DECISION RUBRICLast verified: October 2026

Human vs Agent Swimlanes: When to Put a Step in Which Lane (2026)

The single most-asked operator question. Vendor pages duck it because their product is sold as agent-everywhere. The neutral answer is a four-criterion rubric. Each criterion routes a step to a different placement: pure agent lane, pure human lane, agent lane with human review, or agent lane with human escalation gate.

Decision-rubric swimlane. The same generic three-step process is annotated with the placement outcomes the rubric produces. Step 1 (low cost of error) sits in the agent lane. Step 2 (cost of error material, action reversible) sits in the agent lane with a human review gate. Step 3 sits in the system lane (deterministic persistence).Source: rubric synthesised from Anthropic, “Building Effective Agents” (Schluntz, December 2024) and the OECD AI Exposure Measure task taxonomy.

The four criteria

1. Autonomy required

Does the step need genuine decision-making, or is it rule-following? Anthropic's “Building Effective Agents” (Schluntz, 2024) frames the distinction as workflow versus agent. A workflow is deterministic, controlled by predefined paths. An agent is non-deterministic, choosing its own steps within a tool budget. Rule-following work belongs in either an agent lane (if a model is the cheapest implementation) or a system lane (if a deterministic engine is). Decision-under-uncertainty work benefits from a human review gate.

2. Reversibility of the action

Can the action be undone if it turns out wrong? Drafting a reply is reversible (the draft can be discarded). Sending money, filing a regulatory return, deleting production data, or sending an external email are not. Irreversible actions warrant a human gate even at high model confidence; the cost of a wrong action exceeds the cost of human review.

3. Regulatory constraint

Is human review required by regulation? Some sectors require an identifiable human in the chain regardless of confidence. The EU AI Act (Regulation (EU) 2024/1689, in force August 2024) requires human oversight for systems classified as high-risk under Annex III, which includes credit-scoring, certain recruitment uses, and administration of essential public services. The US Department of Health and Human Services regulations on clinical decision support carry analogous human-in-the-loop obligations under 45 CFR Part 162. Regulatory constraint is a hard yes/no; it does not depend on the confidence score.

4. Cost of error

What is the downside if the action is wrong? If the cost exceeds the cost of human review, review is the cheaper path even when the model is confident. Klarna is the clearest public worked example in both directions. Its February 2024 press release put two-thirds of support chats in the agent lane. By May 2025 its chief executive told Bloomberg that cost had been “a too predominant evaluation factor” in how that support was organised and that “what you end up having is lower quality”, and the company began recruiting human agents again. Cost of error here was not a regulatory limit but a service-quality one, and it surfaced only after the gate had been set too wide.

The four resulting placements

Pure human lane

The step requires autonomy and is irreversible or regulated. Examples: final hiring decision in EU jurisdictions (EU AI Act Annex III §4), high-value loan approval, clinical diagnosis sign-off. The agent does not appear in this lane; the agent may appear in adjacent lanes drafting summaries that feed the human task.

Pure agent lane

The step is rule-following or low-stakes, and the action is reversible. Classification, summarisation, intent detection, low-confidence routing. The agent acts and writes to a system of record; no human gate.

Agent lane with human review

The action is reversible but cost of error is meaningful. The agent drafts; the human reviews before send. Common in sales outreach, first-draft customer replies for sensitive accounts, and code review before merge. In BPMN, the review is a downstream bpmn:userTask.

Agent lane with human escalation gate

Most cases auto-resolve in the agent lane; ambiguous or high-stakes cases throw a bpmn:signalEvent that a human catches in another lane. The customer-support intake default. Operators rarely publish their escalation rate, and Klarna has not published one: what its February 2024 release gives is the share handled by the assistant, two-thirds of chats in the first month, which leaves roughly a third of that month's volume arriving somewhere other than the agent lane. If you are quoted an escalation rate, ask what the denominator is before you design a gate around it.

Common anti-patterns

Theatre approval

The human is asked to rubber-stamp without time or information to read the underlying material. The signature is performed; the oversight is not. The rubber-stamp is documented as an anti-pattern in the human factors literature on decision support. If the volume is too high for genuine review, the gate is in the wrong place.

Rate-limited oversight

The human becomes the bottleneck. Throughput drops, review queue grows, and the cost-benefit of the agent lane is offset by review backlog. The fix is either to remove the gate (where reversibility and cost-of-error allow), to raise the gate threshold (only escalate ambiguous cases), or to add review capacity.

Loss of oversight skill

As the agent handles more cases, the human loses the routine practice that makes the gate effective. The reviewer can no longer evaluate the agent's output. The fix is rotation: the human reviews a sampled subset (not just the escalated cases) so the skill stays calibrated. Documented in the FAA's pilot-monitoring literature on automation dependence and applies analogously here.

Reference example: customer support escalation

Klarna's February 2024 press release describes a deployment where the AI assistant handled roughly two thirds of customer service chats in its first month at scale. The release puts a broad scope in the agent lane, naming refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies, and keeps one human path open: customers “can still choose to interact with live agents if they'd prefer”. That self-selected path, not a confidence score, is the only escalation route the release actually documents.

The follow-up is the more useful half of the example for anyone placing a gate. In May 2025 Klarna reversed direction and began recruiting human agents again, with its chief executive saying that cost had become too dominant a factor and the result was lower quality, and that there must “always be a human if you want”. The assistant still carried about two-thirds of inquiries at that point. By February 2026 the model was explicitly hybrid: the agent lane for routine questions, a flexible pool of human agents for the rest, with human contact positioned as the premium tier rather than the overflow queue.

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

Handoffs : the agent-to-human handoff is the BPMN-correct shape for the escalation gate.
BPMN with AI agents : the spec-correct element reference for the gate (bpmn:exclusiveGateway, bpmn:eventBasedGateway, bpmn:signalEvent).
agenticorgchart.com / human-in-the-loop : the org-chart-shaped sister view of the same human-gate pattern.
whatisanaiagent.com / how to evaluate an AI agent : failure modes that inform the gate.