Observations
AI changes behaviour before it changes products. These pages trace the shifts that follow in search, interfaces, delegation, and the conditions that hold across them.
AI changes user behaviour before it changes products. People who use generative tools every day don’t go back: they delegate search, expect interfaces that adapt, take it for granted that the system will act on their behalf. These behaviours don’t stay inside the chat — they carry over to any digital product, including ones that have nothing to do with AI. That’s where the effect is most visible, because those products did nothing to provoke it.
Generative AI enters through three routes, at very different levels of maturity
Generative AI enters people’s lives in three distinct ways, and each way changes a different part of the digital experience. One starts in search, one in the chat, and one in delegation to an agent.
1. AI-mediated search — Search becomes conversation
AI Overviews and AI Mode reach billions of people inside a path they already knew. Nobody has to adopt anything: it changes the behaviour of people who were already searching on Google. It’s also the chapter where AI decides how it represents us, because that happens in the same response.
- Intent-based interaction: from intention to result
- The customer journey shortens
- Serendipity declines
- Representation forms where you don’t publish
- The brand reduced to comparable attributes
- When AI gets the brand wrong
- The conversations you cannot observe
2. Conversational chat — The conversational interface
Daily use of ChatGPT, Gemini, and Claude makes the chat itself feel like the new default interface. That shift changes what users expect from any product that asks them to interact step by step.
- The new expectation of zero learning curve
- Chat as the universal default interface
- The user’s language replaces the system’s taxonomy
- From form to progressive refinement
- Memory and context adaptation
- Accessibility and hybrid interfaces
3. Agentic AI — Agentic AI: from showing to doing
Some systems now act on behalf of the user rather than only showing options. That raises a different question: how far can a product go when it is asked to do the work itself?
The order isn’t arbitrary, and it’s the most useful thing to keep in mind: the depth of adoption tells you how urgent the phenomenon is. AI-mediated search is already a settled fact and needs addressing now. Agentic delegation is a bet on the next cycle, and you address it by preparing, not by chasing it.
Cross-cutting themes — Cross-Cutting Themes
Some conditions shape every route: trust, regulation, and social presence all affect how these systems are used and judged.
Every phenomenon forces a decision, and even the absence of a choice produces an effect
None of these phenomena stops at an observation: each one puts the organisation in front of a choice. And if nobody makes it, the choice gets made anyway — by case-by-case development, by the vendor’s default setting, or by the customer who asks for a feature because it’s the trend. Not deciding is a position, not a postponement.
That’s why the decision sits inside the phenomenon that generates it, together with the evidence it rests on and what changes for the professions it touches. The decisions that cut across several phenomena at once are collected in What has to be decided; the reading by role is in What changes for work.
What these phenomena rest on
The material comes from the workshop Shaped by AI — Phase 2: the impact on digital professions; the sources that make up the corpus are in the literature review, and they are the base these pages’ reflections rest on, as far as possible. The pages only make sense if the evidence behind them is visible, so the method matters as much as the theme.
No number goes in without a verifiable source. Where a figure rests on a single vendor study, the page says so — and says who has an interest in that result.
Where estimates diverge, the range is reported, not the most quotable extreme. The decline in organic click-through, for instance, is reported the way the twelve independent analyses actually report it: agreeing on direction, differing by a factor of five on magnitude. A wide range is information; a single well-chosen number is an editorial decision disguised as data.
Where no measure exists, the pages say so instead of simulating one. There is no rigorous, independent, brand-specific measure of how often AI gets brands wrong: solid studies measure news, and the ones about brands come from vendors selling the solution to the problem they measure. That’s why When AI gets the brand wrong doesn’t open with a number but with an asymmetry, which is documented instead.
Two limits remain, stated because working around them would be worse. On brand identity and on the long horizon, the literature offers no metrics, and every decision that follows from it inherits this limit. The professions are a reading of the corpus, not a field survey: if a profession appears rarely, it means the corpus names it rarely, not that the change touches it rarely.
The numbers age, and quite a few of these will. The full method is in methodology; if something here is wrong or has been superseded, the comments at the bottom of every page are the fastest way to flag it.