Skip to content

AI Workflow Design Matrix

Every AI workflow can be described by two dimensions: how much decision-making the AI has and whether a human is in the loop during execution. These two dimensions combine into a 3x2 matrix of six workflow archetypes — a shared vocabulary for classifying, comparing, and designing AI workflows.

A third dimension — Lens — determines the scope of your analysis:

  • Individual — Workflows scoped to one person’s trigger-to-deliverable flow. Focus: personal productivity, task automation, quality improvement.
  • Organizational — Workflows scoped to an end-to-end business process, potentially spanning multiple roles. Focus: value chain optimization, cross-functional efficiency, strategic outcomes.

The lens doesn’t change the matrix — it changes what you’re analyzing. An individual-lens audit surfaces your personal pain points; an organizational-lens audit surfaces value chain processes tied to business objectives. Both produce candidates classified on the same Autonomy x Involvement matrix.

Dimension 1: Autonomy — How Much Decision-Making Does the AI Have?

Section titled “Dimension 1: Autonomy — How Much Decision-Making Does the AI Have?”

Autonomy describes how much latitude the AI has to make decisions during execution. This is about the AI’s role, not the human’s.

LevelAI’s RoleWhat It Looks LikeExample
DeterministicFollows fixed rules — no decisions, no judgmentInput → predefined steps → structured output. Same input produces same output every time.Format a report from a template, extract data from a form
GuidedMakes bounded decisions within guardrailsAI chooses how to accomplish a step, but the human sets direction and reviews output.Draft an email based on meeting notes, research a topic and summarize findings
AutonomousPlans, decides, and adapts independentlyAI determines what to do, uses tools, adjusts its approach based on what it finds.Research agent that finds sources, evaluates quality, and produces a report end-to-end

Key question: How much does the AI decide on its own?

  • Deterministic — nothing; it follows your script exactly
  • Guided — some; it makes choices within boundaries you set
  • Autonomous — a lot; it plans its own approach and adapts

Dimension 2: Human Involvement — Is a Human in the Loop During Execution?

Section titled “Dimension 2: Human Involvement — Is a Human in the Loop During Execution?”

Human involvement describes whether a human participates while the workflow is running — not before (design) or after (review), but during.

ModeHuman’s RoleWhat It Looks LikeExample
AugmentedHuman is in the loop — reviews, steers, or decides at key pointsAI pauses for human input, feedback, or approval before continuing. Human and AI collaborate in real time.Co-writing a document, reviewing AI research before it continues
AutomatedAI runs solo — human reviews only the final outputWorkflow executes end-to-end without human intervention during the run. May be triggered manually or on a schedule.Weekly report generated overnight, prospect list compiled on a schedule

Key question: Does a human participate during the workflow run, or only see the final result?

Combining these two dimensions produces six distinct workflow archetypes:

Augmented (human in the loop)Automated (AI runs solo)
DeterministicHuman triggers a fixed sequence and reviews output before it’s used. AI follows rules; human confirms results.Fixed sequence runs on a schedule or trigger with no human involvement. Same input → same output, every time.
GuidedHuman and AI collaborate — AI researches, drafts, or analyzes; human steers, refines, and decides.AI handles research, drafting, or analysis on its own, applying bounded judgment. Human reviews only the final output.
AutonomousAI plans and executes multi-step work; human reviews at defined checkpoints before the workflow continues.AI runs a full pipeline end-to-end — plans, executes, adapts — with no human intervention until the deliverable is complete.
ArchetypeAutonomyInvolvementDescriptionExample
Deterministic + AugmentedDeterministicAugmentedHuman triggers a rules-based process and confirms results before useRun a data extraction template, review output, then forward
Deterministic + AutomatedDeterministicAutomatedRules-based process runs unattended on a trigger or scheduleNightly report formatting, scheduled form processing
Guided + AugmentedGuidedAugmentedAI drafts, researches, or analyzes; human steers and decides in real timeMeeting prep research, co-writing, competitive analysis
Guided + AutomatedGuidedAutomatedAI applies bounded judgment independently; human reviews the final outputDraft personalized outreach emails overnight for morning review
Autonomous + AugmentedAutonomousAugmentedAI executes a multi-step pipeline with human approval gatesMulti-agent research → write → edit pipeline with a review checkpoint
Autonomous + AutomatedAutonomousAutomatedAI runs a full autonomous pipeline end-to-end without human involvementContinuous monitoring agent that detects, analyzes, and alerts

These three worked examples from the Build section illustrate different matrix positions:

ExampleArchetypeWhy
Deterministic AutomationDeterministic + AutomatedFixed rules, structured input/output, runs the same way every time with no human steering
AI CollaborativeGuided + AugmentedAI researches and drafts; human reviews, refines, and decides what to use
Autonomous AgentAutonomous + AugmentedMulti-agent pipeline executes autonomously with one human review gate before publishing
  1. How much should the AI decide on its own? → Determines your autonomy level (Deterministic / Guided / Autonomous)
  2. Does a human need to be involved during the run? → Determines your involvement mode (Augmented / Automated)

Most workflows evolve along predictable paths as you build confidence:

  • Deterministic + Augmented → Deterministic + Automated — You start by running the process and reviewing output. Once you trust it, you schedule it to run unattended.
  • Guided + Augmented → Guided + Automated — You collaborate with AI on drafts and research. Once the AI consistently produces good output, you let it run independently and review only the final result.
  • Guided + Augmented → Autonomous + Augmented — The workflow grows more complex. You add tool access and multi-step reasoning, but keep a human review gate for high-stakes output.

The orchestration mechanism describes who drives the workflow. The matrix describes how the AI and human interact. They’re complementary:

Orchestration MechanismTypical Archetypes
PromptDeterministic + Augmented, Deterministic + Automated
Skill-Powered PromptDeterministic or Guided, either involvement mode
AgentGuided or Autonomous, either involvement mode

The seven workflow architecture patterns provide implementation blueprints within each archetype:

ArchetypeCommon Architecture Patterns
DeterministicAugmented LLM, Prompt Chaining
GuidedPrompt Chaining, Routing, Evaluator-Optimizer
AutonomousOrchestrator-Workers, Autonomous Agents