Skill Progression Map
What Level 1, 2, and 3 look like for your role — concretely.
How to Use This Page
The Reference Framework defines three maturity levels. This page shows what those levels mean in practice for specific role families — not in theory, but in the work you do day to day.
For each role family:
- Level 1 (AI-Assisted): You use AI as a tool. Same workflows, faster in spots.
- Level 2 (AI-Integrated): AI is embedded in your workflow. Some of your work has been redesigned around what AI can do.
- Level 3 (AI-Native): You define specifications and judge results. AI handles execution.
The individual tier scale adds intermediate stages: Tier 0.5 (AI-Curious), Tier 1.5 (AI-Building), and Tier 2.5 (AI-Advanced). If you find yourself between two levels below, you're likely at the .5 stage – actively transitioning. That's progress, not a gap.
Find your role family below. Identify where you are. Then use Transforming Your Role for the transition process and the Recognizing Your Pattern section to understand which structural forces are acting on your role.
A note on depth. The Engineering column has the most operational substance because that's where AI-native maturity is best documented. The Customer Service, Marketing, Sales, and Design columns reflect the framework's current understanding — directionally right, but lighter. Parallel readiness models for the discrete-task domains where this maturity exists most clearly — engineering, customer service, finance operations, legal review, knowledge research — are on the framework's roadmap; until they ship, treat the non-engineering columns as the floor of what your role looks like at each level, not the ceiling. See also the framework's broader note on this asymmetry.
Engineering
Dominant pattern: Elevation — from writing code to specifying what the code should do.
Level 1 — AI-Assisted
You use AI for code completion and quick lookups. Copilot or ChatGPT suggests lines; you accept or reject.
What this looks like:
- AI autocompletes code as you type
- You paste code into ChatGPT to debug or explain
- AI outputs require significant manual review and editing
- No shared configurations or prompt templates across the team
- Your workflow is fundamentally the same as before AI
The data: 84% of developers use or plan to use AI tools, 51% daily. But trust has fallen to 29%, and 66% report spending more time fixing AI-generated code than they save (Stack Overflow, 2025). This is the Level 1 experience: AI helps in spots, but the net gain is uncertain because the workflow hasn't been redesigned.
Level 2 — AI-Integrated
AI is part of the development workflow, not just a helper. You direct multi-file changes, review AI-generated code at the PR level, and maintain shared context files.
What this looks like:
- AI generates code from descriptions; you review and iterate
- Shared prompt templates and context files exist for the codebase
- AI handles tests, documentation, and boilerplate systematically
- You spend more time on architecture and review, less on typing
- Removing AI would break your development velocity
The shift: At Level 2, you accept ~30% of AI suggestions but retain 88% of generated characters (GitHub/Accenture, 2024). The skill is knowing what to accept, what to reject, and how to direct the generation.
Level 3 — AI-Native
AI is the execution layer; humans give direction and validate. Work is structured around a recurring operational unit (Context → Clarification → Execution → Validation → Recovery) and your value concentrates at the boundaries. This corresponds to Rungs 4–5 on the engineering scale.
Unit of work: the feature or user story, handled end-to-end inside one agent loop (architect → implement → review → merge). Not the line of code. Not the prompt.
Cycle time: stories ship in hours-to-days; features in days-to-weeks. "Behind on a deliverable" loses its old meaning — when a project stalls at L3, the cause is rarely human capacity.
What this looks like:
- You define specifications, acceptance criteria, and constraints
- AI produces the implementation, runs tests, opens the PR, and resolves review comments
- A separate agent reviewer validates the PR; you intervene only on flagged issues or final UX validation
- Validation gates are risk-graded — agent-only review for reversible work, human approval for irreversible changes (production deploys, sensitive data, customer-facing communications)
- See the AI Lab for the operational unit in detail
Failure mode: when a deliverable stalls, the cause is usually the AI bottleneck — the agent has hit a structural limit (wrong direction, ambiguous spec, subjective edge case it cannot resolve alone). The recovery is recalibration, not debugging: a brainstorm session that rebuilds the AI's understanding of the problem, often with multiple humans bringing different perspectives. Throwing more humans at execution doesn't help.
Day shape: A typical L3 day concentrates work at the front and back boundaries. Morning: review yesterday's overnight agent output, validate two PRs the agent reviewed, run first-user UX testing on a feature that just shipped. Midday: write specs for two new stories; engage a clarification dialogue with the agent until no material ambiguity remains. Afternoon: a recalibration session on a stuck story; refine acceptance criteria for next sprint. Almost no time is spent watching the agent execute.
Metrics: throughput (PRs merged per week, stories shipped), quality (defects per story, scenario coverage), cost (token cost per merged PR, AI gross margin) — not "time saved by AI." See AI economics at maturity.
The critical warning: Level 1 without progression actively degrades quality. Analysis of 211 million lines of code shows that AI-assisted development without skill progression caused refactoring to drop from 25% to under 10% of changes, while code churn nearly doubled (GitClear, 2025). The tools make it easy to produce code and hard to produce good code. The productivity story is also bimodal: empirical studies measuring individuals at L1–L2 adding AI to existing workflows find gains are spurious or negative; case studies of teams restructured around AI execution (AMPECO, Monte Carlo, Every) report 4× gains and 73% PR-rate increases. Level 2 and 3 skills — review judgment, specification quality, test design, process design — are what unlock the second mode and prevent the first.
Self-assessment
| Question | Level 1 | Level 1.5 | Level 2 | Level 2.5 | Level 3 |
|---|---|---|---|---|---|
| How do you start a new feature? | Open editor, start coding, use AI for completion | Experiment with AI for parts of the feature, building and testing prompt workflows | Describe the feature to AI, review the output, iterate | Write detailed specs with constraints, moving toward test-verified output | Write a specification with constraints and test cases, let AI implement |
| What happens when AI code is wrong? | Fix it line by line | Iterate on the prompt, starting to build reusable templates | Improve the prompt/context and regenerate | Improve scenarios and verification systems | Recalibrate (rebuild the AI's understanding via brainstorm + re-spec) before debugging — the spec or context is usually the actual cause |
| What do you share with teammates? | Nothing AI-specific | Experiments that worked, prompt drafts | Prompt templates, context files | Specification patterns, verification approaches | Specification patterns, scenario libraries |
Marketing
Dominant pattern: Specialization — shedding content production, deepening strategic judgment.
Level 1 — AI-Assisted
You use AI for first drafts and idea generation. Every output gets manually edited.
What this looks like:
- AI generates blog post drafts, email copy, or social media posts
- You edit 80%+ of AI output before publishing
- No systematic workflow — AI is used ad hoc
- Each team member uses AI differently (or not at all)
- Campaigns are still planned and executed the traditional way
The data: 91% of marketing leaders say their teams use AI, with content creation (43%) as the top use case. But 86% edit AI-generated content before publishing (HubSpot, 2025). And 68% receive no formal AI training (Marketing AI Institute, 2025).
Level 2 — AI-Integrated
Campaign workflows are redesigned around AI. AI doesn't just draft — it generates variants, handles research, and produces analysis as a systematic step.
What this looks like:
- Shared prompt libraries encode brand voice and positioning
- AI generates campaign variants; you select and refine
- Research, competitive analysis, and reporting are AI-first workflows
- New team members are onboarded into AI-integrated processes
- The team produces more with fewer people
The shift: You stop writing content and start directing content systems. Your value moves from production speed to strategic judgment: which angle, which audience, which positioning.
Level 3 — AI-Native
You define strategy, positioning, and constraints. Systems produce campaigns, variants, and reports.
What this looks like:
- You specify the campaign: target, positioning, constraints, success metrics
- AI produces the creative, copy, and distribution plan
- You review, select, and adjust — not produce
- The marketing team is significantly smaller but produces significantly more
- Your role is strategy and taste, not execution
External validation: The Marketing AI Institute's own maturity survey maps almost directly to these levels: 40% of marketing teams are at Experimentation (Level 1), 26% at Integration (Level 2), 17% at Transformation (Level 3) (Marketing AI Institute, 2025).
Self-assessment
| Question | Level 1 | Level 1.5 | Level 2 | Level 2.5 | Level 3 |
|---|---|---|---|---|---|
| How do you create a campaign? | Plan it, then use AI for some drafts | Test AI for specific steps, building prompt libraries | Define the brief, AI generates variants, you curate | Define strategy and constraints, AI produces most deliverables with light editing | Define the strategy and constraints, AI produces the campaign |
| What's your bottleneck? | Writing and production | Finding which AI workflows stick | Review and strategic decisions | Defining the right constraints for consistent quality | Defining the right problem to solve |
| How much do you edit AI output? | 80%+ | 50–70% (improving as workflows mature) | 30–50% | 15–25% (mostly selecting, not rewriting) | 10–20% (selecting, not rewriting) |
Customer Service
Dominant pattern: Elevation shifting to Convergence — from answering tickets to designing service systems.
Level 1 — AI-Assisted
AI suggests responses. Agents copy, paste, and edit. The workflow is the same, slightly faster.
What this looks like:
- AI drafts reply suggestions for agents
- Agents handle the same volume and types of interactions
- AI handles only the simplest, most scripted inquiries
- No role changes — everyone still does the same job
- Quality depends on individual agents, not systems
Level 2 — AI-Integrated
AI handles routine inquiries autonomously. Agents shift from answering to training, reviewing, and handling complex cases. New roles emerge.
What this looks like:
- AI resolves the majority of routine tickets without human involvement
- Agents spend more time training AI systems than doing traditional support
- New roles emerge: conversation analysts, knowledge managers, AI operations leads
- Escalation logic is designed and documented, not improvised
- The team handles significantly more volume with stable or reduced headcount
The data: 82% of support teams feel positive about AI collaboration. 60% say roles are evolving. 40% of teams report agents spend more time training AI systems than doing traditional support (Intercom, 2025). This is Level 2 in action.
Level 3 — AI-Native
Humans define service strategy, escalation logic, and quality standards. AI executes the vast majority of interactions.
What this looks like:
- You define: what constitutes good service, when to escalate, what quality looks like
- AI handles 80%+ of interactions
- Human agents exist for judgment calls, relationship moments, and cases the system can't handle
- The team is a fraction of its previous size, but service quality is equal or better
- Your role is system design and quality ownership, not ticket resolution
Customer service is often the first function to reach Level 3. It shows the highest actual AI task coverage (Anthropic, 2026) and generates the largest share of AI value (38%, per BCG, 2025).
Self-assessment
| Question | Level 1 | Level 1.5 | Level 2 | Level 2.5 | Level 3 |
|---|---|---|---|---|---|
| What do you do most of the day? | Answer tickets | Test AI on ticket categories, build response templates | Train AI, handle escalations, review quality | Design escalation logic, monitor AI quality metrics | Design service strategy and escalation rules |
| What happens when AI gives a bad answer? | Fix it and move on | Build a better template or knowledge base entry | Update the training data or knowledge base | Redesign the quality criteria or training data | Redesign the escalation logic or quality criteria |
| How is your performance measured? | Tickets resolved, response time | Template quality, AI adoption rate | AI deflection rate, escalation quality | System design quality, quality at scale | Service quality metrics, system design effectiveness |
Sales
Dominant pattern: Specialization — shedding administrative overhead, deepening relationship and deal judgment.
Level 1 — AI-Assisted
AI helps with email drafts and basic research. Sellers still spend most of their time on non-selling tasks.
What this looks like:
- AI drafts cold emails and follow-ups
- Research is semi-manual with AI assistance
- CRM is updated by humans
- 70% of time goes to non-selling tasks (Salesforce, 2024)
- The sales process hasn't changed, just individual tasks
Level 2 — AI-Integrated
AI automates research, outreach sequencing, and CRM enrichment. Sellers focus on relationships and complex deal strategy.
What this looks like:
- AI handles prospecting research, outreach sequences, and follow-up timing
- CRM is enriched automatically with AI-gathered data
- Sellers focus on high-value conversations: qualification, negotiation, closing
- AI users are 2.4× less likely to feel overworked
- The non-selling overhead drops significantly
The shift: The value moves from activity volume (calls made, emails sent) to deal quality (pipeline accuracy, win rate, deal size). Level 2 sellers don't work harder — they work on the right things.
Level 3 — AI-Native
Humans define qualification logic, deal rules, and escalation thresholds. AI produces outreach, proposals, and pipeline analysis.
What this looks like:
- You define: ideal customer profile, qualification criteria, pricing rules, escalation conditions
- AI produces: outreach, follow-ups, proposals, competitive analysis
- Your time goes to relationship building, strategic accounts, and judgment calls
- By 2027, 95% of seller research workflows are predicted to begin with AI (Gartner, 2025)
- Sellers who partner with AI are 3.7× more likely to meet quota
Self-assessment
| Question | Level 1 | Level 1.5 | Level 2 | Level 2.5 | Level 3 |
|---|---|---|---|---|---|
| How much time do you spend on admin? | 70%+ | 50–60% (actively automating tasks) | 30–40% | 15–25% (most admin is system-handled) | Under 15% |
| How do you research a prospect? | Manually, with some AI help | Build AI research workflows, test automation | AI produces the research brief, you review | AI handles end-to-end research, you review and strategize | AI identifies and qualifies prospects, you handle relationships |
| What's your competitive advantage? | Activity volume | Workflow experimentation | Deal judgment | System design for the sales process | Specification of what "good" looks like |
Design
Dominant pattern: Elevation — from pixel production to system direction.
Level 1 — AI-Assisted
AI generates mood boards, initial concepts, or copy. You refine everything manually.
What this looks like:
- AI produces inspiration: mood boards, concept variations, style explorations
- All production work (layouts, components, assets) is done manually
- AI is a starting point, not a workflow participant
- The design process is unchanged — AI adds a brainstorming step
Level 2 — AI-Integrated
AI handles production work. You shift to system thinking, brand direction, and quality judgment.
What this looks like:
- AI generates layouts, asset variations, and responsive adaptations
- You define design systems and brand constraints; AI operates within them
- Production time drops dramatically; review and direction time increases
- Entry-level production roles contract as AI absorbs that work
- 71% of UX professionals believe AI will shape the future of UX (UX Design Institute, 2025)
The shift: Your value moves from craft execution to taste and system design. You're not less of a designer — you're more of an architect.
Level 3 — AI-Native
You define systems, constraints, and brand rules. AI produces the artifacts.
What this looks like:
- You specify: design system, brand parameters, constraints, quality criteria
- AI produces: mockups, components, responsive layouts, asset libraries
- You review, curate, and refine — not draw
- Outcome-oriented design replaces pixel-level work
- "Systems Architects" and "AI Directors" emerge as the high-value design roles (NN/g, 2025)
By Q3 2025, "manual pixel-pushing had effectively ended for commercial production" (UX Design Institute, 2025). The progression from Level 1 to Level 3 is happening faster in design than in most other functions.
Self-assessment
| Question | Level 1 | Level 1.5 | Level 2 | Level 2.5 | Level 3 |
|---|---|---|---|---|---|
| What do you produce? | Pixel-perfect deliverables | Mix of manual and AI-assisted deliverables | Design systems and direction, AI produces deliverables | Design systems and constraints, AI produces most assets | Specifications and quality criteria |
| What's your bottleneck? | Production time | Finding which AI tools work for your workflow | Making the right design decisions | Defining constraints that produce consistent brand quality | Defining the right constraints |
| What skills are growing? | Tool mastery | AI tool integration, prompt design | System thinking, brand judgment | Constraint specification, quality judgment at scale | Specification engineering, taste at scale |
Cross-Cutting: The Skills That Matter at Every Level
Regardless of your role family, certain skills compound across the progression:
Level 1 → Level 2: The critical skill is recognizing which parts of your work are legacy patterns — repeatable execution that AI can absorb. The transition is about seeing the opportunity, not just using the tool.
Level 2 → Level 3: The critical skill is specification engineering — writing clear enough instructions that AI can execute without real-time supervision. This is the Universal Translation Rule applied to your individual work.
At every level: The five irreplaceable functions — Direction, Judgment, Taste, Relationship, Accountability — define what stays human. Your progression isn't about doing less. It's about concentrating on what only you can do.
Workers in AI-exposed roles earn up to 30% salary premiums (PwC, 2025). The market is already pricing the progression.
← Back to home · Transforming your role · AI Execution Standards · Specification Guide