar: price below cost to build dependency, then adjust once switching costs are locked in.
Uber ran the same play. The company lost over thirty billion dollars in its first decade, subsidizing rides below cost to build market share. Average Uber prices rose ninety-two percent between 2018 and 2021 once the subsidy phase ended. Riders who had reorganized their lives around cheap rides — sold second cars, moved to transit-poor neighborhoods, abandoned taxi accounts — absorbed the increase because the switching costs were real.
The LLM subsidy is structurally identical. Companies are building workflows, training employees, rewriting processes, and — critically — firing the humans those workflows replaced, all on pricing that reflects a vendor's growth strategy, not the actual cost of inference. When the vendor pivots from growth to profitability, the price adjusts. But the humans are already gone.
The difference between Uber and LLM vendors is the irreversibility of the customer's decision. When Uber raised prices, riders could take a taxi. When AI vendors raise prices, the customer service team that was laid off cannot be reconstituted by changing a subscription tier.
The Token Paradox
Per-token inference costs have dropped roughly tenfold annually. GPT-4 equivalent performance that cost twenty dollars per million tokens in late 2022 now costs around forty cents. This is the number that appears in pitch decks, board presentations, and restructuring memos. It is also the wrong number.
The right number is cost per resolution — what it actually costs to solve a customer's problem end to end. And that number is going up, not down, for three reasons.
First, frontier models consume three to ten times more tokens per interaction than their predecessors. The models are smarter, but they are also hungrier. A customer service interaction that required a few hundred tokens on GPT-3.5 may require thousands on a frontier model that reasons through the problem, considers edge cases, and generates a more nuanced response. The unit price fell. The units consumed exploded.
Second, the easy cases were automated first. The interactions that remain — and the ones companies are now trying to automate — are the complex, ambiguous, multi-step problems that consume the most tokens and fail the most often. Gartner's Patrick Quinlan put it directly: AI simply isn't mature enough to fully replace expertise.
Third, and most fundamental: generative AI is non-deterministic. The same prompt produces different outputs. A resolution that works once may not work the next time. Failed resolutions generate rework, escalation, and the most expensive outcome of all — a customer who has to explain the problem again to a human who was not supposed to be needed. The cost of unreliability compounds on every interaction the system handles.
The Hardware Clock
The infrastructure behind AI inference is not a one-time capital expenditure. It is a recurring obligation with a clock that runs faster than most corporate planning cycles account for.
AI chips have a functional lifespan of one to three years. Not because they break — because they become obsolete. Each generation of model demands hardware that did not exist when the previous generation was deployed. The chip that runs today's frontier model will not run next year's. This is not depreciation in the accounting sense, where an asset gradually loses value over its useful life. It is replacement — full-cost, every cycle.
The facilities that house these chips are under their own pressure. Data center electricity costs have increased by up to two hundred percent in hotspots like Virginia and Oregon. States that once welcomed data center construction are now blocking power companies from passing infrastructure costs to residential consumers, which means the data center operators absorb the increase directly.
The five largest US technology companies are expected to spend roughly ninety percent of their operating cash flow on capital expenditure in 2026. Trillions of dollars are being deployed to build and maintain the physical infrastructure of AI — and that infrastructure requires continuous reinvestment, not maintenance.
For the companies using AI to replace human workers, these costs are upstream and invisible. They show up as a subscription fee that has, so far, reflected vendor subsidies rather than vendor economics. When the subscription price adjusts to reflect the actual cost of chips that last two years, electricity that costs twice what it did, and models that consume ten times the tokens — the adjustment will not be incremental.
The Twenty Percent
Here is the number that reframes the entire restructuring narrative: as of October 2025, only twenty percent of customer service leaders had actually reduced headcount due to AI.
Twenty percent. Not the fifty or seventy percent you might infer from the headlines. Not the sweeping transformation that earned Block a twenty-four-percent stock premium when it cut nearly half its workforce. Not the signal that thirty-five CEOs sent when they named AI as the reason for layoffs totaling over twenty-two thousand displaced workers in 2026.
The gap between the headline and the survey is the gap between narrative and operations. Most companies that deployed AI in customer service used it to handle more volume with the same headcount, not to cut headcount. They found what Gartner's Emily Potosky described plainly: Relying solely on AI right now is premature.
The companies that did cut — the ones whose announcements moved markets — made a different bet. They bet that AI would be cheaper than humans not just now, when vendors subsidize ninety percent of costs, but permanently. They bet that frontier models would get cheaper per resolution, not just cheaper per token. They bet that the subscription price they pay today reflects the mature economics of AI, not the introductory pricing of a market-share land grab.
The twenty percent who cut are the ones holding the invoice when it arrives.
The Convergence
Two independent research firms have converged on the same prediction: by 2027, half of the companies that cut staff due to AI will rehire.
Gartner and Forrester arrived at this number through different methodologies and different data sets. The convergence is the signal. When two firms that compete for the same clients independently conclude that the AI labor substitution is going to reverse at scale, the question shifts from whether to at what cost.
The cost will not be the original salary. When a company fires a department and then, eighteen months later, tries to reconstitute it, the labor market has repriced. The experienced workers have moved on, retrained, or retired. The remaining talent pool is thinner and more expensive. Gartner's own analysis notes that regulatory pressure — rules requiring human agent availability — will force rehiring at premium salaries. Assisted service volume is projected to rise thirty percent by 2028 as customers, given the choice, default to requesting a human.
This is the full shape of the invoice. The original cost saving was calculated against subsidized AI pricing. The rehiring cost is calculated against a depleted labor market. The spread between the two is the actual price of the restructuring — and it is larger than the salary line item that was cut.
The Announcement and the Arithmetic
The market has developed a clear reward function for AI restructuring. Block eliminated nearly half its workforce and surged twenty-four percent. The market did not ask whether Block's AI costs were subsidized. It did not model the per-resolution economics of frontier models versus offshore agents. It did not discount for the non-determinism that makes every automated resolution a probability rather than a certainty. It rewarded the announcement.
Goldman Sachs says AI contributed basically zero to US GDP in 2025. The St. Louis Fed found AI-related job postings surged five hundred percent while measurable productivity impact remained negligible. The five largest tech companies are spending ninety percent of operating cash flow on AI infrastructure while the economic return remains, by Goldman's measure, approximately nothing.
There is a name for the gap between what the market rewards and what the economics deliver. It is not a bubble — bubbles are about price. It is a subsidy — specifically, a subsidy on the cost of AI that makes permanent restructuring appear rational on a spreadsheet built with temporary numbers.
The invoice is the moment the temporary numbers become permanent. Gartner says that moment arrives by 2030 — three dollars per resolution, more than offshore humans. Forrester and Gartner both say half the companies that cut will reverse course by 2027. The twenty percent who actually reduced headcount will discover whether the market that rewarded the cut also rewards the rehire.
The restructuring was priced. The invoice has not been.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)