Redesign a Process Around Agents with Autonomy Tiers, Exceptions, and Escalation

Workflow area: AI Requirements & Agent Redesign · Tier: Advanced · Hands-on time: ~50 minutes

Workflow redesign is the strongest predictor of AI impact in the McKinsey data, and it is the step most organizations skip — only about a fifth redesign end-to-end. The reason is that redesign is analyst work, and most AI initiatives are staffed with engineers and a vendor. This playbook is the BA claiming that gap: the to-be process is a model, autonomy is a per-step design decision with entry criteria, and the exception path is not an afterthought — in an agent process, the exception path is where the humans live.

When to use — and when not to

• Use it when the opportunity scan (pb-36) proposed an agent treatmentbefore tool selection, because the redesign scope determines what the tool must do, not the reverse (pb-37's rule).
• Use it when an existing AI deployment underperformsthe usual diagnosis is an agent grafted onto an unredesigned process; the to-be map makes the layering anti-pattern visible.
• Use it to make oversight concreteautonomy tiers per step are how pb-45's oversight requirements land in the actual flow, with volumes and SLAs attached.
• Skip it for single-step assist toolsa drafting copilot inside one task changes a task, not a process; a job aid and a quality gate suffice.
• Skip it if the as-is process was never validatedredesigning from an unwalked map compounds fiction; run the SME walkthrough (pb-23) first.
• Do not design autonomy from trust you don't have yetstarting a new agent at act-and-log because the demo was impressive is how incidents get their root cause; tiers are earned through measured performance, and the design must say how.

The recipe

1. 1.

   Re-segment the process before assigning tiers. The agent-era process is rarely the human process with substitutions. Ask of the validated as-is map: which steps exist only because work moved between humans (batching, queuing, status meetings — candidates for elimination); which sequential steps can run in parallel once an agent holds the whole case; where should one agent step replace three human micro-steps? Draft the re-segmented flow first — this is the anti-layering move — then tier it.

2. 2.

   Assign an autonomy tier to every step, with rationale. Three tiers cover almost everything: suggest (agent drafts, human decides and acts), approve-then-act (agent prepares the action, executes on human approval), act-and-log (agent executes, humans review samples and exceptions). Have AI produce the first pass so you can spend your time correcting it:

   Copy

   For each step in the process below, propose an autonomy tier
   (SUGGEST / APPROVE-THEN-ACT / ACT-AND-LOG) as a table:
   Step | Tier | Rationale (impact, reversibility, volume,
   verification cost) | Entry criteria to move UP a tier
   (measured, e.g. "≥98% approval rate over 4 weeks, zero Sev-1")
   | Exception conditions that route to a human regardless of tier
   Rules: irreversible or customer-visible actions start no higher
   than APPROVE-THEN-ACT. If volume makes human approval
   infeasible, say so — do not silently pick ACT-AND-LOG.
   Process and volumes: [PASTE VALIDATED MAP + DATA]
   

   Every tier you change during review, record why — those corrections are the design rationale the steering committee will ask for.

3. 3.

   Design the exception path as a first-class flow. In an agent process, exceptions are the human workload, so model them with the same care as the happy path: named exception queues per reason code (low confidence, validation failure, customer dispute, policy edge), a triage step with routing rules, SLAs per queue, and — critically — the feedback loop: resolved exceptions become golden-set cases (pb-48) so the agent's weakest categories are exactly where the eval set grows. A to-be map whose exception path is one sad box labeled "manual handling" is not done.

4. 4.

   Wire escalation and the kill switch into the model. Escalation is a gateway with conditions, not a cultural aspiration: define what triggers step-level pause (error-rate threshold from monitoring), what triggers process-level stop (incident classes from pb-52), who has authority to pull each, and what the degraded-mode process is — the pre-agent manual path, kept documented and rehearsed, because "turn it off" is only an option if the humans still know how to run the process. Draw degraded mode as a lane, not a footnote.

5. 5.

   Walk the to-be model with SMEs and the future exception-handlers. Same discipline as any process validation: happy path first, then interrogate every gateway — "the agent's confidence is low AND the customer is on the phone: what really happens here?" Pay special attention to the people who will staff the exception queues; their volume estimates and "that reason code is actually three different problems" corrections are the difference between a blueprint and a diagram. Log corrections as findings; stamp the model validated only after the walkthrough.

Make it yours

• Free modeling path: Mermaid or draw.io via LLM generates the first BPMN draft from your step table; Signavio or Miro AI if licensed — the walkthrough discipline is identical and non-negotiable either way.
• Tier ladder as a rollout plan: launch every step one tier below its target and promote on the entry criteria you wrote — the design doubles as the phased rollout and change-management story (pb-49).
• Orchestration-tool alignment: if the org runs Celonis or a workflow engine, express queues, SLAs, and routing rules in the tool's vocabulary so the blueprint is implementable configuration, not a translation project.
• Capstone pairing: this blueprint plus the requirements pack (pb-43–45, pb-47) is the two-artifact deliverable that defines the agent-era BA portfolio.

Quality check

• The process was re-segmented, not decoratedat least one handoff eliminated or steps parallelized; if the to-be is the as-is with an agent icon, the layering anti-pattern won.
• Every step has a tier, a rationale, and promotion criteriaautonomy is earned by measured performance, and the measurements are named.
• Irreversible actions start conservativenothing customer-visible or hard to undo launches at act-and-log.
• Exception queues have reason codes, SLAs, and staffing mathexpected volumes per queue are estimated against real exception rates, not hoped small.
• Escalation, kill switch, and degraded mode are modeledwith named authority and a manual path that still works.
• SMEs and exception-handlers walked the modelcorrections logged, validation stamped, exactly as you would for any process map.