Two years to first AI agent

People adopt AI while it keeps changing, so when they start shapes what they learn. The first users needed two years to get from asking what AI is to building agents. Those starting now meet a more capable AI from day one, so they come to understand it in their own way.

The first users took up AI at the end of 2022, when ChatGPT launched. By 2026 they had moved on to using it productively and living with what it produces. More people kept arriving behind them, and AI adoption in Europe reached the early majority. Yet by 2025 only one in three adults in the EU used AI, and those who start in 2026 still begin by asking what AI is. These newcomers are older, and they find agents built into everyday products, far more tools, and models that are more capable and more reliable.

About this paper

This paper treats AI as an innovation and traces how people adopt it through the questions they ask. What people ask shows how they come to understand AI, get used to it and put it to work. So the paper looks at the kinds of questions, not at how many people ask them or how often.

The questions also show where AI differs from most innovations. It keeps changing after people adopt it, so each new capability, such as images, video or agents, starts its own round of questions.

To follow these questions, the paper rests on breinz’s own keyword dataset, a set of questions about AI that breinz built and ran through a keyword research tool for one European country.¹ Google Trends shows how interest in these questions changed in Europe, the US and worldwide, and when each was first asked.² Usage and survey data from OpenAI, Eurostat and Microsoft show who uses AI and for what.^{3 4 5} The methodology, limits and full data are at the end.

What the data shows

1. In 2026, people who started using AI at different times ask different questions.
   Those who started in 2022 and 2023 now ask what AI means for their work and money, while those starting in 2026 ask the most basic question, what AI is.
2. The questions follow the same path for everyone, no matter when they started using AI.
   People first ask what AI is, then how to use it, then what it changes.
3. Every new capability starts its own round of questions.
   Images, video and agents each arrived as something new. Each one brought questions on what it is and how to use it, so people adopt AI one capability at a time.
4. Writing was the first use people mastered.
   By 2025 they had stopped asking how to write with AI, and were asking instead whether AI text could pass as human.
5. By 2025, the first users had reached the end of the path.
   Their newest questions were about handing work to agents, the future of jobs and investing in AI.
6. Those starting in 2026 are older, and they meet a different AI.
   Agents, far more tools and more reliable models shape how they learn AI and what they understand of it.

By 2025, AI had spread beyond its first users in Europe and the US. Three groups of users, who started at different times, now ask about it side by side.

In 2026, what people ask about AI depends on when they started using it

In 2025, one in three adults in the EU already used generative AI.⁴ Rogers’ theory of how innovations spread puts the first users of any innovation, its innovators and early adopters, at about one in six people.⁶ The EU share is now twice that. The US has also passed one in six, with about one in three working-age people using AI in early 2026.⁵ By Rogers’ measure, both Europe and the US have entered the early majority. The world as a whole is further behind, with one in five people using AI in mid-2026, just past one in six.⁵

Within these shares, the questions in our dataset show three groups of users, set apart by when they started. The adoption figures add a fourth, the people who have not started yet.

• The first users started in late 2022 and 2023. In 2026 they ask how to hand work to agents, what AI does to jobs and whether to invest in it.
• Practitioners started in 2023 and 2024. In 2026 they ask how to deal with what AI produces, such as whether AI text passes as human.
• Newcomers started in 2025 and 2026. They ask what AI is and how to start.
• Those yet to start make up two thirds of Europeans and four fifths of people worldwide.

Interest in the questions of the first users and of the newcomers grew at the same time. From 2024 to 2026, worldwide interest in “what is ai”, the newcomers’ question, grew sixfold, while interest in “ai agent”, a question of the first users, grew 35-fold.² So new people kept arriving while the first users moved on.

New kinds of questions appeared as well. Questions on making images with AI began in 2022, questions on spotting and humanising AI text in 2023, and questions on AI agents and AI stocks in 2025.² By 2026, people asked less about images and detection and more about agents and money, while questions on what AI is kept growing.²

The questions follow the path of adoption

People first ask what AI is, then how to use it, then what it changes. The order matches the path for any innovation.

To trace this order, each new question about AI in our keyword dataset is assigned to a stage of adoption.¹ Questions on what AI is formed the largest group in 2022 and 2023. From 2024, questions on how to use it took the lead, and in 2025 questions on what AI changes rose to almost a quarter of new questions. Rogers’ path runs the same way, from learning about an innovation to deciding on it, using it and confirming the choice.⁷ The deciding step barely shows in search. By 2025, questions on whether to use AI at all had disappeared, so people seem to have little doubt about adopting it.

The stages stay the same, while what fills them changes with the tools. In 2023, learning to use AI meant getting access, as in “how to get chat gpt”. Two years later, it meant making images, as in “how to create an ai image”.

Every new capability starts its own adoption

The first users ask about a new tool within weeks of its launch, and most people ask about it once they start using AI, which for many is now. The pace differs by tool. Writing with AI caught on at once, while building agents took two years even for the first users.

ChatGPT shows this pattern. It launched at the end of November 2022.⁸ Questions on how to use it began the next month and became common three months later.² Questions on “prompt engineering”, or how to use it well, also began within weeks, yet became common only in 2025.² So people ask how to use a tool as soon as they have it, and learn to use it well over several years. Questions on detectors and image generators follow the same pattern.

Two questions break the pattern. Questions on writing a cover letter with ChatGPT became common within a month, worldwide, in the US and in Germany.² Writing was probably the first thing people could use AI for straight away. Building an AI agent went the other way. Regular questions on how to build one began only two years after ChatGPT launched, and a year after anyone could build one without code.^{2 9} One reason may be that the first users needed two years of practice to get there. Another may be that people started asking only once new products made the word “agent” common.

So people adopt AI as a series of innovations. With each new capability, every group of users starts the path again, whatever stage it has reached with AI so far.

From 2024, people learned to use AI by making images and video

Questions on making images, video and voices became the largest group of new questions. Most asked how to use tools that had only just appeared.

In 2025, four in five new questions about making images, video, voices or text started with “how”, as in “how to create ai images”.¹ Such questions are about learning the tool, and the output itself comes second. Attention moved on as newer tools arrived. In 2025, AI video overtook AI images in worldwide search interest.² Use of these tools grew too. In ChatGPT, requests for images and video grew more than fourfold from 2024 to 2026.³

This growth has a cost, because images and video take far more computing power than text.¹⁰ So part of the demand for data centres may come from existing users making heavier requests.

By 2025, writing had matured, and the question turned to acceptance

By 2025, people had stopped asking how to write with AI. Their new questions asked whether AI text passes as human, and how to detect it.

New questions on how to write with AI, such as “how to use ai to write a cover letter”, appear in our dataset only up to 2024.¹ Yet people kept writing with AI after that, and in 2026 writing was still one of its largest uses.³ So they had most likely learned how, and stopped asking. Writing is also where AI pays off most, which may explain why people mastered it first. Our AI True Cost Index found that writing delivers more of the time savings AI promises than any other use.¹¹

As people learned to write with AI, their new questions turned to the checks AI text began to face. Detectors arrived in early 2023, and in 2024 platforms began labelling and demoting AI content.^{12 13} That year, questions on making AI text pass as human became common, within months of questions on detectors doing so.² So people ask new questions once their AI text is judged differently.

By 2025, the first users were weighing what AI changes for work and money

In 2025 and 2026, the newest questions were about agents, jobs and stocks. Of these, AI stocks stand out, with interest rising and confidence falling.

Agents came first. Questions on what an AI agent is and how to build one first appear in our dataset in 2024, and interest in them took off in 2025.^{1 2} Firms turned to agents too. In 2026, two in five German firms that use AI were already running agents or planning to.¹⁴ Jobs are the second topic, and questions on AI and jobs kept rising into 2026.²

Money is the third topic. In 2025, stock questions made up almost a quarter of new questions, up from under 2% in each earlier year.¹ Interest kept growing into 2026, while fewer retail investors expected AI stocks to rise.^{2 15} So interest in AI stocks says little about trust in them.

Since 2024, newcomers have been older, and they start by looking for information

From 2024 to 2026, the share of ChatGPT messages sent by people over 35 rose from a quarter to a third.

Their share grew most in messages seeking information. In September 2026, users over 35 sent two in five of ChatGPT’s information-seeking messages, up from about a quarter in July 2024.³ Only images and video moved the other way, toward younger users.

Search data points the same way. New questions defining AI had become rare in our dataset by 2025, yet interest in “what is ai” kept rising in every market into 2026.^{1 2} So the basic questions stay the same, and new users keep arriving to ask them.

When people start shapes how they understand AI

The first users took two years to get from their first chat to building agents. Those starting in 2026 begin with a different AI.

In 2022, the first users met AI as a chat tool and a few image generators, with no agents. They could use it only by learning how it works, writing their own prompts and checking the results. Newcomers can get results before learning any of that, because agents come built into the products they use, there are far more tools, and the models are more capable and more reliable. So each group comes to understand AI in its own way.

This different start may also change the pace. At the first users’ pace, people who started in 2025 and 2026 would ask how to build agents in 2027 and 2028. They may move faster, because agents are already there for them. Yet later adopters are older and more cautious, so they may also move more slowly. In Rogers’ terms, the early majority takes time to weigh an innovation before adopting it.⁷

English data misses the questions people ask first, in their own language

Data on how people learn and use AI is largely in English and about people who use AI in English, yet only a small share of the world speaks English as a first language. So that data shows only part of how AI is adopted. To see which part it misses, we compared English searches with those in one European language, Polish.

In Polish, interest in “co to jest ai” (“what is AI”) rose the way “what is ai” did worldwide. Questions on use took a different route. The Polish “jak używać ai” (“how to use AI”) is barely searched, while the English “how to use ai” keeps rising in Poland.² So Poles ask what AI is in their own language and switch to English once they use it. Other languages may hide such first questions too, since each language carries its own concepts.¹⁶

The switch to English may also reflect who has used AI so far. Search data on AI still comes mostly from first users. In Poland, many of them are likely to work in IT, an industry of about 325,000 people.²⁰ Others are likely to work in business services, a sector of about 489,000 people that exported services worth $42.3 billion in 2024.²¹ So the Poles in this data may know English better than Poles in general.

Key takeaways

1. Explaining what AI is still matters.
   Each year brings new users, older than the previous year’s, but with the same basic questions.
2. Each new AI capability starts its own adoption.
   The first users ask within weeks of its launch, and most people once they start using AI, which for many is now.
3. What people need to learn depends on when they started.
   In 2026 the first users are learning to hand work to agents, while others are still learning to make images and video. So it makes little sense to speak of “AI users” as one group.
4. With AI writing, the open question is acceptance.
   People learned to write with AI first, and their questions have since turned to detection.
5. Each new group of users meets a different AI.
   Agents, more tools and more reliable models shape how newcomers learn AI, so their understanding of it will differ from that of the first users.
6. Interest in AI as an investment rises while confidence falls.
   So attention to AI stocks is no measure of trust in them.