The thesis
AI agents absorb the execution layer of knowledge work — assembly, reconciliation, drafting, analysis. What they do not absorb is judgment: knowing what should exist, what is good enough, what to kill, what a market signal means, and what must never break. The empirical finding (Hitzig et al. 2026) is that domain judgment, not execution skill, is the multiplier on agent labor — and the redesign evidence (BCG 2026; Bain) says organizations capture that multiplier by redesigning processes, not by buying tools.
The unit of analysis is therefore the process / value stream, not the department. One canvas per process, with every functional seat in the room.
Five judgment types
Every process consumes some mix of these. Each one consumed must have a named human carrier — a person, never a committee, never an agent.
What should exist
The unscored new bet, the thing no scenario model proposes. Strongest pre-product-market-fit.
Good enough / architecture
Fund which bet at what level; what quality bar clears; which structure holds.
Kill / simplify
What to unship, stop, or collapse. Grows in importance as products mature.
What the signal means
Reading demand, attribution, and segments — deciding what the numbers are telling you.
What must never break
Cash floors, compliance, committed spend, integrity constraints. Peaks at maturity and sunset.
The canvas — nine blocks, one process
Filled in a half-day workshop with all functional seats present. Three colors, three layers: human judgment, agent execution, audit / control.
| # | Block | Key question | Output |
|---|---|---|---|
| 1 | Trigger | What starts the process today — and what should, once a continuous agent substrate exists? | Trigger definition |
| 2 | Accountability map | Which function certifies what? Accountability never moves to agents. | Certification register |
| 3 | Execution layer | Which activities are agent-delegable now / in 12 months? | Delegation backlog |
| 4 | Judgment layer | Which of the five judgment types does this process consume, and which seat carries each? | Judgment coverage matrix |
| 5 | Encoded judgment | Which judgments become agent rules? Every rule needs an owner and a review trigger. | Rules-of-engagement registry |
| 6 | Option architecture | What option sets should agents pre-generate? Where are the human choice points? | Session design |
| 7 | Control plane | Six checks: reconciliation · forecast drift · assumption freshness · override log · option-set hygiene · semantic consistency. Who audits? | Control panel + audit owner |
| 8 | Staffing shape | Consolidators to redeploy; stewards to create; judgment seats to protect. | Role transition plan |
| 9 | Learning loop | Where do juniors form judgment once execution is absorbed? | Apprenticeship design |
JDEA — the actor grid that replaces RACI
RACI assumed every actor was human. Inside a redesigned process, every task gets four seats:
Judges
Human. Owns the decision and its consequences. Exactly one J per task — a task with no J is an orphaned decision (bug B3).
Directs
Human steering agents: specifies, decomposes, recovers stuck runs.
Executes
Agent, human, or human-directing-agent. Where J and E are the same human, that task is a delegation candidate.
Audits
Control-plane owner. Must be independent of J and D — segregation of duties survives the agentic shift.
The coverage map — lifecycle × judgment intensity
Each lifecycle stage requires a different judgment mix. A required cell with no named carrier is a coverage gap — a hiring or assignment conclusion, not an observation. Headcount follows judgment gaps, not workload.
| JUDGMENT TYPE | PRE-PMF | GROWTH | MATURE | SUNSET |
|---|---|---|---|---|
| TastePrototyper | ●●● | ●● | ● | |
| Trade-offBuilder | ●●● | ●●● | ●● | ● |
| SubtractionSweeper | ●● | ●●● | ●●● | ●●● |
| MarketGrower | ●● | ●●● | ●● | ● |
| RiskMaintainer | ● | ●● | ●●● | ●●● |
Nine bugs, six twists
The framework distinguishes bugs — failure modes to detect and fix, each wired to a control-plane signal — from twists — expected role mutations to track and support. The classic failure is treating twists as bugs (resisting role change) while missing the real bugs.
| Bug | Detection signal | |
|---|---|---|
| B1 | Fossilization — encoded rules outlive their validity | Stale last-reviewed dates |
| B2 | Automation bias — humans certify agent output unread | Override log ≈ zero |
| B3 | Orphaned judgment — a consumed judgment has no human owner | JDEA task with empty J |
| B4 | Apprenticeship gap — juniors never form judgment | Declining exception-detection |
| B5 | Accountability drift — "the agent did it" | Certifications lapse |
| B6 | Option monoculture — all scenarios are variants of one plan | No unscored wildcard in sessions |
| B7 | Consolidator vacuum — analysts cut before stewards staffed | Rule registry unmaintained |
| B8 | Trust asymmetry — functions reject each other's pre-validations | Duplicate analyses reappear |
| B9 | Constraint conflict — two functions' rules contradict silently | Agents deadlock or oscillate |
| Twist — role mutation to expect | |
|---|---|
| T1 | Controller → harness auditor — certifies the machinery that produces the plan |
| T2 | Analyst → agent steward — maintains the function's agent model, rules, data hygiene |
| T3 | Middle manager → player-coach — models agent direction, coaches recovery |
| T4 | Specialist cluster → integrated lead — one full-stack judgment seat |
| T5 | Seniority inversion — judgment fit + agent fluency outvalue execution mastery |
| T6 | New seat: rules-of-engagement referee — owns cross-functional agent interfaces |
The implementation sequence
Order matters: control before delegation, stewards before redeployment.
- Pick one process with a sharp before/after — quarterly budgeting is ideal.
- Run the canvas with all functional seats present; produce the JDEA grid.
- Pay down workflow debt first — stabilize and document before delegating.
- Stand up the control plane and name the auditor before switching execution to agents.
- Redesign sessions around option-choice points; protect the wildcard slot.
- Re-staff by judgment coverage — stewards before redeploying consolidators.
- Instrument from day one — decision log + override log are the early-warning system.
- Set the review cadence — constraints quarterly, bug scan monthly, twists each planning cycle.
Metrics that matter (enterprise level, not micro-productivity): decision-cycle time · forecast accuracy drift · override-pattern health · judgment-coverage gaps closed · share of leader time at choice points vs. assembly review.