Join my new list, CaseBliz

Everyone assumed AI would be cheaper. Goldman Sachs ran the actual numbers. The answer is not what the vendors are pitching.

The whole case for replacing contact centre agents with AI rests on one assumption: that AI is dramatically cheaper. Goldman Sachs ran its own internal cost model and found that a real-time voice AI agent costs roughly $92 per day, all-in. A human agent costs roughly $90. That's not a typo. The technology pitched as a cost-cutting play is currently more expensive than the people it's supposed to replace.

The gap exists because voice AI is genuinely harder than text. Latency requirements force premium models onto the critical path. Authentication, compliance, escalation logic, and multi-system integration all add cost that doesn't show up in a per-call benchmark. Opus Research noted that model prices will keep falling and the economics will shift — but right now, $92.90 is the honest number, not the vendor's slide deck number.

None of this means AI in service is a bad investment. Salesforce's survey of 2,025 agentic AI decision-makers, fielded through May and June this year, found that companies already running agents in production hit ROI in about eight months, with a 29% average lift in customer satisfaction. But the companies reaching those numbers share one thing: they sorted their data before they deployed. IBM's research on 1,200+ Salesforce customers found 64% can't even determine whether their AI agents are saving money or inflating costs because their total cost of ownership is unclear. They deployed without knowing what they were measuring.

The $2 daily difference between an AI agent and a human one will close. But the ROI gap between organisations with clean data and those without? That one is widening.

That's Goldman Sachs's estimated all-in daily cost for a real-time voice AI agent in a contact centre, according to internal modelling cited by AmplifAI. A human agent costs $90. The reason voice AI is more expensive than most people expect: latency requirements force premium models, and customer service calls carry compliance, authentication, and escalation overhead that text-only AI workloads don't. Goldman treats this as a snapshot, not a permanent ceiling. Costs will fall. But the current number strips out a lot of vendor optimism.

Both numbers are true. That's the point everyone is missing.

SurveyMonkey's study of 2,017 Americans found that 79% strongly prefer talking to a human over an AI agent. The same research, and a parallel Zendesk dataset, found that 51% prefer a bot when they want an immediate answer. These findings aren't contradictory. They describe the same person in two different emotional states: one frustrated, one in a hurry.

The industry keeps treating this as a binary. Either customers want humans and AI is the wrong call, or customers want speed and bots win. The actual preference is conditional. Fast resolution from any source beats slow resolution from a sympathetic one. And 81% of customers already believe companies deploy AI to cut costs, not to improve service. That's not a perception problem. That's a trust deficit companies built themselves by deploying AI on the cheap and calling it innovation. The technology isn't the issue. The motive is.

Going back to re-read every AI cost estimate I've trusted this year. The Goldman $92 number has me questioning all of them.