― Human + AI Delivery Model
The Wrong Distribution Is Expensive
Put a human specialist on routine volume and the operation costs too much. Put AI on a judgment call and the outcome is wrong. The delivery model exists to prevent both: matching the right capability to each case so routine work moves fast and complex work gets the expertise it requires.
That matching is not static. As the Learning Loop captures patterns, work that once required specialist judgment becomes automatable. The delivery model is designed to shift: L1 absorbs more, L2 focuses on harder cases, and L3 shapes the rules that make the whole system sharper. The economics follow the shift.
L1: AI Workers
Handle volume and resolve routine exceptions inside defined rules.
L1 AI workers handle the bulk of routine workflow execution: classification, data extraction, pattern matching, and resolution of cases that fall within set parameters. When a known exception type has a documented resolution path, L1 handles it without escalation, with the resolution logged for traceability. A duplicate invoice resolved through a matching rule, a missing field populated from a known source, a routing correction based on established criteria: these are resolutions, not just triage. The more volume L1 handles, the more specialist time concentrates on the cases where judgment changes the outcome.
When an output falls below confidence thresholds, triggers an approval gate, or encounters missing context that L1 cannot resolve from available data, the case escalates to L2. L1 does not make judgment calls. It resolves what the rules cover and flags what they do not.
L2: Human Specialists
Resolve exceptions where judgment determines the outcome.
L2 specialists resolve the exceptions that L1 cannot handle: ambiguous cases, incomplete context, policy interpretation, edge cases that fall outside the playbook. They operate within documented guardrails but apply judgment that AI cannot yet replicate.
What makes L2 valuable is process knowledge. An L2 specialist with finance operations experience resolves an accrual exception differently than someone working from a generic playbook. One with marketing operations experience can spot a data-quality issue that an AI flags as a model-confidence problem. The judgment at L2 comes from knowing how the business process actually works, not just how the workflow is configured.
L2 specialists also surface patterns back into the system. When an L2 specialist resolves the same benefits-enrollment exception three times in a week because the eligibility rules do not account for mid-year transfers, that pattern gets flagged for the Learning Loop: the eligibility rule gets updated, and the next transfer clears at L1 instead. Every resolution that follows that path makes the operation faster and less expensive.
L3: Domain Experts
Shape the rules the operation runs on.
L3 domain experts handle the cases with the highest stakes and the most complexity: strategic edge cases, process redesign decisions, policy exceptions that set precedent, and situations where the business context requires senior judgment.
L3 also shapes the operating model itself. When a customer service workflow keeps escalating loyalty-discount exceptions because the existing rules cannot handle how promotional offers and loyalty rewards interact, L3 redesigns the discount policy. The thresholds, the escalation criteria, the playbook standards: these are L3 outputs. Every rule L3 writes expands what L1 and L2 can handle without escalation.
The Ratio Shifts as the Operation Matures
Early in an operation, a significant share of exception volume requires L2 or L3 involvement. The playbooks are thin. The rules are new. The Context Layer is still being populated with business logic that the team discovers through operating, not through planning.
As the Learning Loop captures repeated patterns, that distribution changes. Rules get tighter. Routing improves. Playbooks get sharper. Exception types that once required specialist judgment become automatable. L1 absorbs what specialists used to handle manually, and specialist time concentrates on the cases where domain expertise changes the outcome.
This is the mechanism behind the cost curve described in Build, Run, Improve: each improvement cycle reduces the share of exceptions that require human involvement. The operation gets less expensive per unit of throughput as it matures, and the maturity comes from running the work, not from planning it.
Rate Versus Volume
A lower exception rate does not always mean less exception work.
When throughput grows faster than the exception rate improves, the absolute number of exceptions still rises. Better accuracy at higher volume can mean more exception work, not less. The delivery model must handle both dimensions: improving rates while absorbing growing volume
This is why the tiered structure matters at scale, not just at launch. L1 handles the volume increase. L2 and L3 handle the complexity that remains. The Learning Loop keeps shifting work downward so the cost of handling that volume drops even as the volume itself grows.
The wrong distribution is expensive. The right one gets less expensive as the operation matures.
Whether the team needs a managed operation or is building the capability internally, the distribution model shapes everything that follows.
