Your next user isn’t human

An AI agent, in the plainest terms, is software that pursues a goal in steps: it navigates, reads, fills forms, compares options and completes transactions, deciding its own next move as it goes. That is Nielsen Norman Group’s definition, and this spring they published something more interesting than a definition: a piece treating agents as a user group, with needs, blockers and behaviour patterns, the way we would write about any audience.

How does an agent see your screen? Three ways, roughly. Some look at screenshots, which is slow and expensive. Some call APIs and skip the interface entirely. The interesting middle is the accessibility tree: the agent reads your interface the way a screen reader does, as a structured list of named things. Your beautiful screen, to this user, is exactly as good as its semantics.

If this sounds speculative, look at who is preparing for it. Visa has shipped Intelligent Commerce, a product line for letting AI agents pay. Stripe maintains agentic commerce documentation for developers wiring agents into checkouts. When the card network and the payment processor both build for a user type, that user type is not hypothetical. The traffic is small today. So was mobile, right up until it wasn’t, and I sat through that transition too.

Watch one work a checkout and your assumptions get rearranged fast. The agent does not infer. A button that is only an icon has no name in the accessibility tree, so it functionally does not exist. A price that appears on hover is a price that was never published. A confirmation that flashes in a toast and vanishes is state the agent never saw; it will sit there, politely re-checking a page that says nothing, then give up. “Fees may apply” is not a sentence an agent can act on, and unlike a tired human, it will not pretend otherwise.

Here is the part I find genuinely funny: that list is also a list of things that fail stressed humans. The unnamed button, the vanishing state, the fog around fees, these are the same failures I keep writing about from the human side, in the bad-day note and the accessibility one. NN/g’s observation lands the same way: interfaces built accessibly are already legible to agents. The machine user did not invent new requirements. It made the old ones enforceable, because it fails loudly and measurably where humans fail silently and drift away.

When I rebuilt the Lightyear help centre, the entire bet was that a stressed person skims: front-load the answer, name every action, prefer structure to cleverness. Support contacts dropped 15%. That same structure is precisely what a machine parses. I did not know it at the time, but I was designing for a species that had not arrived yet.

One boundary, before the checklist: some walls are load-bearing. You want agents stopped at authentication, at consent, at anything that commits money without a mandate. The waste is not the wall at the vault; it is fighting robots on your pricing page, where the only thing being defended is the fog.

The moment an agent can spend, the design problem stops being legibility and becomes mandate. What may it do without asking? What always escalates? Where does the human sign? These are the same “scary steps” I have always sweated in fintech, except the confirm screen now has two readers: the agent that acts, and the human principal who reviews what was done in their name. A receipt stops being a record and becomes an interface between three parties.

The sensible pattern emerging across the industry is progressive delegation: the agent starts with a narrow allowance, small amounts, reversible actions, and earns scope the way a new employee does. Anthropic’s guide to building effective agents makes the engineering case for the same humility: simple, auditable loops before autonomy. Design has to hold that line on the interface: a visible log of what the agent did while you slept, checkpoints at money moments, an unmissable difference between “the agent suggests” and “the agent has done”.

And there is a quieter consequence for every fintech marketing page. Agents and AI answer engines quote you. If your fee is phrased three different ways on three pages, some machine will confidently serve a customer the wrong one, at the worst moment, citing you as the source. The fix is editorial, not technical: one canonical sentence per fact, price, fee, timeline, kept true. In the onboarding note I argued your KYC flow is a landing page; your fee schedule is now an API, whether you meant it to be or not.