AI wins on the simple, repetitive questions. But for anything complex, its self-resolution rate drops and it becomes an expensive router to humans — and Gartner projects generative-AI cost per resolution will pass $3 by 2030, surpassing many offshore human agents. Add implementation, platform fees, and churn, and “cheap” AI can cost more than a good human team.
There’s a story everyone has heard by now: AI is going to replace call centers and slash your support costs. Parts of it are true. But the version being sold — that automation is automatically cheaper — is quietly falling apart for a lot of companies, and it’s worth understanding why before you bet your customer experience on it.
This isn’t an anti-AI argument. AI is genuinely excellent, and cheaper, for a huge slice of support. It’s an argument about a specific, under-discussed truth: the more complicated your product and the more your support actually matters, the faster AI’s cost advantage disappears — and the more valuable an affordable, skilled human team becomes.
01The sticker price vs. the real bill
On paper, AI support is astonishingly cheap. Per-resolution pricing across the major platforms looks like this in 2026:
| AI platform | Headline price | What’s on top |
|---|---|---|
| Intercom Fin | $0.99 / resolution | Helpdesk seat fees |
| Zendesk AI | $1.50–$2.00 / resolution | $55–$169/agent/mo + AI add-on |
| Salesforce Agentforce | $2.00 / conversation* | $175+/user/mo + $50k–$150k build |
| Ada | $1.00–$3.50 / interaction | ~$30k/yr contract |
| Decagon | ~$0.50 / resolution | ~$50k/yr platform fee |
*charged whether or not the issue is resolved. A dollar or two per resolution sounds unbeatable. But that number is the tip of the bill. The costs that don’t appear on the rate card are the ones that get you:
- Implementation & integration — enterprise AI deployments routinely run $50,000–$150,000 before they answer a single ticket.
- Platform & licence fees — $30,000–$50,000 a year for the serious tools, on top of per-resolution charges.
- Ongoing engineering — fine-tuning, knowledge retrieval (RAG), prompt upkeep, and guardrails never stop needing attention.
- Governance & compliance — logging, auditing, and monitoring, especially in regulated industries.
- Escalation infrastructure — the human hand-off workflows for everything the AI can’t handle.
- Vendor price normalisation — today’s AI pricing is subsidised; vendors are steadily raising it toward profitability.
02The complexity trap
Here’s the mechanism most cost comparisons miss. AI only saves money on the tickets it actually resolves on its own — its “containment rate.” For simple questions (“where’s my order?”), containment is high and AI is genuinely cheap. But the more complicated your product — technical setups, financial edge cases, multi-step troubleshooting, anything with nuance — the lower that containment rate falls.
And when containment is low, the economics invert. The AI doesn’t replace the human; it becomes an expensive extra layer in front of the human. The customer fights the bot, gets frustrated, and ends up with an agent anyway — so you’ve now paid for the AI resolution attempt and the human resolution. For a complicated product, that’s not a saving. It’s a surcharge.
03The cost that never shows up on the invoice
The biggest cost of bad AI support isn’t on any bill — it’s the customers who quietly leave. A frustrating support experience is one of the fastest ways to lose someone, and no per-resolution price captures the lifetime value that walks out the door.
In 2024, Klarna proudly announced its AI was doing the work of 700 agents, handling two-thirds of customer chats. By 2025 it had reversed course — quality had dropped, customers were dissatisfied, and the company resumed hiring humans. CEO Sebastian Siemiatkowski’s conclusion: “investing in the quality of human support is the way of the future,” and customers should always be able to reach a person. The poster child for AI-replaces-humans became the poster child for why it doesn’t.
04The crossover is already on the horizon
This isn’t just today’s problem. The trend line points the wrong way for AI-on-price. As data-centre costs rise, vendors move from subsidised to profitable pricing, and use cases get more complex, the total cost of ownership climbs.
> $3
Gartner’s forecast for generative-AI cost per resolution by 2030 — a level that surpasses many offshore human agents. The gap that makes AI look cheap today is projected to close.
05Where AI genuinely wins (let’s be fair)
None of this means AI is a mistake — using it badly is. AI is the right tool, and the cheaper one, for a large and permanent category of support: high-volume, repetitive, low-stakes questions with clear answers. Order status, password resets, store hours, simple returns — that work isn’t going back to humans, and it shouldn’t. For those interactions, AI will almost always be cheaper than a person, and faster too.
The mistake is treating AI as a wholesale replacement for human support rather than a layer of it. The companies getting burned are the ones who cut the humans first and discovered — like Klarna — what those humans were actually doing.
06The model that actually pencils out
The winning setup in 2026 isn’t “AI or humans.” It’s AI for the simple, humans for the rest — let automation clear the easy volume, and put skilled people on the complex, emotional, and revenue-critical conversations where they change the outcome.
The catch everyone hits: skilled human support in the US or UK is expensive, which is what pushed companies toward “replace it all with AI” in the first place. That’s exactly where an offshore call center changes the equation. A South African call center gives you genuinely skilled, neutral-accent human agents at roughly 40–60% below Western costs — the affordable human layer that makes the blended model work without blowing the budget. You get AI’s efficiency on the simple tickets and real people on the ones that matter, at a price that beats trying to force AI to do everything.
Deep dive What a South African call center actually costs per seat →Get the human layer, without the Western price tag
Tell us your market, seat count, and the kind of support you need. We’ll match you with a vetted South African call center that fits — free.
Get my free match →07Frequently asked questions
Is AI actually cheaper than a call center?
For simple, high-volume questions, usually yes. But the all-in cost of AI — platform fees, implementation, ongoing engineering, and human escalations — can exceed a cost-effective human team for complex or high-touch support. Gartner projects generative-AI cost per resolution will exceed $3 by 2030, surpassing many offshore human agents.
What are the hidden costs of AI customer service?
Implementation and integration (often $50,000–$150,000), platform fees ($30,000–$50,000 a year for enterprise tools), ongoing fine-tuning and guardrail engineering, governance and compliance, human-escalation infrastructure, and the revenue lost to churn when the experience is poor.
Why did Klarna reverse its AI customer service?
Klarna said its AI handled the work of about 700 agents, but quality dropped and customers were dissatisfied. In 2025 it resumed hiring human agents; the CEO said customers should always be able to reach a human and that quality human support is the way forward.
When is a human call center better than AI?
For complicated products, emotional or high-stakes issues, and brand-critical interactions. The more complex the support, the lower AI's self-resolution rate — so it often becomes an expensive router to humans, and you pay for both.
So what's the most cost-effective setup?
A blend: AI for simple, repetitive questions and skilled humans for everything that matters. A South African call center supplies that human layer at roughly 40–60% below Western costs, keeping the blended model affordable without sacrificing experience.