Samuel writes on judging AI without becoming technical on samuelpouyt.com

Follow-up to How AI Assistants will impact businesses and consumers (Boye & Company blog, October 2018).

Synthetic users and AI personas came up as something In October 2018, fresh from hosting the Emerging Tech track at the Boye conference, I wrote a post on this blog about how AI assistants would change business. Assistants, at the time, could set a timer and mispronounce your calendar. Writing about them reshaping commerce felt slightly reckless.

Eight years is long enough. Most predictions die politely - nobody checks them, including their authors. I'd rather grade mine in public, because the pattern of what held and what didn't turns out to be more useful than any of the individual claims.

What held

No ads inside the assistant.
In 2018 I wrote that there would be "no real estate for ads" in an assistant interface, and that platforms could not betray user trust by favoring paid recommendations. In 2026 this is the live strategic debate of the industry: some AI companies are experimenting with advertising, others have publicly staked their brand on rejecting it, and the argument on both sides is the one from 2018: the moment a user wonders whether an answer serves them or an advertiser, the product breaks. The least often expressed allegation in the post was this one. It aged most sensually.

And in 2026 the second shoe is dropping, one I didn't call: it's not only that the assistant has no room for ads, it's that the ads made for humans are increasingly shown to no one. An agent fetching a page on your behalf does not see the banner, cannot be retargeted, and will never click; yet its visit can still be counted as an impression on somebody's invoice. An advertising economy priced on human eyeballs is discovering that a growing share of its audience isn't human.

Trust as the whole game.
I argued that trust builds in a virtuous circle of successful experiences, and that each failure sends the assistant back to the toy drawer. Siri was my example. That is precisely the story of the last eight years: the assistants that won are the ones that crossed the trust threshold, and every public failure, an invented citation, a confident wrong answer, costs a user more than ten good answers earn back. Trust turned out to be the product. Everything else is packaging.

A handful of platforms as the new middlemen.
I predicted market consolidation into a few dominant AI platforms positioned between brands and consumers, like the credit card networks before them. Consolidated it has, though not around the names I would have bet on, which is its own lesson filed below.

A new profession: "AI platform optimization."
In 2018 I wrote that a job would emerge paralleling the SEO specialist: people whose work is making a brand visible to AI platforms rather than to search crawlers. That industry now exists, has conferences, and has an acronym war (AEO, GEO, LLMO). I will confess the prediction became personal: I have spent part of this year doing exactly this work on my own website, marking up pages so that assistants can cite them. When your old blog post describes your current weekend, the forecast held.

Purchases delegated to the assistant.
Routine buying handed to an agent—the toothpaste, the groceries, the travel—was the post's central scenario, along with the observation that consumer loyalty is mostly laziness and that an agent that compares without fatigue would expose it. “Agentic commerce” is now on every retail board agenda, and brands are discovering that an assistant doing the choosing does not care about their retail theater. The 2018 framing needed no update, only patience.

The claim beneath the claims.
Read together, eight years later, these five turn out to be one prediction wearing five coats: the assistant is becoming the interface. Not a feature, not a channel, the layer that sits between a person and everything digital, with the power of life and death over what gets through. When your product is a blog post, the assistant decides whether it is cited. When your product is toothpaste, the assistant decides whether it is in the basket. That is why a profession now exists to negotiate with it. In 1998 you fought for a shelf; in 2010 you fought for a search ranking; in 2026 you fight for a place in the answer. The assistants are quietly becoming the new web.

A word on words before the misses. In 2018 I wrote “assistant”; the industry now says “agents.” I'll keep both, deliberately. The assistant is the one you talk to, the trusted interface everyone imagining Jarvis dreams of. Agents are the errands it sends into the world. The assistant holds the relationship; the agents do the walking. The distinction matters because trust concentrates in the first while traffic comes from the second.

What missed

The channel.
The post was soaked in voice. I cited the era's numbers with a straight face: smart speakers in 55% of US households by 2022, voice shopping growing from $2 billion to $40 billion by 2022, half of all search queries by voice by 2020. None of that happened. Smart speakers plateaued into kitchen timers and radios. Voice shopping never found its $40 billion. The capable assistant, when it finally arrived, came through a text box—reading, writing, and reasoning—not through a speaker on the counter.

The timing.
I assumed the assistants of 2018 would improve gradually into the assistants I was describing. Instead, nothing much happened for four years, and then everything happened at once, through a technology shift almost nobody had scheduled. The forces were pointed the right way the whole time; the vehicle was one I didn't see coming.

The incumbents.
“94% of all devices in use are from Google and Amazon,” I noted, and treated the device duopoly as the seed of the platform oligopoly. The actual winners included companies that didn't exist or didn't matter in the assistant conversation of 2018. Distribution mattered less than capability. I would not have bet that way, and I should say so.

The pattern

Put the two lists side by side, and the pattern is hard to miss. Every prediction about incentives and human behavior held: trust, laziness, middlemen, the incompatibility of ads with advice, and delegation of the boring. Every prediction carrying a number, a form factor, or a date missed the households, the billions, the speakers, and the timeline.

I don't think I was unusually wise about the first list or unusually careless about the second. I think that is simply the shape of foresight. Forces are predictable because people are; forms are not, because forms depend on breakthroughs that don't announce themselves. The corollary for anyone making AI decisions in 2026: when a strategy document tells you what will happen, weigh it; when it tells you when and in what packaging, treat it as fiction with a spreadsheet.

And one more corollary, which is really the reason this post exists. The only reason I can grade these claims is that they were written down, in public, with a date on them. Every AI strategy adopted this year contains predictions about vendors, about capabilities, about what your customers will tolerate. Most of them will never be graded, because they were never written down as claims. Write them down. Date them. Put a reminder in the calendar for 2028. Being wrong in specific, checkable ways is how judgment compounds; being vaguely right forever is how it doesn't.

What I'm writing down for 2034

So let me eat my own cooking. Dated August 2026, offered for public grading:

1. The interposition completes slowly, then suddenly.
Today the assistant is not yet fully between you and your mail or between you and your toothpaste. By 2030 it will be the first reader of most professional inboxes, triaging, summarizing, and drafting the boring half, and brands will learn to write email for an audience of agents.

2. The inversion completes and takes the ad-funded web with it.
This one is no longer a prediction so much as a trajectory: Cloudflare reported this summer that automated traffic has passed human traffic, a crossing their own CEO hadn't expected before next year. Brands are watching their site traffic evaporate while their products are discussed more than ever inside answers they never see. So the claim for 2034: the page view dies as a metric; the website splits into theater for people and structured truth for machines; display advertising, priced on human eyeballs, shrinks toward the few places humans still visit in person; and the fight moves to the answer itself, which is where the 2018 no-ads claim faces its final test. “Does the assistant cite you?” becomes a board question the way “Are you on Google?” once was.

3. Voice comes back through the side door.
I buried voice in the miss column, but watch where it went: engineers now describe a feature out loud to a coding agent and watch software appear. Voice didn't die with the smart speaker; it was waiting for something worth talking to. Dictating intent to an assistant that can act will spread from engineers to the rest of white-collar work; the countertop cylinder was the wrong body for the right idea.

4. The last mile stays human longer than capability allows.
The assistant that sends the mail unread, moves the money, and reorders without asking. That lag is not about intelligence; it is liability and trust. My 2018 assistant that “resigns on our behalf” will still not be routine in 2030, and the gap between what agents can do and what we let them do will be the most interesting line in business.

5. Rules for agents become a governance artifact.
Companies will keep written rules, automated checks, and named human approval points for their agents the way they keep accounting controls, and a company without them will look, to its board and its insurers, like a building without locks.

By the pattern above, the forces in these five should hold, and at least one form or date will be wrong. My money says the dates are too conservative rather than too bold, which is itself a prediction, so it's written down too.

I'll see you back here in 2034 for the grading.