AI Agent Customer Support Intake Workflow: Cited Swimlane (2026)

Customer support intake is the most-deployed process for AI agents at consumer scale. The canonical pattern is a tier-1 classifier and replier handling routine queries, with a strict escalation path for ambiguous, high-stakes, or out-of-scope cases. The constraint that shapes the process is volume: tier-1 must handle the long tail at near-zero marginal cost, and the human queue must absorb only the residual.

One named case study

Klarna's AI assistant launched at scale in February 2024. The press release reports that the assistant had 2.3 million conversations in its first month, two-thirds of Klarna's customer service chats, doing “the equivalent work of 700 full-time agents”. It was available in 23 markets, around the clock, in more than 35 languages, with OpenAI as the underlying model provider. The release also gives the two metrics an operator would ask for: resolution time fell from 11 minutes to under 2, and repeat inquiries dropped 25 per cent, with customer satisfaction described as on par with human agents.

The scope put in the agent lane was wide. The release names refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies as things the assistant manages, which puts payment queries inside the agent lane alongside the rest. The one human path the release documents is customer-initiated: customers “can still choose to interact with live agents if they'd prefer”.

What happened next is the part worth modelling. In May 2025 Klarna reversed direction and started recruiting human agents again. Its chief executive told Bloomberg that cost had been “a too predominant evaluation factor when organizing this” and that “what you end up having is lower quality”, adding that it was critical customers know “there will be always a human if you want”. The assistant still handled about two-thirds of inquiries at that point; what changed was the standing of the human lane. By February 2026 the company was running a deliberately hybrid model, using the agent lane for routine questions and a flexible, remotely-recruited pool of human agents for everything where the relationship is the product. That is a cost-of-error finding from the human vs agent decision rubric arriving late, after the gate had been set too wide.

Sources: klarna.com / klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month (Klarna press release, 27 February 2024); Bloomberg (8 May 2025); Customer Experience Dive (9 May 2025) and Customer Experience Dive (20 February 2026). All accessed October 2026.

Where the human gates sit

One gate appears in the diagram: the bpmn:exclusiveGateway after the classifier, routing cases by model confidence. Confidence routing is our modelling of the pattern rather than a disclosed internal; operators at this scale do not publish their thresholds. The second route into the human lane is the one the source does document, and it is not a decision the process owns at all: the customer asks for a person. In BPMN that is an interrupting event raised from the customer pool, not a gateway in the service pool, which is why it is easy to leave out of a diagram and easy to leave out of a deployment.

Where the handoffs sit

The escalation is the canonical bpmn:signalEvent handoff documented on the handoffs page. The reply from the agent back to the customer is a bpmn:messageFlow across the customer pool boundary. There is no agent-to-agent handoff in this simplified intake; in production, intent-specific sub-agents add an internal agent-to-agent handoff before the response is composed.

Workforce-impact note

The Klarna note frames the throughput as equivalent to the workload of 700 full-time agents. Whole-role replacement claims at this scale belong on a calculator built for the question; aijobimpactcalculator.com covers the defensible methodology (task-level, not role-level).

Related pages

• Human vs agent swimlanes : the rubric that places each gate.
• Handoffs : the escalation pattern in detail.
• BPMN with AI agents : the spec-correct element reference.
• agenticorgchart.com / supervisor-pattern : the org-chart-shaped sister view.