Skip to content
Primores AI wiki

The Five Levels of AI Enablement

The Five Levels of AI Enablement

TL;DR: AI adoption progresses through five levels — from basic prompting (75% of users) to anticipatory AI that acts before you ask. Most organizations plateau at Level 1; the real value unlocks at Level 3+. The jump isn’t technical — it’s psychological.

The Framework

LevelNameHuman EffortWho’s HereKey Insight
1Done By You100%~75-80%“I haven’t seen ROI” = stuck here
2Done With You50%~10%Standardization across teams
3Done For YouMinimal~5%Agentic execution from specs
4Done Without YouNone~1%Fully autonomous operation
5Done Anticipating YouNone<1%Predictive, proactive AI

Level 1: Done By You

What it is: Direct interaction with AI tools. You prompt, you get output, you use it.

Characteristics:

  • Lowest cost ($20/month access)
  • User performs all work
  • Copy-paste workflows
  • No standardization across team

Examples:

  • Prompting ChatGPT for blog posts
  • Manual interactions with Claude
  • One-off content generation

The Problem: When people say “I haven’t seen the ROI of AI,” they’re stuck here. Each interaction starts from zero.

How to progress: Learn prompting fundamentals, understand model weaknesses, provide quality data to minimize hallucinations.

Level 2: Done With You

What it is: Pre-configured tools (Custom GPTs, Claude Projects) with standardized processes embedded.

Characteristics:

  • 50/50 labor split between human and tool
  • Standardization across teams
  • Still requires active user involvement
  • Prevents “reinventing the wheel”

Examples:

  • Sales playbook GPTs distributed across teams
  • Recipe-maker Gems with built-in rules
  • Custom Projects with pre-baked logic

Value: When anyone on the team uses the same GPT, they get consistent quality without being prompt experts.

How to progress: Build and distribute custom assistants; establish team-wide consistency.

Level 3: Done For You

What it is: Agentic AI systems executing complex tasks from specifications without continuous prompting.

Characteristics:

  • Minimal human intervention
  • Works from comprehensive project plans
  • Self-contained execution
  • Where you need to be in 2026

Examples:

  • Claude Code writing documentation from source files
  • AI systems updating website copy autonomously
  • Content generation following the 5P Framework (Purpose, People, Process, Platform, Performance)

Practical Application: Provide a 300-page sales playbook + landing page → system reorganizes it following best practices without further prompting.

The Transition: The jump to Level 3 is not technical — it’s psychological. It’s about trusting AI with execution.

How to progress: Create comprehensive specifications with success criteria. Shift from “copy-paste monkey” to specification focus.

Level 4: Done Without You

What it is: Fully autonomous systems operating without human-in-the-loop oversight.

Characteristics:

  • System sets its own agents and processes
  • Only high-level objectives required
  • Requires careful permission modeling
  • Poses existential threat to agencies and contractors

Examples:

  • Systems autonomously optimizing websites
  • AI writing legal documents (NDAs, contracts)
  • Autonomous content marketing operations

The Math: “A six-to-nine-million-dollar project you can do in six to nine hours for six to nine dollars.”

Critical Design Challenge: Engineering what systems should NOT do autonomously. Building trust and permission frameworks.

How to progress: Focus on permission modeling — what should AI be allowed to do without asking?

Level 5: Done Anticipating You

What it is: Persistent-memory AI systems that identify needs before you express them.

Status: Component technologies exist; full implementation expected by end of 2026.

Characteristics:

  • Always-on persistent memory
  • Predictive capability
  • Makes proactive recommendations
  • No explicit request required

Examples:

  • System notices you frequently create sales playbooks → proactively builds new ones for emerging products
  • Monitors market news → recommends portfolio rebalancing without prompting
  • Identifies optimization opportunities from your data patterns

Enabling Technologies:

  • Persistent memory systems (ByteDance Open Viking, Serena MCP)
  • Pattern recognition across interaction history
  • Proactive notification systems

Critical Insights

Leapfrogging Is Possible

Organizations stuck at Level 1 can skip directly to Level 3 or 4 as tools like Claude’s agentic features mature. Level 2 may become optional.

The Trust Problem — Three Psychological Frictions

Most organizations plateau at Levels 1-2 not because of technical limitations but because they can’t psychologically hand over control. Research from Wharton + Science Says (surveying 700,000+ employees across Google, Zapier, ServiceNow, and others) identifies three specific frictions:

FrictionQuestion User AsksWhat Blocks Adoption
Perceived Competence”Can this agent actually do this?”Users won’t delegate to agents they perceive as incompetent
Trust”Should I trust it with this specific task?”Vague agents that hide limitations erode confidence
Delegation of Control”How much autonomy should I give?”Too much autonomy = anxiety; too little = micromanagement fatigue

The Fixes:

  1. Show reasoning, not personality. Counterintuitively, agents with a friendly/warm tone are perceived as less competent. AI agents should explain their reasoning and cite criteria — clarity builds confidence. (This is the “Pratfall Effect” — too personable reduces credibility in professional contexts.)

  2. Be explicit about limitations. Agents that transparently state what they can and cannot do earn higher trust than those that overpromise. This connects to the glossary/rumpelstiltskin-effect — naming limitations is itself a trust-building act.

  3. Operate in the Goldilocks Zone. The optimal autonomy level is moderate — present options, let humans approve. This maps to Levels of Automation theory (Sheridan & Verplank, 1978): middle levels consistently outperform full automation or full manual control.

Implication for Level Progression:

  • Level 1→2 stall: Users don’t trust standardized AI because they haven’t seen the reasoning
  • Level 2→3 stall: Users won’t delegate execution because autonomy triggers anxiety
  • Level 3→4 stall: Full autonomy requires trust that agents will correctly identify their own limits

The technical capability exists. The blocker is psychological. See glossary/tpb — “perceived behavioral control” is the same construct as delegation anxiety.

Value Chain Elevation

As automation commodifies knowledge work, professionals must move up the value chain:

Commodity → Brand → Service → Experience → Transformation

If AI can do your work at Level 4, your value must come from somewhere AI can’t reach.

First-Mover Advantage Compounds

Companies reaching Level 4 first develop:

  • Superior training data from real operations
  • Better understanding of edge cases
  • Institutional knowledge of what works

This makes catch-up progressively harder.

Budget Doesn’t Determine Level

Cheap models (Minimax, Nemotron, Mistral) can support Level 3-4 deployment. Success depends on initiative and risk tolerance, not budget size.

Self-Assessment Checklist

QuestionIf Yes…
Do team members prompt AI individually with no standards?Level 1
Do you have custom GPTs/Projects shared across the team?Level 2
Can you give AI a spec and walk away for hours?Level 3
Does AI operate on your behalf without daily check-ins?Level 4
Does AI suggest actions before you think to ask?Level 5

Key Takeaways

  • 75-80% of professionals are stuck at Level 1
  • The jump to Level 3 is psychological, not technical
  • Level 4 threatens traditional service businesses (agencies, contractors)
  • Leapfrogging Levels is possible as tools mature
  • First-movers at higher levels compound their advantage

Sources