Seven levels of working with AI. Results from 82 people at AI × Voices #1

The quiz was part of my talk "AI Powered Company" at the first AI × Voices, a monthly AI conference in Kraków. The room answered on their phones at q.aixv.ai, with no login. I showed one question at a time on the screen, and the room saw live how many people picked each answer. After every question that could change a level, the screen showed the stairs with the whole room.

Everyone got the same 11 questions in the same order and picked one answer each. The phone asked at the start: "Pick what you actually do, not what you know how to do". So the scale describes practice, not knowledge about AI.

A level says how far AI reaches into someone's work: from no contact, through pasting answers, to work that AI starts and finishes on its own. The last column is the thing for tomorrow from the talk, one step of about fifteen minutes.

1. 1

   I don't use AI, but I'm at AI × Voices

   You are at an AI conference, but you don't use AI day to day yet.

   Example

   you write emails and texts yourself, without an AI chat

   For tomorrow

   Install an AI app and make a funny picture.

2. 2

   I'm trying

   You use AI now and then, but its answers don't make it into your work yet.

   Example

   you ask ChatGPT out of curiosity, but don't use the result

   For tomorrow

   Give AI a boring task and grade it like an intern's work.

3. 3

   I ask

   You ask AI and paste its answers into your work.

   Example

   you paste an email into ChatGPT and the answer back

   For tomorrow

   Tell AI who the text is for and what the reader should do.

4. 4

   Step by step

   You tell AI what to fix and refine the result until it fits.

   Example

   you tell AI "shorter", "no emoji" or "use the formal form"

   For tomorrow

   Write one procedure down as a skill, a file of instructions for AI.

5. 5

   I create with an agent

   You give AI a task, and it does several things in a row on its own: reads, checks and saves.

   Example

   you hand AI a task, and it reads your email and replies itself

   For tomorrow

   Ask AI for three weak spots in what it did.

6. 6

   AI works, I check

   AI starts work on its own, and you check the result.

   Example

   AI prepares a reply on its own, and you check it

   For tomorrow

   Write down what you approve without reading.

7. 7

   AI works without me

   AI finishes work without you, and you read the summary.

   Example

   Gryzik answers a student on its own when it is sure

   For tomorrow

   Add an alert to whatever runs on its own.

1. 1 · I don't use AI, but I'm at AI × Voices2 people · 2%
2. 2 · I'm trying3 people · 4%
3. 3 · I ask2 people · 2%
4. 4 · Step by step14 people · 17%
5. 5 · I create with an agent33 people · 40%
6. 6 · AI works, I check6 people · 7%
7. 7 · AI works without me22 people · 27%

Shares are out of the 82 people who answered at least one question. Levels are ways of working, not points, so there is no average.

Levels 6 and 7 are open only to people whose AI starts work on its own. 28 people stand on them, 22 of them on level 7, where AI also finishes the work without a person. Levels 1 to 3, from no AI to pasting answers, hold 7 people together.

According to CBOS, in May 2026 36% of adults in Poland had used generative AI in the previous three months. In the room it was 98%, meaning any answer to the first question other than "Not once". The period is the same, three months, but the questions are worded differently. The room was people who signed up for an AI conference themselves, not a representative sample. The comparison shows a difference, not its cause.

Each answer carries its value: the level that answer confirms. An answer with no value neither raises nor lowers the level. The bar shows how many people picked that answer, out of those who answered the question.

During the talk the whole room answered, including people whose level a question could no longer change. At questions 4, 7, 9, 10 and 11 I asked people who don't use AI to pick a specific answer, such as "I don't use AI". So the share of those answers also says how many such people were in the room.

1. Question 1 of 11 · 81 answers

   How often in the last three months did you use AI?

   For example ChatGPT, Gemini, Claude, Copilot, Meta AI in WhatsApp, on your phone too.

   What it measures: How often someone uses AI. This question sets the path: "Not once" is level 1, "A few times, to try it" is level 2. Frequency alone gives at most 3, because more often does not mean differently.

   • ANot once

     sets level 1
     1 · 1%

   • BA few times, to try it

     sets level 2
     0 · 0%

   • CA few times a month, but not regularly

     value 2
     2 · 2%

   • DRegularly, for example every week

     value 3
     3 · 4%

   • EAlmost every day

     value 3
     9 · 11%

   • FEvery day

     value 3
     66 · 81%

   66 of 81 people, or 81%, use AI every day, and 9 almost every day. One person did not use it once in the last three months.

2. Question 2 of 11 · 81 answers

   What do you think about AI?

   For example about chats that write texts and answer questions. This answer does not affect your level.

   What it measures: Attitude to AI. It does not change the level. Answers A and B are marked as sceptical.

   • AI don't trust it, it makes things up

     sceptic, no effect
     0 · 0%

   • BI'm afraid it will take my job, my ideas or my thinking

     sceptic, no effect
     4 · 5%

   • CI'm curious, but careful

     no value
     20 · 25%

   • DIt helps me a lot, I use it

     no value
     57 · 70%

   Nobody picked “I don't trust it, it makes things up”. 57 of 81 people say AI helps them a lot, and one in four is curious but careful. Four people fear AI will take their work, ideas or thinking.

3. Question 3 of 11 · 80 answers

   You have a complex task to do. How do you do it?

   For example a reply to a difficult email, a graphic for an invitation or a piece of code.

   What it measures: How someone does a complex task: without AI, with a hint, pasting the result, refining it with AI, or handing over the whole task.

   • AI do it without AI

     no value
     3 · 4%

   • BI ask AI how to do it, then do it without AI

     value 2
     8 · 10%

   • CAI does it, and I paste the result where it's needed

     value 3
     5 · 6%

   • DI ask AI for changes until the result is what I want"shorter", "other colours", "this doesn't work, fix it"

     value 4
     49 · 61%

   • EI give AI the task, say "do it", and it does the restit gathers what it needs, does it and finishes, and I don't even approve

     value 5
     12 · 15%

   • FI don't do things like that

     no value
     3 · 4%

   Most often the room refines the result together with AI: 49 of 80 people (61%). 12 people hand the whole task to AI, and 3 do it without AI.

4. Question 4 of 11 · 79 answers

   You ask AI for something and the result doesn't fit. What do you do?

   For example the email came out too long, the graphic has the wrong colours, the code doesn't work.

   What it measures: Whether someone fixes the result together with AI, or gives up on AI, or fixes it by hand.

   • AI give up on AI and do it without it

     value 2
     2 · 3%

   • BI take AI's result and fix it by hand

     value 3
     4 · 5%

   • CI tell AI what to change and get a new version"shorter", "no emoji", "use the formal form"

     value 4
     28 · 35%

   • DI give AI more information, for example a file or a linkor tell it to check something else

     value 4
     45 · 57%

   • EI don't use AI

     no value
     0 · 0%

   Only 4 of 79 people fix the result by hand. More people give AI extra material such as a file or a link (45 people, 57%) than only tell it what to change (28 people, 35%).

5. Question 5 of 11 · 79 answers

   You need to prepare a letter, an offer or a presentation. How do you do it?

   For example an offer for a client, slides for a meeting, a poster for an event.

   What it measures: Whether AI produces a finished document or only suggests content.

   • AI do it without AI

     no value
     1 · 1%

   • BAI suggests ideas, I do the rest without it

     value 2
     8 · 10%

   • CAI writes the content, and I paste it into the file

     value 3
     3 · 4%

   • DI refine it with AI until it fits"simpler", "add a chart"

     value 4
     25 · 32%

   • EAI makes the finished file itself, and I check the resultfor example 10 slides from my notes

     value 5
     41 · 52%

   • FI don't do things like that

     no value
     1 · 1%

   Over half the room, 41 of 79 people, get a finished file from AI and only check the result. Another 25 people refine the text with AI until it fits.

6. Question 6 of 11 · 79 answers

   You're planning something bigger. How do you do it?

   For example a trip, a move, a renovation or a party.

   What it measures: How far AI reaches in a private matter, from ideas to booking and paying.

   • AI plan without AI

     no value
     1 · 1%

   • BAI suggests ideas, and I choose without it

     value 2
     14 · 18%

   • CI take the plan from AI

     value 3
     1 · 1%

   • DI ask AI follow-up questions and fix the plan until it fits

     value 4
     32 · 41%

   • EAI compares offers and puts the plan together itselfand I book and pay without it

     value 5
     31 · 39%

   • FAI books and pays itselffor example tickets, a place to stay or a renovation crew

     value 6
     0 · 0%

   In private matters the room splits almost evenly. 32 people refine the plan with AI, and 31 let it compare offers and build the plan on its own. Nobody lets AI book and pay.

7. Question 7 of 11 · 79 answers

   What does AI have access to for you? Pick the answer that gives it the most access or abilities.

   For example a plain chat knows only what you type in. You can connect it to your email, drive or calendar.

   What it measures: What AI has access to: only what someone pastes in, email and drive, the browser, or payments. The only question that says which answer to pick when several fit.

   • AOnly what I give itI type, paste or upload a file

     no value
     31 · 39%

   • BMy email, drive or calendar

     value 4
     15 · 19%

   • CThe browser or the whole computer, it opens pages and clicks buttons itself

     value 5
     33 · 42%

   • DPaying for mefor example with a card saved in the browser

     value 6
     0 · 0%

   • EI don't know

     no value
     0 · 0%

   Here too the room splits almost evenly. 33 of 79 people give AI access to the browser or the whole computer, and 31 only to what they give it. Nobody gives it access to payments, the same as in question 6.

8. Question 8 of 11 · 79 answers

   Does AI regularly do something for you, not just answer in a chat?

   For example every morning it lists tasks from your email, every week it makes a summary, every month it prepares a report.

   What it measures: Whether AI regularly works for someone outside the chat.

   • ANoAI only answers me in a chat, or I don't use it

     no value
     17 · 22%

   • BIt happened once or twice

     value 5
     15 · 19%

   • CYes, every week or more often

     value 5
     47 · 59%

   For 47 of 79 people (59%), AI does more than answer in a chat every week or more often. 17 people answered “No”.

9. Question 9 of 11 · 78 answers

   Does AI start work for you on its own, before you ask for anything?

   For example an email comes in and a reply is already waiting for you. A finished report waits every morning. Tasks appear on their own after a meeting.

   Does not count: Keyboard word suggestions, short reply suggestions under an email and notification summaries do not count.

   What it measures: Whether AI starts work on its own, without being asked. Answer B or C opens questions 10 and 11 and is required for levels 6 and 7.

   • ANo, I always start

     no value
     42 · 54%

   • BYes, it prepares, and I check every item

     value 6
     12 · 15%

   • CYes, and some things it finishes itself, without my approval

     value 7
     23 · 29%

   • DI don't use AI for work

     no value
     1 · 1%

   Levels 6 and 7 depend on this question. For 42 of 78 people (54%), a person always starts the work. For 35 people AI starts on its own, and for 23 of them it also finishes some things without their consent.

10. Question 10 of 11 · 78 answers

    How many things did AI finish completely in the last month, without you checking?

    For example a reply reached a client, a report reached the team, an invoice reached accounting, a fix reached the website. It doesn't have to be an email.

    What it measures: How many things AI finished without being checked in the last month. It counts only for people whose AI starts work on its own.

    • ANone, I check every one before it reaches people

      value 6
      36 · 46%

    • BOne or a few

      value 7
      18 · 23%

    • CA dozen or more

      value 7
      23 · 29%

    • DI don't use AI for work

      no value
      1 · 1%

    23 of 78 people (29%) say AI finished a dozen or more matters in the last month without their check. The whole room answered this question, but it counts towards the level only for the 35 people whose AI starts work on its own (question 9). So “None” was also picked by people who have nothing to check.

11. Question 11 of 11 · 78 answers

    How do you know AI did well what you don't check?

    For example you read the overnight report in the morning, an automated check tests a fix, a second AI reads a text before it goes out.

    What it measures: How AI's work is checked when nobody looks at it before it goes out. It counts only for people whose AI starts work on its own.

    • AI check before sending, every item

      value 6
      31 · 40%

    • BI check after, for example in a daily summary

      value 7
      8 · 10%

    • CAn automated check or a second AI does it for me

      value 7
      31 · 40%

    • DI don't check

      value 7
      3 · 4%

    • EI don't do things like that

      no value
      5 · 6%

    Two answers drew 31 of 78 people each. Some check every item before sending, others leave the check to an automated check or a second AI. Three people do not check at all. Like question 10, this question counts towards the level only for the 35 people whose AI starts work on its own.

Take a sheet of paper and go through the questions above. Write down the value of each answer. Then:

1. If you pick "Not once" in question 1, you are on level 1. If you pick "A few times, to try it", you are on level 2. Counting stops there.
2. Questions 10 and 11 count only if you pick B or C in question 9.
3. Your level is the highest number from 7 down that meets three conditions at once:
   • You have at least two answers worth that much or more. One answer alone never makes a level.
   • From level 4 up, you also have at least two answers from questions 3–8, about your own work, worth at least one less than that level. For levels 6 and 7 it is enough that they are worth 5 or more. This is a separate condition: these answers do not have to be the ones from the point above.
   • From level 6 up, at least one answer from questions 9–11 is worth 6 or more: AI starts or finishes work on its own.
4. If no number from 3 up meets them, you are on level 2.

In a team, count how many people stand on each level. An average says nothing here, because levels are different ways of working, not points.

The scale measures what people say they do. It does not measure competence, talent or an employee's worth, so it is not a tool for screening candidates. It helps in four situations:

1. Before you buy AI training, check where the team stands. Someone on level 2 needs a first habit. Someone on level 5 needs access to tools and rules about what AI may do on its own.
2. In a job interview, ask about practice, the way the quiz does. "Tell me about the last task AI did for you" shows how someone works; "Do you know ChatGPT?" only shows they know the name. Questions 3, 5 and 9 work in an interview almost unchanged.
3. Collect answers anonymously. When a rating depends on the result, it is easier to answer the expected way than the way you actually work.
4. Run the quiz again after a quarter and compare the distribution, not an average. Progress shows as fewer people on the lower levels and more on the higher ones.

Each level has one thing for tomorrow in the table. To start, it is enough that everyone does theirs.

The scale shows where people get stuck. Each step up needs something different:

• From 3 to 4: talking to AI instead of pasting. The habit of telling AI what to fix and who the text is for is enough.
• From 4 to 5: whole tasks instead of edits. Level 5 is confirmed by answers in which AI does several steps on its own: a finished file, comparing offers, work outside the chat, or a browser in which AI clicks by itself (questions 3, 5, 6, 7 and 8). Email connected to AI alone is still level 4.
• From 5 to 6: AI starts on its own. This is a technical step: a schedule, a reaction to a new email, a report every morning (question 9).
• From 6 to 7: oversight instead of checking every item. People stop checking everything before it goes out once they have a daily summary, an automated check or a second AI that reads the result (question 11).

Access alone does not make a level. AI that can pay with a card (question 7, answer D) gives at most level 5 until it starts or finishes work on its own.

The counting rules are in "Count your own level". After each question the quiz counted the level only from the answers so far, and later questions only added evidence. That is why nobody moved down the stairs.

• The room selected itself. These were people at an AI conference, not a representative sample.
• These are declarations. Nobody checked whether someone really works this way. The two-answer condition limits the effect of one inflated answer but does not remove it.
• Each question takes one answer. Someone who does different things differently had to pick one. Only question 7 said which to pick.
• The room saw everyone's answers live and could change their own before I showed the stairs. Some people may have followed the majority.
• The quiz was anonymous, so I don't know jobs, ages or companies and don't split the results into groups. The same person may have joined twice, for example from a second phone.
• With 82 people, the share of one answer can be off by up to about 11 percentage points, and levels 6 and 7 may hold single people. A difference of a few points is not a trend.
• People who joined late did not answer the earlier questions, so their level may have come out lower than it really is.

• CBOS, research report no. 64/2026 "Korzystanie ze sztucznej inteligencji" (Use of artificial intelligence), fieldwork May 2026, N = 1041
• The questions, answer values and counting rules come from the quiz code as used on 6 October 2026 (043d708).