Field Guide · Reference Index
Enterprise taxonomies.
Ten lookup tables, designed to be quotable as shorthand — “that’s a V3 task at A2 autonomy.” They compose: an action’s risk class plus its verification class determines its allowable autonomy tier and required oversight mode.
T1 — AI OPERATING MATURITY (M0–M4)
| Level | Name | Description | Telltale |
|---|---|---|---|
| M0 | Exploratory | Individuals use AI tools ad hoc; no institutional footprint | Shadow ChatGPT usage |
| M1 | Assisted | Sanctioned copilots; humans do the work, AI accelerates it | Licenses purchased, workflows unchanged |
| M2 | Orchestrated | Deterministic workflows with embedded model steps | First real engineering; evals appear |
| M3 | Delegated | Agents own bounded tasks within tiered autonomy | Action inventories, provenance, oversight design exist |
| M4 | Engineered | AI portfolio managed on unit economics, eval coverage, and absorption rate as standing functions | AI indistinguishable from "how operations works" |
The load-bearing distinction is M1→M2 (tool adoption becomes engineering) and M2→M3 (engineering acquires governance). Most enterprises claiming M3 are at M1 with demos.
T2 — AGENT AUTONOMY TIERS (A0–A5) — assigned per action class, never per system
| Tier | Name | Behavior |
|---|---|---|
| A0 | Advisory | Produces drafts and recommendations only; humans execute |
| A1 | Gated | Executes, but every action requires pre-approval |
| A2 | Exception-gated | Executes freely; policy triggers route defined cases to humans |
| A3 | Sampled | Executes; humans review a statistical sample post-hoc |
| A4 | Monitored | Executes; automated monitoring with human escalation on anomaly |
| A5 | Delegated | Outcome-level review only |
A single agent legitimately operates at different tiers simultaneously — A4 for reading and drafting, A1 for payments. A single system-wide tier does not show whether an action inventory was performed — ask for it.
T3 — HUMAN OVERSIGHT MODES
| Mode | Human position | Appropriate when |
|---|---|---|
| In-loop | Approves each action before execution | Risk class IV; early trust gradient |
| On-loop | Watches live, can intervene | Transitions; high-tempo workflows |
| Sampled | Reviews statistical sample after execution | Risk class II with evidence base |
| Exception-only | Sees only policy-triggered cases | Mature A2/A4 operation |
| Outcome-only | Reviews results, not actions | Risk class I; verification is mechanical |
The design failure to name in workshops: nominal in-loop — a human formally in the loop who, once approval volume outruns the attention available, is functionally outcome-only while carrying in-loop accountability.
T4 — ENTERPRISE KNOWLEDGE CLASSES
| Class | Description | Agent-readiness |
|---|---|---|
| K1 Structured | Systems of record; transactional truth | High — the trust anchor; everything derives from here |
| K2 Recorded | Documents, wikis, policies | Medium — abundant, stale, contradictory; needs source-of-truth tiers |
| K3 Procedural | How work is actually done | Low until made legible; the target of legibility programs |
| K4 Tribal | Undocumented know-how in tenured heads | Zero until extracted; the largest hidden dependency |
| K5 Judgment | Tacit criteria for good decisions | Partially extractable as resolved ambiguities and graded examples; the rest is what humans are for |
The strategic insight: RAG programs concentrate on K2, the least differentiating class. Competitive advantage lives in converting K3/K4 to legible form and encoding K5 into evaluation criteria.
T5 — AI GOVERNANCE CONTROL TYPES
| Type | Nature | Example |
|---|---|---|
| Structural | What the system can do | Scoped credentials, egress controls, tool allowlists |
| Runtime-preventive | Deterministic gates on actions | Threshold checks, policy-as-code, approval routing |
| Detective | Discovers deviation | Evals, canaries, distribution monitoring, reconciliation |
| Corrective | Undoes or contains damage | Rollback, compensating transactions, kill switches |
| Procedural | Human process | Review boards, attestations, training |
| Documentary | Enables reconstruction | Provenance, traces, decision records |
Governance maturity is the migration of weight leftward: immature programs are procedural and documentary; mature ones are structural and runtime, with procedure reserved for judgment.
T6 — EVALUATION MATURITY (E0–E4)
| Level | Name | Description |
|---|---|---|
| E0 | Anecdotal | Demos and spot checks; confidence by vibes |
| E1 | Static | Fixed test set run at release time |
| E2 | Representative | Production-sampled inputs, graded scoring, periodic refresh |
| E3 | Continuous | Runs on every change — including vendor model updates; drift alarms; correlated-grader controls |
| E4 | Institutional | Eval assets versioned and owned; grader–human agreement tracked; production disagreements feed the suite; eval results gate change management |
E3 is the minimum for anything in risk classes III–IV. The E2→E3 jump is organizational, not technical: it requires someone to own evaluation as a standing function.
T7 — WORKFLOW COMPLEXITY CLASSES (W1–W5)
| Class | Shape | Automation posture |
|---|---|---|
| W1 | Linear, deterministic | Classic automation; models add nothing structural |
| W2 | Branching, rules-expressible | Orchestration with model steps for perception/extraction |
| W3 | Exception-heavy | Value concentrates in exception mass; census before building |
| W4 | Judgment-dense | Agent prepares, human decides; encode judgment as eval criteria over time |
| W5 | Open-ended | Genuine agentic planning; smallest category, highest risk, most hype |
The most common scoping error is treating W3 as W2 (the exceptions surface in production) and W4 as W5 (deploying open-ended agents where judgment placement was the actual problem).
T8 — VERIFICATION COST CLASSES (V0–V4)
| Class | Verification | Examples | Implication |
|---|---|---|---|
| V0 | Mechanical, free | Compiles, schema-validates, reconciles | Long chains safe; automate aggressively |
| V1 | Rule-checkable | Policy lint, balance checks, format contracts | Cheap gates; near-V0 economics |
| V2 | Spot-checkable | Statistical sampling suffices | A3 autonomy viable; verification budget models the sample |
| V3 | Full expert review | Legal positions, strategic analysis | Verification ≈ generation cost; agent drafts, human owns |
| V4 | Outcome-delayed | Advice quality, credit decisions | Correctness observable only later; requires counterfactual baselines and cohort monitoring |
This is the taxonomy that does the most portfolio work: V-class × risk class predicts deployability better than any capability benchmark.
T9 — ACTION RISK CLASSES (I–IV)
| Class | Reversibility | Blast radius | Minimum posture |
|---|---|---|---|
| I | Reversible | Contained | A3–A5; outcome/sampled oversight |
| II | Reversible | Broad | A2–A3; pre-execution checks, monitored rollout |
| III | Irreversible | Contained | A1–A2; runtime gates, full provenance |
| IV | Irreversible | Broad | A0–A1; human approval always, dual control, documentary completeness |
The bands above are risk-only: they assume verification is not the binding constraint. Where it is, the composed rule governs — the Workflow Diagnostic assigns autonomy on risk class and verification class together, and tightens below these bands as verification gets more expensive. A Class I action that requires full expert review to check (V3) does not get A3–A5; it gets A2. The published matrix is on How this tool decided. Where this table and that matrix disagree, the matrix is what actually runs.
T10 — OPERATIONAL INTELLIGENCE LEVELS (OI1–OI5)
| Level | Capability | Question answered |
|---|---|---|
| OI1 | Descriptive | What happened? |
| OI2 | Diagnostic | Why did it happen? |
| OI3 | Predictive | What will happen? |
| OI4 | Prescriptive | What should we do? |
| OI5 | Closed-loop | The system acts — governed by autonomy tiers and risk classes |
OI5 is where operational intelligence merges with agent governance: the moment analytics acquires hands, every taxonomy above applies to it.