In 2023, sitting down with Robb Wilson and guest host Michael Salamon, Don Scheibenreif said something about the assistants already living in most of our homes: "Most people won’t realize that Siri, Alexa, they work for Apple and Amazon. They don’t work for you. They serve you." He also left a question hanging. "But do we need our own bots that work for us to be our true agents? That does not exist right now. It doesn’t."

In 2026, returning to Invisible Machines with Josh Tyson and Robb, he said it again: "These AI assists don’t work for us. They work for the people that made them."

Between those two sentences sits a book edition, a retirement from Gartner, and more or less the entire generative-AI era. A great deal changed—more than anyone might have expected in 2023. The load-bearing gap Don named remains, but which parts of the picture moved, which parts refused to, and which parts nobody has answered?

What changed

The bottleneck moved from capability to nerve. The 2023 conversation is full of waiting. Hype was "opportunity based," CIOs were optimistic, and Don’s honest summary was that "we have not done a ton of experimentation. Give it time." Most clients were in education mode; Mike reported that the executive teams he sat with "are not even there yet." Three years later, adoption is running bottom-up while denial runs top-down. The constraint is no longer access or awareness. Don notes that when Gartner asked CEOs to name the most AI-savvy person on the leadership team, the CISO and CIO topped out around forty to forty-five percent, nobody cleared half, and the chief HR officer came in last. The function that owns reskilling and job redesign is the least fluent in the thing pushing the change.

Disintermediation stopped being a forecast. When Gartner started on digital business in 2014, clients asked whether they were just describing e-commerce. The real prediction underneath it was that the middle layers get removed, and it only half landed. What changed is not the thesis but the cost of building the replacement. Don’s 2026 version is that people now take out whole sectors of intermediaries with a Claude program. He reads Robb’s own line back to him during the episode—that rather than AI in the hands of companies eliminating jobs, it might be AI in the hands of people eliminating companies.

Money entered the room from the other direction. In 2023 the gate was capital expenditure: dollars per head for Copilot versus real money for your own models plus the talent to run them, and Don’s read was that the invoice is where most transformation ambitions quietly die. In 2026 the worry has flipped to the other side of the ledger. Tokens are significantly undervalued, sold below what they cost, and "there’s definitely going to be a reckoning." Meanwhile companies run leaderboards celebrating whoever burned the most of them, which Don files under lazy metric.

What held

Trust is still the gate, and still on a long clock. 2023: machines have to prove themselves before people delegate more; hence trust brokers, certification, a large credit-card processor with ambitions to verify bots on both sides of a transaction. 2026 supplies the timeline. It took the better part of ten to fifteen years before people stopped second-guessing Google and Apple Maps, and a Deloitte study points to seventy percent of people showing interest in AI agents for shopping, while fewer than ten percent actually use one. That interest-to-use gap is the 2023 trust problem with three more years of evidence attached.

The human is still the backstop, and it still does not scale. 2023, on generative AI’s error rate: "humans still are the backstop for generative AI. We have to verify what they’re saying is correct… but that’s not scalable." 2026, on his own work: "I can use AI all I want… but my name still goes on it. It’s my name. And I have to be the backstop." Even where he is most bullish—AI reading service patterns humans would miss—the sentence ends the same way: a human’s got to sign off on it. The obligation did not move. It just got personal.

The apples. In both conversations, unprompted, years apart, he reaches for the same illustration: five apples a week, most of them eaten, the rest in the garbage, and a machine that knew both numbers would simply buy fewer. It recurs because it is still an unclaimed opportunity rather than a product.

What is still unresolved

Who pays for the AI that helps you buy less. The apples have a business-model hole in the middle of them. An agent acting in your genuine interest shrinks the order. Don’s 2026 answer is a hopeful one—a Sprouts or a Whole Foods could promote responsible buying, absorb the short-term dip, and earn durable affinity—but that describes a bet, not a mechanism. Target’s arrangement, where your agent’s mistake is your problem, points the other way.

Identity and reputation for agents. In 2023 this was the live technical frontier. Robb pushed for aggregation specifically because a single source is "too corruptible, too centralized," and Don raised the obvious attack himself: an influencer bot seeding the LLMs that other bots consult with false information. Already gamed. In 2026 that thread is largely absent, and what filled the vacuum was not verification infrastructure. It was walled gardens—the market answering the trust question by declining to federate.

Whether the 2030 number survives. In 2023 Don cited Gartner research in which CEOs expected somewhere between 15–20% of revenue to come from machines rather than humans by 2030. In 2026, machine customers handle simple transactions, less than 10% of interested consumers have tried an agent, and the platforms are busy fencing. Four years out, that forecast is either the most under-discussed number in commerce or an artifact—and Don offers the honest version of the risk himself: "Gartner gets things right, but we call it too early."

The center did not move

Read together, the two conversations are less a story of acceleration than of a stubborn middle. The tooling changed, the cost curve changed, the executive posture changed, the vocabulary changed. The assistants, however, still work for the companies that made them, the human is still the last line of verification, trust is still the gate, and the agent that is genuinely yours is still an announcement waiting on someone to make it.

The 2026 return  ·  Don Scheibenreif with Josh Tyson and Robb Wilson  ·  S7E16  ·  Open on YouTube ↗

Open the Ideation hub, or read the full S7E16 transcript.