Decisions
AI makes explicit choices that could previously remain implicit. Every missed choice still produces an effect, and it is easier to manage when that is recognised early.
AI makes explicit choices that could previously remain implicit. What used to be a question of degree becomes a policy, a boundary, or an effect that still shows up when nobody makes the call.
Many of these decisions already existed as margins of choice, and AI turns them into forks. How far to adapt to the user was a matter of degree; today it is a policy to be written. Whether a customer can get a quote without going through us was a matter of channel; today it is a choice of identity.
They all share one feature: even the absence of a choice produces an effect. If nobody decides, the decision gets made anyway — by development case by case, by the vendor’s out-of-the-box settings, by the intermediary that brokers the relationship, or by the client who asks for a feature because it is the trend. Not deciding is a position, not a postponement.
The decisions collected here do not belong to a single phenomenon: they are made once and then hold across every chapter. The ones anchored to a specific phenomenon sit inside that phenomenon, in the “The decision” block.
Each card states the question, the effect that still appears if nobody answers it, and what either branch costs. Who ought to decide is a different question: it belongs to professions rather than to the decision, and it is answered in Who decides.
The decision that generates the others
AI can do four things to a business, and for each one the same question applies — how much do we want AI to do this job? The four answers are not independent of one another, but neither are they the same answer repeated four times. Each has a different constraint, and each has a different default if you do not make the call.
Attract
Citation, recommendation, search, discovery. How much do we want AI to bring people towards us?
The constraint the literature documents is that position is no longer governable. Only 38% of citations in AI Overviews come from the organic top 10, against 76% in the preceding period; when Google changed its default model in January 2026, roughly 42% of the cited domains were replaced; and the same question asked a hundred times does not produce the same list twice — “any tool selling you a ranking position inside AI is selling an artefact”.
So this choice does not regulate ranking, because ranking cannot be regulated. It regulates how much to invest in a channel that is volatile by construction.
→ Falls on Intent-based interaction.
Interpret
Brand, content, positioning, reputation. How much do we accept that AI says who we are?
This is the choice with the least control, and it is worth writing without softening: only 1% of the sources cited by LLMs comes from brand-owned sites, the rest is earned. Brand representation is built where you do not publish.
It has a minimum position that cannot be zeroed: you can choose not to tend it, you cannot choose not to be interpreted.
→ Falls on Representation forms where you don’t publish and The brand reduced to comparable attributes; and, on the search side, on Serendipity declines — both decisions — and Intent-based interaction.
Assist
Chat, personalisation, progressive refinement. How much do we want AI to stand between us and the user during use?
This is the choice where the default comes from the client rather than from the technology — “in most cases it is the trend driving it, not a real need” — and it is the only one with a documented upper limit: beyond the optimum point of adaptivity users report “more negative perceptions”. The limit exists, but only if you set it: no tool flags it for you.
→ Falls on Zero learning curve, Chat as the universal default interface, From form to progressive refinement, Memory and context adaptation.
Act
Agentic interaction, quote, booking, purchase, service. How much do we want AI to do things on our behalf and on the user’s behalf?
This is the choice where the price is most concrete. Opening up exposes you to prompt injection and to commoditisation — “the agent compares attributes, not brands” — closing down means leaving the channel where the comparison will happen anyway.
And it is the only one with infrastructure already in the field: Universal Commerce Protocol presented at NRF in January 2026, agentic checkout in AI Mode. Meanwhile OpenAI withdrew Instant Checkout on 24 March 2026, five months after launch, and only 17% of shoppers say they are comfortable completing a purchase with AI. Direction set, adoption uncertain.
→ Falls on The user expects the system to act and Making a traditional product usable by an agent.
The four positions together are a profile
And different profiles are legitimate. Four examples, to show there is no correct configuration:
- a luxury brand can keep attract high and act at zero: being found yes, being transacted by an intermediary no;
- an editorial retailer can keep interpret high and accept a loss of precision, because discovery is its lever;
- a media company may not be able to close interpret down, because it is its product;
- a high-volume, low-margin service can open act completely and accept commoditisation, because price is already its lever.
The profile has a price, and the cost items are the other cards in this chapter: they only say that there is a cost, and that choosing not to choose means paying it without knowing.
The other seven decisions
Where we are, and through whom
How much we let it do, and who answers
What we keep counting
One stated limit, valid for every card: on brand identity and on the long horizon the literature offers no metrics. Where the measure does not exist, the cards say so instead of simulating it.