Cognitecture manifesto

AI Leverage Matrix

A framework for deciding where AI runs the show — and where you still need to.

Two questions determine how much leverage AI gives you on any task: Can the context be captured structurally? And is there a feedback loop to measure success?

When both answers are yes, AI can run end to end. When neither is, you're the one doing the work. Most interesting tasks live in between.

Think in tasks, not jobs

AI doesn't replace jobs. It replaces tasks. Every role is a shifting mix — some tasks need your judgment, some need your taste, some are just you copy-pasting between systems.

Which tasks AI can take over depends on the answers to two questions. Those same two questions determine what your role becomes for each task.

The two questions

"Context is capturable"

Context is everything the worker needs to do the task correctly: goals, constraints, domain knowledge, inputs from systems, tacit rules, edge cases, and prior decisions. Context is capturable when you can put it somewhere AI can reliably access and interpret without the human restating it each time.

That means it's accessible (lives in systems AI can reach), structured enough (fields, rules, examples, KB articles), stable enough (doesn't change minute-to-minute without being updated), and complete enough that missing context is rare and detectable.

"Feedback loop"

A feedback loop exists when the system can get a signal that meaningfully indicates success or failure — and can use it to improve. That signal can be automated tests, objective metrics, human review with a rubric, user behavior, or ground truth comparisons.

A loop is strong when it's fast (minutes, not months), frequent (enough volume to learn), aligned (measures what you actually care about), and actionable (you can change the process based on it).

What the matrix tells you

This is not "where AI is possible." It's where AI can be trusted with autonomy and improve itself over time.

The matrix isn't static. Your job as a cognitect is to move tasks toward the top-right — by making context capturable and feedback loops measurable. Every task you move is leverage you keep.

Quadrant playbooks

Each quadrant has a different operating model. Not just what AI can do, but what the human's job becomes.

Human Leads. AI Assists.

Context: No / Feedback: No

The work relies on tacit knowledge, lived experience, or nuance. "Good" is hard to measure and disagreement is common. This is where AI hallucinates most confidently.

AI is useful for:

  • Generating options (not answers)
  • Summarizing discussions and producing first drafts you'll heavily edit
  • Exploring implications ("what could go wrong?")

Your job:

  • Set direction and intent
  • Apply judgment and accountability
  • Handle sensitive nuance and stakeholder dynamics

Watch out for:

Over-trusting confident text. Substituting AI output for leadership judgment. Using AI to "decide" when values and tradeoffs are the real work.

You're The Context Layer.

Context: No / Feedback: Yes

One of the most important quadrants — where people feel AI is close to autonomy, but it isn't. The feedback loop works, but context is the bottleneck. AI can iterate, but only if you continuously feed it situational inputs.

The workflow:

  • Human provides a "Context Pack" for each run (goal, current state, constraints, examples, what changed)
  • AI produces an output
  • Feedback signal is observed
  • Human updates context, AI tries again

Watch out for:

Context drip-feeding ("oh, also...") without structure. Stale context. AI optimizing the metric but missing the real goal.

Upgrade path:

This quadrant is screaming for instrumentation. Capture context automatically from systems. Use standard intake forms. Build structured memory (project KBs, decision logs). Turn repeated context into schemas, required fields, and retrieval sources.

You Judge. AI Executes.

Context: Yes / Feedback: No

The creative and strategic quadrant. AI can access the context (brand voice, product info, prior examples), but success is subjective. You can't safely let AI self-correct without human evaluation.

The workflow:

  • Provide AI with context (retrieval + rules + examples)
  • Ask for multiple candidates (3-10)
  • Evaluate with a human rubric
  • Pick one, request a revision pass
  • Log your judgment (why it was good/bad)

Watch out for:

"Looks good" approval without a rubric leads to inconsistent outcomes. Feedback never captured means no learning. Hallucinated facts hiding inside polished writing.

Upgrade path:

Collect judgment labels at scale (approve/reject + reasons). Create QA checklists. Define proxy metrics — engagement, revisions needed, time saved. Turn subjective judgment into structured rubrics.

AI Runs It. End To End.

Context: Yes / Feedback: Yes

True leverage. Context is accessible and structured, outcomes are measurable, and iterative improvement doesn't require constant human involvement. AI operates like a system.

Your job shifts to governance:

  • Define success metrics and constraints
  • Design guardrails and escalation paths
  • Monitor performance and drift
  • Handle exceptions
  • Continuously improve context sources and evals

The production loop:

  • Ingest context (RAG / tools / structured data)
  • Generate with constraints
  • Validate (automated checks)
  • Execute (apply, send, publish)
  • Measure outcomes
  • Learn (update prompts, rules, retrieval, tests)

Watch out for:

Feedback measuring the wrong thing — AI "wins the metric" but loses reality. Silent drift in context sources causing regression. Missing stop conditions leading to runaway loops.

Push tasks right and up

The matrix isn't a classification you accept. It's a map of where to invest. Every task you move toward the top-right is leverage you permanently unlock.

Move right: make context capturable

  • Build intake forms — force explicit context
  • Standardize templates and schemas
  • Centralize docs into a single source of truth
  • Add examples (good and bad) and documented edge cases
  • Build retrieval with citations so AI can pull context itself
  • Assign ownership and an update cadence to each context source

Move up: create a feedback loop

  • Define success metrics, even proxies
  • Add automated validators (format checks, constraints, tests)
  • Capture human approvals and rejections as labeled data
  • Run A/B tests where possible
  • Build a gold dataset of representative cases
  • Set up postmortems that turn failures into test cases

Start with one task

Pick a piece of work that's frequent, repetitive, or expensive in time. Define the input, the output, and what failure looks like. Ask the two questions. Place it in the matrix. Then either work the quadrant playbook or invest in moving it toward the top-right.

This isn't about replacing people. It's about knowing exactly where AI gives you leverage — and where you're still the one who matters.