A Leader's Map for the AI Portfolio

The AI Cognitive Quadrant

Where rules, ML, GenAI, and agents actually fit — and what kind of human oversight each one really demands.

Q1 · Codify — Rules / RPA Q2 · Compute — ML Models Q3 · Compose — GenAI Q4 · Orchestrate — Agents

Two axes, four kinds of system

Most organizations talk about “AI” as if it were one thing.

It isn't. A pricing rule, a demand model, a copy-drafting assistant, and an agent that coworks on a quarterly business review are fundamentally different systems — different failure modes, different governance needs, different demands on the humans around them. Plot two axes and four quadrants fall out, each mapping almost exactly to a generation of AI.

X-Axis

Complexity of Execution

How many steps, systems, and dependencies must be coordinated for the task to complete? Low is a single contained action. High is multi-step, multi-tool work that unfolds over time.

Y-Axis

Ambiguity of Judgment

How much does the “right answer” depend on interpretation, context, and values? Low ambiguity means the correct output is knowable and checkable. High ambiguity means reasonable people could land in different, still-defensible places.

Each quadrant maps almost exactly to a generation of AI technology — and to a distinct human role.

The four quadrants in depth

Four systems, four human roles

Each quadrant carries its own failure modes and its own demand on the people around it. The technology it maps to is only half the story — the other half is what the human is there to do.

Q1 · Codify

Deterministic logic

Rules engines · RPA bots · validation checks

Low execution · Low ambiguity

The oldest form of “AI” in the enterprise and often the most under-appreciated — not exciting, but cheap, auditable, and fast. One contained action, one correct answer.

◆ Use-case scenario

A pricing rule automatically blocks any price change that would push an item's margin below a set floor, and routes it for review. The logic is fixed and the outcome checkable — the rule either fired correctly or it didn't.

◆ Human's role — Designer of the rule

Judgment is front-loaded into writing the rule correctly; after that, oversight is mostly exception handling. This quadrant needs the least ongoing human judgment of the four.

Q2 · Compute

Classic ML / AI

Forecasting · elasticity · churn · recommenders

High execution · Low ambiguity

Execution is genuinely complex — feature engineering, training, tuning, serving at scale — but the objective function is clear and the output is checkable against ground truth.

◆ Use-case scenario

A demand-forecasting model predicts next week's unit sales per SKU per store to drive replenishment. Building it is hard, but “was the forecast accurate?” has a factual answer once the week closes — so it can be measured, retrained, and trusted on merit.

◆ Human's role — Validator & model owner

Humans define the objective, choose features and constraints, and judge whether accuracy and behavior are fit for purpose. The judgment is technical — “is this good enough and safe enough to trust” — not interpretive.

Q3 · Compose

GenAI in its common form

Drafting · summarizing · creative variants

Low execution · High ambiguity

The execution itself is simple — one inference call — but there's no single correct output. Tone, framing, and appropriateness are all judgment calls.

◆ Use-case scenario

GenAI drafts product descriptions and marketing copy for thousands of catalog items. Any one draft returns in seconds, but whether it's on-brand, accurate, and compelling is a judgment no metric fully settles — so a human has to shape and approve it.

◆ Human's role — Editor & curator

The job shifts from writing the answer to evaluating and shaping it — prompting well, reviewing critically, deciding what “good” looks like in context. Genuinely new work; prompt design and evaluation are becoming real job functions.

Q4 · OrchestrateThe Frontier

AI agents that cowork

Plan · call tools · coordinate multi-step workflows

High execution · High ambiguity

Both axes maxed out: execution spans systems and time, and there's no ground truth to check each judgment call against in the moment.

◆ Use-case scenario

An analyst partners with an agent to build a quarterly business review. The agent pulls numbers from the BI platform, gathers context from documents and prior decks, drafts the narrative, flags anomalies, and iterates over several rounds. Genuine cowork — the analyst steers, challenges, and owns the final narrative.

◆ Human's role — Supervisor & accountability holder

Humans can no longer review every action. The job moves to designing guardrails, defining escalation thresholds, auditing outcomes in aggregate, and remaining accountable. “Human-on-the-loop,” not “human-in-the-loop” — and it needs logging, escalation paths, and kill switches most orgs haven't built.

The pattern across the quadrant

The human arc

The human doesn't disappear as you move up and to the right — the nature of their judgment changes.

1
Q1 · Codify

Designer of the rule

Judgment front-loaded and technical.

2
Q2 · Compute

Validator of the model

Judgment technical, on merit.

3
Q3 · Compose

Editor of the output

Judgment ongoing and interpretive.

4
Q4 · Orchestrate

Supervisor of the system

Judgment at the level of behavior.

In Q1 and Q2, human judgment is front-loaded and technical. In Q3 and Q4, it becomes ongoing and interpretive — eventually exercised at the level of the system's behavior rather than any single output. That's a genuinely different skill set, and usually not the one already sitting in an organization's rules-engine or data-science teams.

The adoption cost of each quadrant

Change management underneath each quadrant

The quadrant is also a map of how much an organization must change to adopt each type well — and the change gets harder as you move up and to the right.

Q1

Displacement & trust in automation

Mostly operational. Redeploy the people whose manual checks the rule now performs — don't just tell them the bot has their job. The trap is quiet resistance (shadow manual processes “just in case”). The fix: transparency and involving process owners in writing the rule.

Q2

From intuition to model-driven decisions

The hard change is cultural, not technical: getting experts to trust a forecast over their gut. Position the model as a decision aid the expert owns — able to see why it said what it said and to override with a reason. Explainability isn't a nice-to-have; it's the adoption strategy.

Q3

The new skill of editing & evaluating

Writing content becomes reviewing and directing it — a different skill. People need training and permission to reject AI output, plus clear quality bars. Watch two failure modes: over-trust (shipping unreviewed output) and over-rejection (quietly not using the tool).

Q4

Supervising a system you don't control step-by-step

The deepest change, because it touches accountability. It works only with trust built incrementally, clear ownership of the agent's output, and visible guardrails. Rush it and you get rejection (“not putting my name on it”) or abdication (“not my problem”) — both dangerous.

The cross-cutting theme

Change management shifts from retraining tasks (Q1–Q2) to rebuilding trust and redefining roles (Q3–Q4). The adoption bottleneck is almost always human — so the change effort should scale with the quadrant, not with the size of the model.

For the C-suite

What AI leadership should take from this

Five moves that follow directly from knowing which quadrant a piece of work belongs in.

1

Match governance to ambiguity, not hype

A high-complexity forecast (Q2) and a low-complexity GenAI draft (Q3) can carry very different risk profiles. Governance should scale with how much interpretation is involved — not with how impressive the technology sounds.

2

Don't govern every quadrant the same way

Q4-level scrutiny on a Q1 rule kills velocity for no safety benefit; Q1-level scrutiny on a Q4 agent is how orgs get surprised. Most governance failures are a mismatch of this kind — not a lack of governance overall.

3

Build the muscle you're missing

Most enterprises are well-staffed for Q1 and Q2. Few have invested in Q3 evaluation design or Q4 agent-guardrail design as disciplines. If agentic AI is on your roadmap, the gap is usually not the model — it's the muscle to supervise it.

4

Change what you measure as you move right

Q1 and Q2 are measured on accuracy and throughput. Q3 and Q4 need judgment-quality measures — escalation rates, override rates, outcome audits — because “did it execute” stops being the interesting question once ambiguity is high.

5

Expect a portfolio, not a single answer

Most real functions need all four quadrants working together: rules for guardrails, ML for forecasts, GenAI for content, and agents to orchestrate across the whole. The leadership job is knowing which quadrant each piece belongs in — and resourcing the human role it demands.

From map to build

Now the concrete part: how each quadrant actually works

The framework is the easy half. The hard half is what it takes to build in each quadrant — and what specific engineering you add to move up (more ambiguity) or right (more execution, toward true co-work). Below: an engineering spec per quadrant, then the two moves broken down to the component level, then a real agent decomposed end-to-end.

The engineering underneath

A spec sheet for each quadrant

Same four quadrants — but now the stack, the data, how you prove it's “good,” the dominant failure mode, and the infrastructure each one demands. The pattern to notice: as you move up-and-right, verification is replaced by evaluation, and a single call is replaced by a governed loop.

Q1 · Codify

Deterministic logic

Rules engines · RPA · validation
Stack
Decision tables & if/then rules (Drools-style), schema validators, RPA scripts. No model.
Data / input
Structured, typed, deterministic conditions. Nothing learned.
Proof of “good”
Verification. Unit tests & assertions — the rule fired or it didn't. 100% reproducible, binary correct.
Failure mode
Brittleness & rule sprawl; silent drift when reality changes but the rule doesn't.
Infra
Rule version control, audit log, an exception queue for human review.
Q2 · Compute

Classic ML / AI

Forecasting · elasticity · churn
Stack
Supervised / statistical models (gradient boosting, regression, deep nets), feature pipelines, MLOps serving.
Data / input
Labeled history + engineered features. Quality of data is the product.
Proof of “good”
Verification vs ground truth. MAPE, AUC, precision/recall; backtests, holdout sets.
Failure mode
Distribution shift & drift, data leakage, bias, quiet accuracy decay over time.
Infra
Feature store, model registry, retraining pipelines, drift & performance monitors.
Q3 · Compose

GenAI generation

Drafting · summarizing · variants
Stack
Foundation model + prompt template; often RAG for grounding; structured / constrained output.
Data / input
Prompt + context + retrieved documents. Behavior tuned by prompt & examples, not retraining.
Proof of “good”
Evaluation, not verification. Rubrics, LLM-as-judge, human preference labels, golden sets, red-teaming.
Failure mode
Hallucination, off-brand tone, inconsistency, prompt injection.
Infra
Prompt versioning, an eval harness, output guardrails (content / PII filters), a human review queue.
Q4 · Orchestrate

Agentic systems

Plan · tools · multi-step cowork
Stack
LLM + tool/function calling + a planning loop (ReAct / plan-execute) + memory + orchestration (state machine / graph).
Data / input
A goal, a tool/API set, retrieved context, and running state carried across steps.
Proof of “good”
Trajectory + outcome evals. Not just the final answer — were the actions safe? Escalation, override, and aggregate outcome audits.
Failure mode
Compounding errors across steps, tool misuse, runaway loops, unsafe actions, cost blow-ups.
Infra
Orchestration + tracing/observability, guardrails, HITL checkpoints, escalation paths, sandboxed tools, budget caps, kill switch.

How you actually cross the map

Two moves — and they're different engineering problems

You don't drift diagonally. You make two distinct investments: moving up means engineering for ambiguity; moving right means engineering for execution and, ultimately, co-work. Most teams underestimate the first and rush the second.

Move up ↑

Engineering for ambiguity

Q1→Q3 and Q2→Q4. The right answer stops being checkable, so you can no longer assert correctness — you have to evaluate quality.

The core shift: from verification (a test that passes) to evaluation (a judgment of quality across a distribution).

Move right →

Engineering for co-work

Q1→Q2 and Q3→Q4. One contained action becomes a multi-step process that spans systems and time, and takes actions on your behalf.

The core shift: from a single call to a governed loop of planning, tool use, memory, and human checkpoints.

The upward move, concretely

What you build to survive more ambiguity

In Q1/Q2 you write a test and it passes. In Q3/Q4 there is no single correct output to test against — so you replace the test suite with an evaluation system and constrain the space so the model can't wander. Six concrete additions:

1

Grounding

Constrain the answer space with retrieval: RAG, citations, structured context. Less room to invent = less ambiguity to manage.

2

An eval harness

Golden datasets + rubric scoring + LLM-as-judge (calibrated against human labels). You measure a distribution of quality, not pass/fail.

3

Human preference signal

Capture edits, approvals, and rejections as labeled data. Pairwise comparison and preference data become your ground-truth proxy.

4

Guardrails

Input/output filters, policy & PII checks, prompt-injection defense — because you can't pre-verify every generation.

5

Calibration & abstention

Let the system say “I'm not sure” and escalate. Confidence thresholds turn ambiguity into a routing decision, not a silent guess.

6

Feedback loop

Route human corrections back into prompts, few-shot examples, or fine-tuning. The system improves on the exact judgments it gets wrong.

The rightward move, concretely

From Compose to co-work: the capability ladder

Getting from a GenAI draft (Q3) to an agent that coworks on a QBR (Q4) is not one leap — it's a stack of capabilities, each adding real power and a new class of risk you have to control. This is the concrete path from “one inference” to “governed loop.”

Start · Q3 ComposeEnd · Q4 Orchestrate
1

Single inference Q3 baseline

One prompt in, one output out. Stateless. The human triggers it and edits the result — classic Compose.

Adds: nothing new. Ceiling: can't act, can't use fresh data, can't take more than one step.

2

+ Retrieval (RAG) grounding

Pull real, current context from your systems (vector search, BI, docs) into the prompt so the output is grounded in truth, not the model's memory.

Adds: accuracy & freshness. New risk: retrieval quality — wrong context in, confident wrong answer out.

3

+ Tool / function calling it can now act

The model calls typed functions with validated arguments — run a SQL query, fetch a report, generate a chart. This is the line where it stops talking and starts doing.

Adds: real actions across systems. New risk: tool misuse & side effects — every tool needs scoped, sandboxed permissions.

4

+ Planning & reflection loop multi-step

Decompose a goal into steps and iterate — ReAct or plan-and-execute: reason → act → observe → re-plan. The agent now chains many calls toward an objective.

Adds: autonomy over sequences. New risk: compounding errors & runaway loops — needs step limits and checks.

5

+ Memory & state continuity

Working memory across steps and durable memory across sessions, so the agent carries context, prior decisions, and the analyst's earlier steering into later rounds.

Adds: coherent multi-round cowork. New risk: stale or poisoned memory quietly corrupting later decisions.

6

+ Orchestration control flow

A supervising layer (state graph / controller, e.g. LangGraph-style) coordinates tools and sub-agents with branching, retries, and error handling — turning a loose loop into a reliable workflow.

Adds: reliability at scale. New risk: emergent behavior across components that no single step reveals.

7

+ Guardrails, HITL & observability Q4 co-work

The controls that make supervision real: human-on-the-loop checkpoints, escalation thresholds, action-scoped guardrails, budget/rate caps, full trace/observability, and a kill switch. This is the rung that makes co-work safe enough to trust.

Adds: accountable autonomy — the analyst steers and owns; the agent does the legwork. Without it: you have autonomy no one can supervise.

A real Q4 system, decomposed

Worked example: the QBR co-work agent

The article's quarterly-business-review agent, drawn as an actual architecture. The analyst sets the goal and owns the narrative; the agent runs the loop below — retrieving, acting, drafting, and pausing at human checkpoints — inside a cage of guardrails.

🧠
1 · Plan

Decompose the QBR goal into steps & data needs.

🔧
2 · Retrieve & act

Query BI, pull prior decks & docs via scoped tools.

3 · Draft & flag

Write the performance narrative; surface anomalies.

👤
4 · Human checkpoint

Analyst reviews, challenges, steers the story.

5 · Iterate & own

Refine over rounds; analyst owns the final QBR.

↻ Steps 2–4 loop across several rounds — the agent re-plans as the analyst reacts. No human scripts each step; no single correct QBR exists.

🛡 Guardrails

Data-access scoped to the analyst's permissions; no external send; policy & PII filters on every draft.

⏱ Escalation

Low-confidence numbers or unexplained anomalies are flagged up, not silently included.

🔍 Observability

Every tool call, prompt, and decision is traced — so outcomes can be audited in aggregate.

🛑 Kill switch

The analyst can stop or override the run at any checkpoint and remains the accountable party.

Underneath the ambiguity axis

Not all uncertainty is the same

The Y-axis quietly blends two fundamentally different kinds of uncertainty. Telling them apart is the difference between a problem you can solve and one you can only manage — and it's the single most useful idea for anyone building non-deterministic systems.

Epistemic

Reducible uncertainty

You can shrink it — so reduce it

You're unsure because you lack knowledge: not enough data, features, or model capacity. There is a right answer; you just can't see it clearly yet. More and better data collapses it toward that answer. This is the uncertainty of Q1→Q2 — a fuzzy forecast sharpens with more history.

How you attack it: collect more data, better features, retraining, ensembles, active learning, RAG grounding, confidence intervals. Effort converges on truth.
Aleatoric

Irreducible uncertainty

You can't shrink it — so manage it

The task itself is genuinely non-deterministic: inherent noise, or many equally-defensible “right” answers. No amount of data collapses it, because there is nothing to collapse to. This is the uncertainty of Q3→Q4 — “the best on-brand tone,” “the right QBR narrative.”

How you manage it: human ownership, present a range or options not a point answer, calibrated abstention, guardrails on the acceptable band, preference learning to encode taste (shrinks perceived ambiguity, never to zero).
⚠ The cardinal mistake

Teams treat aleatoric ambiguity as if it were epistemic — pouring more data, bigger models, and stricter evals at Q3/Q4 expecting the output to converge on one “correct” answer. The residual is irreducible, so it never converges. That is precisely why the top of the quadrant needs a human owner and a defined “good-enough range,” not a hunt for a ground truth that doesn't exist.

The “when” of human judgment

You can shape an agent at three different times

Human curation isn't a single moment. There are three distinct windows to inject judgment — and most teams only use one. The higher the quadrant, the more you need all three working as a flywheel, because you can't front-load every decision into code the way Q1 does.

Clock 1 · before it runs

Design-time

Humans as architects of the envelope

Bake judgment into the system up front: system prompts & policies, the tool / action allow-list, guardrail rules, eval suites, few-shot exemplars, fine-tuning & RLHF, the “constitution.”

Q1 lives entirely here — write the rule once. In Q4 it sets the boundaries the agent may never cross.
Clock 2 · while it runs

Runtime

Humans as pilots

Steer in the moment: approval checkpoints (human-on-the-loop), live interventions and overrides, answering the agent's escalations, correcting mid-task before errors compound.

The heart of Q4 co-work — barely exists in Q1. This is where the analyst steers the QBR agent round by round.
Clock 3 · after it runs

Learning-time

Humans as editors of its future

Curate outcomes to improve the next run: audit trajectories, label good/bad, update evals & prompts, capture overrides as preference data, widen or tighten the action scope on track record.

Closes the loop back to design-time — learning-time evidence is what earns an agent a wider envelope.

Design-time → Runtime → Learning-time  ↻  a flywheel: each run's outcomes reshape the next run's envelope.

How the mix shifts across the quadrant

In Q1, judgment is almost entirely design-time — you write the rule and walk away. As you climb toward Q4, judgment spreads across all three clocks: set the envelope, steer inside it, and reshape it from outcomes. Supervising an agent means owning all three, not just the code review.

Q1 · Codify
Design 88%
Q2 · Compute
Design 52%Learning 36%
Q3 · Compose
Design 42%Runtime 40%18%
Q4 · Orchestrate
Design 30%Runtime 42%Learning 28%
Design-timeRuntimeLearning-timeIllustrative proportions

Four lenses the 2×2 doesn't show

Advanced operating concepts for the top-right

The quadrant tells you what kind of system you have. These four sharpen how you actually govern the hard corner — and each one cuts across all four quadrants rather than sitting in one.

🎟

Graduated (earned) autonomy

Autonomy isn't binary — it's a dial the agent earns. Move it one notch at a time: assistdraft-with-approvalact-within-scopeact-and-report. Calibrated confidence and track record (from learning-time) unlock a wider action scope.

Trust is granted incrementally, never by default.
🚪

Reversibility & blast radius

The hidden third axis. Two-way doors (reversible, cheap to undo) can run with light gating even in Q4. One-way doors (money moved, message sent, data deleted) need a hard human gate regardless of quadrant. Govern by consequence, not just ambiguity.

Ask not only “how ambiguous?” but “how reversible?”
📊

The oversight budget

Human attention is finite and does not scale with agent throughput. You can't review every action, so spend oversight where marginal value is highest: low-confidence, high-impact, irreversible, or novel cases. Sampling plus escalation — not blanket review.

Attention economics: supervise the exceptions, audit the rest.

Trust calibration

The goal isn't maximum trust — it's calibrated trust: reliance that tracks real reliability. Over-trust ships unreviewed errors; under-trust means shadow manual work and wasted capability. Confidence signals, transparency, and audit history keep the two aligned.

Match the trust to the track record — both directions are failures.

The next time an initiative gets pitched to you as “AI,” ask one thing first: which quadrant is this really in?

The answer tells you what kind of human oversight it needs — and whether you currently have it.