Ambient Agents: AI That Assists Without Being Asked

Every AI tool you use today waits for you to ask. You open the chat, you type the question, you read the answer. That interaction model — the chatbot — has become so universal that it is easy to mistake it for the only one. But the most interesting AI systems being built right now invert it completely: they listen to the context you are already in, decide for themselves whether they have something worth saying, and stay silent the rest of the time.

That category has a name: ambient agents. And after building one — a sales copilot that listens to live video calls and whispers cues only the salesperson can see — we think it is the interaction model that will matter most for real work. Not because it is more impressive than a chatbot, but because of a simple observation: the moments where help is most valuable are exactly the moments where you cannot stop to ask for it.

The cost of asking

Think about what a chatbot actually demands from you. You must notice that you need help, formulate the question, switch context to another window, type, wait, read, and return. In a quiet moment at your desk, that cost is trivial. In the middle of a negotiation, a live call, a surgery, an incident response — the moments where the right fact at the right time changes the outcome — that cost is prohibitive. Stopping to ask is the same as not asking.

So the assistance either arrives on its own, or it does not arrive at all. That is the entire case for ambient agents in one sentence.

Three decisions that make or break an ambient agent

Building one taught us that the hard problems are not where we expected. Transcribing audio in real time is a solved problem; large language models are astonishingly good. The difficulty is entirely in the judgment layer — the part that decides when to act. Three design decisions ended up mattering more than any model choice.

First: silence is a feature. An assistant that interrupts constantly does not get turned off because it is wrong — it gets turned off because it is exhausting. We measure our copilot on trigger precision: the right card, at the right moment, and nothing the rest of the time. Every intervention it doesn't make is part of the product. This is the exact opposite of a chatbot, which answers every single time you ask.

Second: suggest intentions, not scripts. When the system detects an opening, it suggests "→ ask about last season's losses" — never the full sentence to read aloud. A person reading sounds like a robot; a person reminded of the right point sounds like an expert. This distinction seems small and is everything: it keeps the human being the one who performs, with the machine as memory rather than mouth.

Third: the knowledge has to be yours. An ambient agent that answers with what anyone can google is a novelty. Ours consults the company's own knowledge first — product sheets, competitive comparisons, the account's history — and only complements with the web. The agent's value is exactly the value of the private knowledge behind it. That is also, not coincidentally, what makes it defensible as a product.

The engineering shape

Under the hood, the pattern that works is a two-tier judge. A fast, cheap model watches the transcript continuously and answers a single question: is this moment worth acting on? Only when the answer is yes does a more capable model retrieve, reason and write the cue. Run everything through the big model and you get chatbot latency in a context that cannot tolerate it; run everything through the small one and the cues are not worth showing. Recent research formalizes a version of this — retrieving in parallel while the other person is still talking, with a judge deciding whether to search or stay quiet — and our production experience matches it: the gatekeeper architecture is what makes conversation-speed assistance economically and technically viable.

There is also an invisible requirement that no benchmark captures: the assistance must be private. In our copilot, the overlay is excluded from screen capture at the operating-system level, so the salesperson can share their screen and the cues remain theirs alone. An ambient agent the other party can see is not an assistant — it is a liability.

Where this goes

Sales calls are one instance of a much larger pattern: any situation where an expert performs live and the knowledge that would help them exists but is not reachable in the moment. Support agents with the knowledge base in their ear. Insurance brokers with the policy details surfacing as the client asks. Interviewers with the candidate's file and the right next question in sight. Field technicians, emergency coordinators, teachers.

The chatbot decade taught everyone to go to the AI. The next one, we think, is about AI that comes to you — precisely, briefly, and only when it has earned the interruption.