The Holak (and Gomulski) Scale 3.0

Where the name comes from. I have built this scale with Konrad Gomulski's help from the start. With version 3.0 Konrad got much more involved - his questions, counterarguments and reviews genuinely shaped this model. So we decided it was time to fix the name: from this version it is the Holak (and Gomulski) Scale. The reasoning behind the update is in the post The Holak (and Gomulski) Scale 3.0 - fewer levels, more meaning.

TL;DR

Five levels instead of twelve. The biggest change: a level describes a context, not a person.

0

Resistance

1

Basics

2

Conscious use

3

Advanced processes

4

Autonomy / mature adoption

Why version 3.0

The previous version of the Holak Scale described the detailed road of AI adoption well: from a simple chat, through prompting, instructions, context, skills, tools, agentic workflows, orchestration and an agentic OS. In practice that detail came at a cost. Instead of talking about real changes to their work, people started arguing about level numbers. Are we at 6 or at 7? Does MCP automatically grant level 8? Can an organisation be at 2 if a single person is at 9?

Version 3.0 answers these problems through simplification. Instead of many rungs we have five main levels. The detailed elements of the previous scale do not disappear, but they become diagnostic sub-areas. We do not score them separately. We use them to talk about the maturity of a specific process.

Core principle: a level is contextual

The same person can be:

We do not ask "which level are you at?". We ask "which level is this specific area of AI use at?".

How to read the levels

Sub-levels are not scored. The point is not a 7/10 result or ticking a checklist. Each level has a few maturity sub-areas. Understanding some of them lets you enter a level. Moving to the next one, however, requires understanding and applying the entire level below.

The exception is level 0 - it describes the state before conscious adoption, or a conscious decision not to use AI in a given area.

Level 0 - Resistance

AI is not used, or it is used only accidentally, quietly, without rules or against the official policy. Resistance may stem from a lack of knowledge, fear, bad experiences, regulatory limits, ethical reasons or the absence of a clear use case.

Level 0 is not an insult. Sometimes a conscious "we do not use AI in this process" is more mature than uncontrolled use of AI on sensitive data.

Sub-areas

• No use - AI is not part of daily work or private life.
• Unconscious use - someone uses AI without thinking about what is allowed, what the risks are and where the data goes.
• Emotional resistance - the barrier is not technology, but fear, distrust or a sense of threat.
• Organisational resistance - no policy, owner, rules or space for safe experimentation.
• Conscious opt-out - the risk, cost or regulatory aspect outweighs the potential benefit.

Exit signal

You can move to level 1 when you can point to a safe AI use case and name the basic boundaries: what not to paste, what not to trust without checking, when AI should not be used and who is responsible for the result.

Typical trap: pretending AI does not exist while people use it quietly anyway. That is not security. That is a lack of control.

Level 1 - Basics

AI is used for simple, ad hoc tasks: a question, an answer, a summary, an email, an idea, a quick explanation, a simple text transformation. The user starts to understand that AI is useful, but does not yet treat it as part of a process. Many advanced users return to level 1 for simple tasks - the difference is that they do not stay there permanently.

Sub-areas

• Simple chat - a question and a useful answer.
• First use cases - an email, a summary, ideas, a simple translation, a document draft, an explanation of a concept.
• Basic verification - awareness that an answer may be wrong despite a convincing form.
• Basic data hygiene - personal data, company secrets, passwords, medical and financial data are not pasted thoughtlessly.
• Recognising limits - AI is not a search engine, a source of truth or an expert with professional liability.

Typical trap: "I use AI every day, so I am advanced". You can ask simple questions every day for a year and still be at level 1.

Level 2 - Conscious use

AI stops being treated like a magic box for typing questions. The user starts designing the instruction: a goal, context, role, constraints, input, expected format and quality criteria. They iterate, compare answers, correct the model and use persistent instructions. This is the level where a real jump in quality appears - without building integrations, agents and complex processes.

Sub-areas

• Goal and context - why, for whom, in which domain, with what constraints and success criteria.
• Response format - a table, a risk list, test cases, an action plan, a checklist, JSON, a user story, a report.
• Iteration and critique - improvements, variants, counterarguments, edge cases, assumption checks, risk assessment.
• Templates and frameworks - repeatable work patterns instead of inventing the prompt from scratch.
• Custom instructions - persistent preferences, style, role and standards go into settings, the profile and project instructions.
• Model awareness - one model for code, another for reasoning, another cheap, another local or with a large context.
• Cost and token hygiene - you do not pay every day for the same 200 lines of context.

Typical trap: the prompt fetish. A team builds a library of 200-line "golden prompts" but still copies them by hand instead of moving the persistent context into instructions and documentation.

Level 3 - Advanced processes

AI stops being just a conversation and becomes part of a process. Durable context, organised knowledge, tools, integrations, automations, skills, evaluations and security rules appear. AI reads files, analyses the repository, uses documentation, creates tasks, runs tests and supports the workflow.

Tools before skills. Before you build specialised skills and knowledge bases, you have to understand what AI actually works with, what permissions it has and how to control its actions. A skill without an understanding of tools becomes another nice prompt. A tool without rules becomes a risk. A process emerges only when both layers come together.

Sub-areas

• Durable context - project files, documentation, ADRs, README, AGENTS.md, CLAUDE.md, a knowledge base, the repository.
• Advanced operational instructions - when to ask, when to act, when to escalate, how to report, when to stop.
• Tools, integrations and connectors - access to repositories, documents, Jira, the calendar, APIs, the test environment, CI/CD - limited and justified.
• Hooks and control automations - run tests, check the format, block a forbidden command, ask for approval, write a log, enforce review.
• Skills and repeatable capabilities - log analysis, bug triage, test generation, code review, release notes.
• Knowledge bases and RAG - the knowledge has an owner, a scope, freshness, sources and update rules.
• Evaluations - a few representative cases and an expected result show whether a prompt or skill still works after a change.
• Security and ownership - who is responsible, which data may be used, which actions require approval, where the log is, how to stop the process.

Typical traps

• Documentation graveyard - the context exists, but nobody updates it.
• Integrations without a process - AI has access to tools but no clear workflow.
• Skill bloat - a dozen skills that nobody uses or maintains.
• MCP as a trophy - there are connectors, but no control, metrics or real value.

Level 4 - Autonomy / mature adoption

AI delivers limited goals end-to-end. The human does not steer the model at every step - they define the goal, the boundaries, the success criteria and the way to verify. AI plans, acts and reports, while the human approves the outcome, checks the risks and corrects the system. This is not full, unsupervised autonomy. It is mature adoption, in which autonomy is limited, traceable and reversible.

Sub-areas

• Delegating goals, not steps - "prepare the release notes", "analyse production errors", not "do A, then B, then C".
• Agentic workflows - gather data, analyse, run the steps, produce the result, give the rationale and flag risks.
• Traceability - you can see what AI did, on what data, what decisions it made and what the cost was.
• Reversibility - a wrong action can be undone or its effects contained.
• Human in the loop where needed - approving high-risk actions: production, money, sensitive data, legal and HR decisions.
• Orchestration only where it makes sense - multiple roles only once a single agent reliably delivers.
• Agentic OS / operating model - roles, owners, metrics, costs, logs, reviews, a feedback loop and a stop procedure.
• Ethical and organisational maturity - impact on people, privacy, bias, compliance, the boundaries of automation and the right to refuse AI.

Typical traps

• The illusion of autonomy - AI works "on its own", but the human later fixes 60% of the result.
• Orchestrating mediocrity - many agents cooperate, but none does its own part well.
• An OS without a goal - a flashy platform, but it is unclear which process it actually improves.
• Autonomy without accountability - the agent acts, but nobody owns the consequences.

Migration map from v2.1 to v3.0

The old levels were not thrown away. They were grouped into five new ones. Below is a visual map of the 12 old rungs collapsing into 5 new levels.

0 - Resistance

• 0 - Resistance / no adoption

1 - Basics

• 1 - Basic chat

2 - Conscious use

• 2 - Conscious prompting
• 3 - Prompting frameworks
• 4 - Instructions and token hygiene

3 - Advanced processes

• 5 - Project/home context
• 6 - Advanced instructions
• 8 - Tools, MCP, hooks
• 7 - Skills, knowledge, evaluations

4 - Autonomy

• 9 - Agentic workflows
• 10 - Orchestration
• 11 - Agentic OS

Minimal 3.0 diagnostics

Instead of asking "which level are you at?", use a short diagnosis.

The context matrix

Version 3.0 should be used as a matrix, not a single label. An example of a filled-in diagnosis:

The value and cost curve

Levels do not carry equal value or equal cost. The biggest return for most people and teams comes from reaching level 2. The higher you go, the faster cost and required discipline grow relative to the extra value.

The biggest return is on the transitions 0 → 1, 1 → 2 and 2 → 3. Level 4 is for selected processes, not for everything.

Value Cost and complexity

The key principles of the Holak (and Gomulski) Scale 3.0

1. A level is contextual. You are not "at 3". You can be at 3 in QA, at 2 in document analysis and at 1 in communication.
2. A lower level is not morally worse. Sometimes level 0 is a conscious decision, and level 4 would be irresponsible.
3. You do not have to reach 4. For many people and processes, level 2 or 3 gives the best value-to-cost ratio.
4. Tools do not mean maturity. MCP, connectors, skills and agents are means, not proof of adoption.
5. Autonomy requires context. You cannot responsibly jump from good prompting to a self-driving agent without rules, tools, evaluations and control.
6. The scale measures capability, not ethics. That is why mature adoption must include a separate dimension of accountability, privacy and the boundaries of automation.
7. Evidence beats a declaration. The best test: show a concrete result from the last week and explain what AI actually improved.

Sources and context

This model was created as a simplification and rebuild of the earlier versions:

• The Holak (and Gomulski) Scale 3.0 - fewer levels, more meaning (the update post)
• The Holak Scale v2.1e - enterprise
• The Holak Scale v2.1p - private
• Where the Holak Scale fails - a self-critique

Version 3.0. Co-created by Grzegorz Holak and Konrad Gomulski.