Gartner believes that the promise of cheap, automated customer service through generative AI is running headlong into an uncomfortable reality.
According to new research from the firm, the cost per resolution for gen AI in customer service will exceed $3 by 2030, higher than many B2C offshore human agents.
The prediction challenges the dominant narrative that's driven billions in AI investment across the contact center industry.
Patrick Quinlan, Senior Director Analyst at Gartner, explained that the issue goes well beyond the per-unit consumption costs that vendors typically emphasize.
In an exclusive interview with CX Today, Quinlan explained that “it's not just like the per unit cost from a consumption perspective, that's what vendors usually focus on, and that is often very low.
“But what then ends up happening is that many organizations consume a lot more than they expect, and then they don't account for the total cost of ownership.”
These unaccounted costs include hiring specialized AI talent that commands significantly higher salaries than traditional contact center agents, unpredictable usage patterns that blow through budgets, and a series of infrastructure cost increases that are baked into the technology's future.
The Subsidy Problem
Another significant factor driving costs upward is something most organizations haven't considered: the prices they're paying today are artificially low.
Large language model vendors are currently subsidizing their services by up to 90%, according to some estimates, as part of a growth strategy to build market share. Quinlan claims that this won't last:
“That price is subsidized to drive growth in their user base, and this is a common strategy. You enter the market with a low price, but that's going to have to change once those companies pivot to being profitable.”
The situation gets worse when you factor in newer models. While per-token costs have dropped for older models, frontier models consume three, five, or even ten times more tokens for similar interactions.
The net result is that queries on newer models end up costing more, even if the unit price appears lower.
Infrastructure Reality Check
Beyond the subsidy issue, there's a physical infrastructure problem that could be about to hit everyone's wallet.
GenAI doesn't scale the way traditional software does. Adding more users requires an almost linear increase in compute resources, which means data centers need to expand dramatically.
Those data centers need massive amounts of electricity, and the grid can't handle it.
Several U.S. states have already passed laws preventing power companies from passing infrastructure costs onto consumers after electricity bills in data center hotspots like Northern Virginia and Oregon jumped by as much as 200%.
That means LLM vendors will have to absorb those costs themselves.
“That makes their data center investments higher, the cost of electricity higher,” Quinlan said.
“At some point that's going to affect what the LLM vendor charges, which will affect what the application vendor charges.”
Some companies are even investing in small modular nuclear reactors to power their data centers off-grid because the electrical infrastructure can't keep up with demand.
Water access for cooling these facilities is another limited resource that's becoming more expensive.
Then there are the specialized AI chips required to power these workloads, which typically burn out in one to three years and need replacing.