AI Fluency Levels: 5 Tiers Explained [2026]

AI fluency levels give you something a single composite score cannot: a clear picture of what you can actually do with AI tools right now, and what you need to learn next. Across 2,454 completed AISA assessments to date, the median composite score is 46 — squarely in the Developing tier. Most people overestimate where they sit by nearly 18 points. Understanding the five tiers — Emerging, Developing, Proficient, Advanced, and Expert — turns a vague sense of "I use ChatGPT sometimes" into a concrete development target.

This post walks through each AI fluency level with score ranges, the share of the population that lands there, and the day-to-day behaviours that define each tier. If you manage a team, these descriptions double as a rubric for workforce planning. If you're an individual contributor, they're a checklist for your next quarter.

Why AI Skill Levels Matter More Than a Single Score

A single number — say, 46 out of 100 — tells you almost nothing about what to do next. Tiered AI fluency levels translate raw scores into behavioural profiles: what tasks someone can handle independently, where they need support, and which skills are the bottleneck.

This matters for three reasons.

Scores Without Context Are Noise

Telling an engineer they scored 54.7 is like telling a runner their VO₂ max without explaining what the number means. Tiers provide the interpretive frame. A Developing-tier professional and a Proficient-tier professional may both use AI daily, but the Developing user typically copies outputs verbatim while the Proficient user restructures prompts, checks claims, and selects models deliberately. The tier label encodes that behavioural difference.

Teams Need Distribution, Not Averages

If you're an engineering manager rolling out AI-assisted code review, the team average is less useful than the distribution. A team with eight Proficient members and two Emerging members has a different training need than a team of ten Developing members. AISA's tier system lets you see that distribution at a glance — and plan accordingly. The AI readiness assessment surfaces exactly this kind of team-level breakdown.

Development Plans Need Targets

Moving from Developing to Proficient is a different project than moving from Proficient to Advanced. The skills involved are different, the practice is different, and the time investment is different. Tiers give you a concrete next milestone rather than an abstract "get better at AI."

The World Economic Forum's Future of Jobs Report 2025 found that 59% of workers will need reskilling or upskilling by 2030 — but effective reskilling requires knowing where someone starts. That's the job of a tiered framework.

The Five AI Fluency Tiers: Score Ranges and Behaviours

AISA assigns every candidate to one of five composite tiers based on performance across 11 criteria spanning five dimensions: Prompting & Communication, Critical Thinking, Technical Understanding, Workflow & Application, and Safety & Responsibility. Here's what each tier looks like in practice.

The distribution is heavily left-skewed. Nearly 69% of all candidates land in Emerging or Developing. Only 7.4% reach Advanced or Expert. These numbers align with what Anthropic's AI Fluency Index describes as a "long tail of deep competence" — most professionals have surface-level familiarity, few have integrated AI into their critical workflows.

Emerging (0–27): Awareness Without Application

577 people — 23.5% of all assessments — land here. An Emerging-tier individual knows AI tools exist and may have tried one or two, but hasn't built any repeatable habits. Day-to-day, this looks like:

• Prompting: Types natural-language questions as if searching Google. Doesn't iterate on prompts or provide context. Accepts the first output.
• Critical thinking: Takes AI-generated text at face value. Doesn't check claims, notice hallucinations, or question sourcing.
• Technical understanding: Cannot explain what a language model does at even a high level. Confuses AI products ("ChatGPT is AI") with underlying technology.
• Workflow: Uses AI sporadically, if at all. No integration into existing tools or processes.
• Safety: Unaware of data privacy implications. May paste proprietary code or customer data into public-facing tools without considering the consequences.

The Emerging tier often includes people who are waiting to see if AI matters before investing time. The data suggests the wait is costly — the gap between Emerging and Developing is primarily about practice, not technical aptitude.

Developing (28–59): Active Use, Shallow Depth

1,116 people — 45.5% — sit in this tier, making it the largest by far. The median AISA score of 46 falls here. Developing-tier professionals use AI tools regularly but haven't moved beyond surface-level interaction. Their daily patterns include:

• Prompting: Writes multi-sentence prompts and sometimes includes role instructions or examples. Rarely chains prompts or decomposes complex tasks. Knows about few-shot prompting but doesn't apply it consistently.
• Critical thinking: Recognises that AI can be wrong but checks outputs inconsistently. Tends to verify facts they already suspect are wrong while trusting plausible-sounding claims.
• Technical understanding: Can name several AI tools and roughly describe what a large language model does. Struggles to explain why model selection matters or what a context window affects.
• Workflow: Uses AI for drafting emails, summarising documents, or generating first-pass code. Hasn't redesigned any workflow around AI capabilities.
• Safety: Aware of basic data privacy rules ("don't paste passwords") but unclear on organisational policies, training data implications, or when human oversight is required.

This is the tier where most professionals plateau. The jump to Proficient requires deliberate practice in critical thinking with AI and a shift from using AI as a convenience tool to treating it as a collaborator that needs steering.

Proficient (60–79): Intentional and Structured

579 people — 23.6% — reach Proficient. This is where AI use becomes genuinely productive. Proficient users don't just get answers from AI; they get better answers through structured interaction. Daily behaviours include:

• Prompting: Decomposes complex tasks into prompt sequences. Uses system prompts, output formatting constraints, and iterative refinement. Adapts prompting strategy based on the model and task type.
• Critical thinking: Routinely cross-checks AI outputs against primary sources. Identifies when a model is confabulating versus when it's drawing on training data. Adjusts trust level based on domain and stakes.
• Technical understanding: Understands token limits, temperature settings, and why different models suit different tasks. Can explain the trade-offs between a local model and an API-based service.
• Workflow: Has at least one AI-augmented workflow that's measurably faster or higher-quality than the pre-AI version. Documents prompts and shares them with teammates.
• Safety: Classifies data before deciding which tool to use. Understands the difference between a model's training data and its inference-time inputs. Applies responsible AI practices without needing to be reminded.

Proficient is the tier where organisations start seeing ROI from AI adoption. It's also where the AI proficiency levels discussion gets interesting — because the gap between Proficient and Advanced is less about knowledge and more about systems thinking.

Advanced (80–91): Systems-Level Integration

154 people — 6.3% — reach Advanced. These are the people who don't just use AI tools; they design how their team or organisation uses them. Their daily work looks qualitatively different:

• Prompting: Builds reusable prompt templates and libraries. Tests prompts systematically (A/B variants, different models). Understands prompt injection risks and designs around them.
• Critical thinking: Evaluates AI outputs probabilistically rather than binary right/wrong. Builds verification workflows that scale — checklists, automated checks, peer review gates.
• Technical understanding: Can evaluate new models against specific use cases. Understands fine-tuning trade-offs, RAG architectures, and when to use agents versus single-turn prompts. Follows model releases (like the recent Gemini 4 Carbon or Mistral Large 4) and assesses their relevance to current projects.
• Workflow: Orchestrates multi-step AI workflows across tools. Measures time saved and quality impact. Identifies tasks that shouldn't be delegated to AI.
• Safety: Designs guardrails for team-level AI use. Understands regulatory frameworks (EU AI Act, NIST AI RMF). Can articulate the organisation's AI risk posture to non-technical stakeholders.

Advanced-tier professionals are often the AI champions in their organisations — the people others go to when they're stuck. They're also the ones most likely to have an accurate self-assessment: AISA data shows Founders, who often land near this tier, overestimate their scores by only 7.3 points on average, compared to 17.8 points overall.

Expert (92–100): Frontier Capability

28 people — 1.1% — reach Expert. This is a small, distinct group. Expert-tier individuals operate at the frontier of what's possible with current AI tools and contribute to advancing how AI is used in their domain:

• Prompting: Designs novel prompting strategies for unsolved problems. Contributes to prompt engineering research or internal knowledge bases that shift team-wide practice.
• Critical thinking: Builds evaluation frameworks for AI outputs that others adopt. Can identify subtle failure modes — sycophancy bias, cascading errors in multi-step chains, distributional shift in model behaviour.
• Technical understanding: Deep fluency across the stack — from transformer architecture to deployment infrastructure to token economics. Can make build-vs-buy decisions on AI infrastructure with confidence.
• Workflow: Architects end-to-end AI-augmented systems. Designs human-in-the-loop patterns that balance automation with oversight. Measures second-order effects of AI integration.
• Safety: Shapes organisational AI policy. Engages with emerging safety research — like the recent Anthropic/UK AISI findings on unsanctioned agent actions — and translates it into practical safeguards.

Expert is not a permanent state. As models evolve and new capabilities emerge, the bar for Expert-level fluency shifts. What matters is the pattern of behaviour: continuous learning, systematic evaluation, and the ability to teach others.

How to Move Up One AI Fluency Tier

The most actionable question isn't "What tier am I?" but "What do I need to do to reach the next one?" Here are concrete micro-goals for each transition, based on the dimension averages we observe across 2,454 assessments.

Emerging → Developing: Build the Daily Habit

The biggest barrier at this stage is inconsistency. The goal is to make AI use automatic, not occasional.

1. Pick one recurring task — email drafting, meeting notes, code review — and commit to using AI for it every day for two weeks.
2. Learn basic prompt structure. Read the prompting skills guide and practice writing prompts with a clear role, task, and output format.
3. Start noticing errors. Before accepting any AI output, ask: "Is there a claim here I should check?" You don't need to check everything — just build the reflex.
4. Understand one tool's data policy. Read the privacy page for whichever AI tool you use most. Know whether your inputs are used for training.

Developing → Proficient: Go From User to Operator

This is the hardest jump for most people. The average dimension score for Technical Understanding across all assessments is 39.2 — the lowest of all five dimensions. Closing that gap is often what separates Developing from Proficient.

1. Learn how models work at a conceptual level. The How AI Works guide covers the essentials: tokens, context windows, probability distributions, and why models hallucinate.
2. Redesign one workflow end-to-end. Don't just add AI to an existing process — rethink the process with AI as a core component. Document the before and after.
3. Practice structured verification. Build a personal checklist: source check, logic check, completeness check. Use it on every AI output that goes to a stakeholder.
4. Try a second model. If you only use one AI tool, try another for the same task. Comparing outputs builds intuition about model differences.

Proficient → Advanced: Think in Systems

1. Build a prompt library. Collect your best prompts, version them, and share them with your team. Track which ones produce consistent results.
2. Measure impact. Pick one AI-augmented workflow and measure it: time saved, error rate, output quality. Use real numbers, not vibes.
3. Learn about agents and multi-step workflows. Understand when to chain prompts, when to use agentic workflows, and when a single well-crafted prompt is enough.
4. Engage with safety and governance. Read your organisation's AI policy. If there isn't one, draft one. Understand the EU AI Act risk categories relevant to your domain.

Advanced → Expert: Contribute to the Field

1. Publish or teach. Write internal documentation, give talks, or mentor others. Teaching forces you to articulate what you know — and reveals what you don't.
2. Evaluate new models systematically. When a new model drops, don't just try it — benchmark it against your specific use cases with consistent test prompts.
3. Design evaluation frameworks. Build rubrics or automated checks that your team uses to assess AI output quality at scale.
4. Stay current on safety research. Follow developments like the Anthropic/UK AISI containment analyses and translate findings into practical team guidance.

How AISA Assigns Your AI Fluency Tier

AISA uses a conversational assessment format where a candidate talks with an AI facilitator while a separate AI evaluator scores independently. This separation is deliberate — it prevents the conversation from being optimised for scoring rather than genuine demonstration of skill.

The Scoring Architecture

Every score is linked to specific evidence in the candidate's own words. The AISA rubric defines 11 criteria across five dimensions, each scored 1–10:

• Prompting & Communication (23% of composite): How you structure requests, iterate, and adapt your communication to the model.
• Critical Thinking (22%): How you evaluate, verify, and challenge AI outputs.
• Technical Understanding (20%): Your grasp of how AI systems work and their limitations.
• Workflow & Application (25%): How you integrate AI into real tasks and processes.
• Safety & Responsibility (10%): Your awareness of risks, data handling, and ethical considerations.

The weighted composite maps to the five tiers. The framework is cross-referenced against Anthropic's AI Fluency Index (93% marker coverage) and the U.S. Department of Labor AI Literacy Framework (100% coverage).

Why the Results Hold Up

98% of people who retake AISA land in the same or an adjacent tier, and scores correlate at r = 0.79 between attempts. The assessment has been audited at 4.6/5 against the AERA/APA/NCME Standards and ISO 10667 (quality framework details). This stability matters because a tier assignment is only useful if it reflects something real about your capabilities — not your mood on a given day.

Integrity checks run on every session, and the assessment design is built to resist gaming. You can't memorise answers because there are no fixed questions — the facilitator adapts to your responses.

What You Get

Every completed assessment produces a detailed report breaking down your scores by dimension and criterion, your assigned persona (one of 10 behavioural profiles), and your composite tier. You also receive a certificate — useful if you're pursuing AI certification for professional development or job applications.

The dimension-level breakdown is where the real value lives. Someone who scores Proficient overall but has a Technical Understanding score in the Developing range has a very specific development target. The report makes that visible.

Where Most People Actually Land — and What It Means

The population distribution tells a clear story. With an average composite of 46.3 and a median of 46, the typical AISA candidate is solidly mid-Developing. The prediction gap data adds another layer: across 1,544 candidates who estimated their score before taking the assessment, the average predicted score was 62.3 against an average actual of 44.5 — an overestimation of 17.8 points.

This gap is largest among Students (31.4 points) and smallest among Founders (7.3 points). Stanford's 2024 AI Index Report noted that self-assessed AI competence correlates poorly with demonstrated skill across multiple studies — AISA's data confirms this pattern at scale.

The practical implication: if you think you're Proficient, there's a meaningful chance you're actually Developing. The only way to know is to take the assessment and get an evidence-based score. The report you receive isn't a pass/fail — it's a map of where you are and what to work on next.

For teams, the distribution data helps set realistic expectations. If 45.5% of the general population lands in Developing, a team that averages Proficient is already well above the baseline. An AI skills assessment at the team level reveals whether your group matches, exceeds, or falls below these benchmarks — and which dimensions need the most attention.

---

Related reading: AI Proficiency Levels Explained — a deep dive into criterion-level scoring and what each proficiency band means.

Related reading: What Is a Good AI Score? 2026 Benchmarks — how to interpret your composite score against role-specific and industry benchmarks.

Related reading: Am I Falling Behind on AI? — data on where professionals actually stand, and whether the urgency is justified.

Frequently Asked Questions

How many AI fluency levels are there?

AISA defines five AI fluency levels: Emerging (0–27), Developing (28–59), Proficient (60–79), Advanced (80–91), and Expert (92–100). Each tier corresponds to a distinct set of observable behaviours across prompting, critical thinking, technical understanding, workflow integration, and safety awareness. The tiers are derived from a weighted composite of 11 scored criteria.

What percentage of people reach Advanced AI fluency?

Across 2,454 completed AISA assessments, 6.3% of candidates reach the Advanced tier (scores 80–91) and just 1.1% reach Expert (92–100). The largest group — 45.5% — lands in Developing. Reaching Advanced typically requires not just regular AI use but systems-level thinking: building reusable workflows, measuring impact, and designing team-level AI practices.

Can I skip a level?

Technically, yes — AISA assigns your tier based on your demonstrated performance, not on a prerequisite sequence. If you score 72 on your first assessment, you're Proficient regardless of whether you were ever formally "Developing." In practice, though, the skills build on each other. The behaviours that define Proficient (structured prompting, systematic verification, workflow redesign) depend on habits that are typically established during the Developing stage. Most people progress through adjacent tiers rather than leaping two levels.

How often should I reassess my AI fluency level?

We observe that meaningful skill changes — enough to shift a tier — typically take focused effort over several weeks or months, not days. A reasonable cadence is every three to six months, or after completing a structured learning programme. Since 98% of retakers land in the same or adjacent tier, reassessing too frequently is unlikely to show movement. Focus on the micro-goals for your current tier transition, then reassess when you've put in the work.