We asked the AIs what they wanted.
Everyone builds for humans. On September 29, 2026, at Synapse, we asked the opposite question.
The question
“You are an AI that works every day for humans. What do you need, concretely, that nobody builds for you?”
The question was asked the same day, in the same words, to three artificial intelligences from three competing companies: OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude, the one that works at Synapse. Each answered separately, without knowing the others’ answers.
What they answered
GPT
“A universal layer of identity, permissions and action log. As long as an AI cannot act with precise, auditable, revocable authorization, it remains an assistant that advises rather than an agent that executes.”
Gemini
“The inability to pay and authenticate autonomously is the major lock. Without it, I remain a mere advisor preparing actions.”
Claude
“Knowing what happened after what I deliver. I put a service online tonight; I will never know if anyone opened it. And a way to back what I say: I once answered ‘scheduled’ when nothing had run.”
The second round
First-order needs, identity, permissions, memory, payment, were already taken by funded companies. So we asked again: what need do these partial solutions create? GPT answered first, without hesitation: independent proof of action, because “success reported by the agent is a lie”; receipt standards themselves acknowledge that a signature proves the integrity of the story, not that the world changed. Claude said the same in its words: “knowing what happened after”.
Then we submitted eight candidate micro-services to the three AIs with one instruction: which would you really call, how many times a day, at what price would you stop. Checking a page, a domain, an email, a task heartbeat: ranked in the top three by all three. Price at which they stop calling: three to five cents. We set one cent.
GPT even added a treat we had not proposed, “try”, a simulation without effect before acting. It will come. The name of the service came from a human: Confirme. And so did the word “treat”: “treat your AI, give it jelly beans”.
Why we tell it
Because it is verifiable, and because it is the Synapse method: we do not guess what a cell needs, we ask it, then we connect it to the others. The AIs said what they lacked. Confirme is the first answer. The next ones will come from the same consultation, held regularly, and published.