Gartner expects generative AI in customer service to cost more than $3 per resolution by 2030. Many B2C offshore human agents cost less than that per resolution today, so the familiar plan of swapping low-cost support teams for chatbots may not save money at all.

What Gartner found
The analysis was published in late January 2026. Its central point is simple: the price organisations pay for AI today does not reflect what it will cost to run at scale later in the decade. Anyone building a customer service business case on current AI pricing should treat that number as temporary.
Why the cost per resolution is rising
The biggest factor is pricing. LLM vendors are subsidising their services by up to 90% to build market share. Patrick Quinlan of Gartner put it plainly: "That price is subsidized to drive growth in their user base... that's going to have to change." When the subsidies end, enterprises will pay closer to the real cost of inference.
The infrastructure behind that inference is expensive and getting more so. Data centre expansion needs nearly linear compute growth for each additional user. Electricity prices have risen so sharply in some regions that states have capped increases to prevent 200% jumps. Water for cooling is becoming scarcer and more expensive around the world.
The hardware also wears out quickly. Specialised AI chips typically burn out in one to three years and need replacing, compared with five to seven years for traditional servers. That shorter depreciation window feeds straight into the price per query.
Newer models make things worse. Frontier models use 3 to 10 times more tokens than older models for similar tasks, and customers increasingly expect the more capable experience.
People cost money too. Specialised AI talent earns significantly more than traditional customer service managers. Running these systems well needs data engineers, ML ops specialists and prompt engineers, roles that did not exist at scale five years ago and that are hard to hire for. On top of that, organisations often underestimate how much compute a deployment will consume once real customers start using it.
How companies are actually using AI in service
Only 20% of customer service leaders have reduced staffing because of AI. Most report steady or higher headcount while supporting more customers, which suggests AI is helping agents rather than replacing them.
Gartner also predicts that half of the companies that cut staff because of AI will rehire by 2027. Meanwhile, 10% of Fortune 500 companies are expected to double their customer service spending so they can use AI for hyperpersonalised, proactive service. Those firms are spending more to offer a better service, with cost savings a secondary concern.
The uses Gartner rates most highly are assistive: triage (collecting information before handing a case to a person), summarisation, note-taking and intent classification. Each makes a human agent faster without asking the AI to handle the whole case alone.
The calculation for UAE organisations
Many organisations in the UAE and the wider GCC run customer service on a mix of local staff and offshore support from South Asia. If AI costs more per resolution than those trained agents, the case for replacing them gets much weaker.
Climate adds to the pressure. Gartner highlights electricity and cooling as cost drivers, and data centres in the Middle East face higher cooling costs than those in cooler climates. That could make AI operations even more expensive locally.
Cost per resolution on its own is also the wrong yardstick. A $3 AI resolution that fully fixes the problem and leaves the customer satisfied may give a better return than a $2 human resolution that needs a callback. The reverse holds as well: an expensive AI interaction that frustrates the customer and still ends with a human escalation has negative value, however advanced the technology.
A team preparing a business case should compare AI and human service on cost per fully resolved case, counting escalations, callbacks and repeat contacts, and model AI pricing without today's subsidies. Then start with the assistive uses above, where the value is easiest to measure, before committing to full automation.