Klarna's AI assistant did the work of 700 agents, then the company started rehiring humans. The real story is that support AI succeeds in narrow, policy-bounded tiers and breaks hard the moment it touches judgment calls and liability.
In February 2024, Klarna announced that its OpenAI-built AI assistant was doing the work of 700 full-time customer service agents, handling two-thirds of all support chats in its first month and resolving them in under two minutes instead of eleven. Fifteen months later, in May 2025, Klarna's CEO Sebastian Siemiatkowski told Bloomberg the company was rehiring human agents, saying the drive toward "the cheapest possible solution" had cost it on quality and that customers wanted "the option to speak to a human." Both statements were true. That's the actual state of AI in customer support: it works decisively well for a specific slice of the ticket volume, and degrades fast the moment it's stretched past that slice.
The useful question isn't "does AI work for customer support" — it's where in the support funnel it works, and where the economics of getting it wrong (a wrong refund, a wrong policy statement, an angry customer escalating on social media) outweigh the savings. That boundary is now reasonably well mapped, because enough companies have run the experiment in public.
The clearest win is self-service resolution for narrow, well-defined, low-stakes requests: order status, password resets, shipping delays, return eligibility against a published policy, subscription cancellation flows. This is where Intercom Fin built its business — Fin launched in 2023 as a support bot layered on Intercom's existing helpdesk, and its distinguishing move wasn't the model, it was the pricing: $0.99 per resolution, not per seat. Intercom was betting that if the bot actually closed tickets without human involvement, customers would pay per outcome rather than per license, and that bet forced Intercom to be honest internally about resolution rates rather than burying a mediocre deflection number inside a flat subscription fee.
Decagon, a newer entrant used by companies like Duolingo and Notion, plays the same game: an AI agent that sits in front of the support queue and only escalates what it can't confidently close. The pattern across these vendors is consistent — deflection rates in the 50-70% range for FAQ-shaped, policy-bounded queries, with the remainder routed to humans rather than force-resolved. The bots that work well are the ones tuned to escalate aggressively when confidence drops, not the ones tuned to look autonomous.
This tier works because the failure mode is cheap and reversible. If a bot fails to resolve a "where's my package" query, the customer waits a bit longer for a human. The blast radius of a wrong answer is small.
The second application area gets less press but has better unit economics: AI as a copilot for human agents rather than a replacement for them. This is what Salesforce Agentforce and Zendesk's AI features mostly ship — real-time suggested replies, automatic ticket summarization so an agent doesn't have to re-read a twelve-message thread, and retrieval over the knowledge base so the agent isn't hunting through internal docs mid-call. Sierra, the company Bret Taylor and Clay Bavor founded in 2023, sits at the boundary of both models — its agents can resolve tickets autonomously for customers like SiriusXM and ADT, but Sierra's own public positioning leans heavily on the agent staying inside guardrails defined by the client, with human escalation as a first-class path rather than a fallback.
Agent-assist doesn't get the "we replaced N humans" headline, but it's the safer bet precisely because a human stays in the loop on judgment calls. The AI drafts, retrieves, and summarizes; a person still decides whether a refund exception is warranted. It's a smaller productivity claim and a much more defensible one.
Intercom's per-resolution pricing wasn't just a go-to-market choice — it changed what the product team had to optimize for. A seat-based support tool's business model still works even if the AI feature is mediocre, because the buyer already pays for seats. An outcome-based model only works if the bot genuinely resolves tickets, because that's the only thing being billed. That's a meaningfully different incentive structure, and it's part of why the vendors most exposed to actual resolution accuracy (Intercom, Decagon) publish deflection-rate numbers more readily than vendors selling a seat-priced "AI-enhanced" add-on.
It also creates a new failure mode worth watching: a vendor incentivized to count a resolution as billable the moment the customer stops replying, whether or not the underlying issue was actually fixed. "Resolved" and "customer gave up" can look identical in a ticket log. Any team evaluating one of these tools should ask for the definition of "resolved," not just the resolution rate.
The clearest cautionary case is Moffatt v. Air Canada (Civil Resolution Tribunal of British Columbia, February 2024). Air Canada's website chatbot told a customer he could apply for a bereavement fare discount retroactively; the airline's actual policy required the request before travel. Air Canada argued the chatbot was "a separate legal entity responsible for its own actions" — a defense the tribunal rejected outright, holding the airline liable for what its own chatbot told a customer, full stop. The ruling didn't turn on novel AI law; it applied ordinary principles of agency and misrepresentation. A company is on the hook for what its automated systems tell customers, regardless of whether a human or a model generated the statement.
That's the structural reason full autonomy stays rare for anything beyond narrow, policy-bounded requests: the cost of a hallucinated policy statement isn't "the bot was wrong," it's "the company is contractually and sometimes legally bound by what the bot said." Klarna's 2025 walkback points at the same wall from a different angle — not legal liability, but the accumulated cost of degraded quality on retention and trust, which showed up slowly enough that it took over a year to become undeniable in the numbers.
Under the hood, most of these systems aren't just chat wrappers — resolving a real ticket means calling out to order, billing, and refund systems, which is exactly the kind of tool-calling problem protocols like MCP were built to standardize, so the agent isn't hand-rolling a bespoke integration against every backend it needs to touch.
| Application area | Maturity | Representative products | Typical failure mode |
|---|---|---|---|
| FAQ / order-status deflection | High — production-proven | Intercom Fin, Decagon | Over-escalates on ambiguous phrasing (cheap failure) |
| Agent-assist / copilot | High — low-risk, widely adopted | Salesforce Agentforce, Zendesk AI, Sierra | Suggests a reply the agent doesn't vet closely enough |
| Autonomous policy exceptions / refund judgment calls | Low — mostly avoided | — | Hallucinated policy creates real liability (Air Canada) |
| Full conversational replacement of human tier | Medium, contested | Klarna's 2024 rollout, later partially reversed | Quality erosion invisible until retention data lags in |
AI customer support isn't one capability with one maturity level — it's at least four different products wearing the same UI. Deflection for narrow, policy-bounded requests is mature and the economics are proven; the outcome-based pricing models built around it are a reasonable signal of vendor confidence, as long as you interrogate what counts as a "resolution." Agent-assist for human reps is the least flashy and the most durable bet, because a person still owns the judgment call. The failure mode to actually worry about isn't a bot being unhelpful — it's a bot being confidently, liably wrong on something a court will later treat exactly like a human employee's promise.