How a WhatsApp AI Assistant Like Instinct Works
There’s no app. You don’t log into a dashboard. You text a phone number like you’d text a friend, “rebook my Friday flight, I’ll miss the connection”, and some time later a message comes back: done. That’s Instinct, the invite-only AI assistant that went from private beta to a $10 billion valuation in about a year, and it’s the clearest example yet of where personal assistants are heading.
The interesting part isn’t the chat. It’s what happens in the gap between your message and the reply. Instinct has no magic API into the airline. It does the thing you’d do: it opens a browser, logs into your account, clicks through the rebooking flow, and if the website won’t cooperate, it picks up a phone and calls them. Let me walk through what’s actually public about how it works, then how you’d design a system like it.
What it feels like
· searching · reading menus · checking tables
How it actually works (the confirmed parts)
Three facts from Instinct’s own site and the reporting around it tell you almost everything:
It has no new interface. In its own words, “there are no new interfaces. You can text or call it. It can use a phone and computer the same way that you do.” That last line is the whole design. Instead of integrating with a thousand services, it operates a real computer and a real phone like a person, so it works even on a website or a shop that has no API at all.
It connects to your digital life to get the job done. With your permission, it links to your email, messaging apps, and calendar. That’s how it finishes real tasks: when a booking site texts a confirmation code, it reads the code from your inbox and types it in. When a login needs a six-digit authenticator code, it handles that too.
When the web won’t do, it calls. Instinct added a “concierge” that literally phones businesses, to book the restaurant that only takes reservations over the phone, or chase a refund that needs a human on the line.
And the assistants are starting to talk to each other. It has a “trusted person network” where your assistant can coordinate plans with your friends’ assistants, and it’s wiring up to partners through proper tool protocols (for example, booking doctor visits through Zocdoc’s MCP connection) rather than clicking through a browser when a clean API exists.
How you’d design a system like it
Here’s where I’ll be honest: Instinct hasn’t published its architecture, so the pipeline below is how I’d build it, grounded in what’s confirmed, not a leak of their internals. But the shape is forced by the problem. If you want “text a number, it does real things,” you need roughly these stages.
How it connects, where the data comes from
The “no API needed” claim rests on four ways in. Most tasks use a mix.
Where it stores your data, and why that’s the scary part
To act for you, it has to remember you, your preferences, your history, the accounts it’s connected to. This is both the feature and the risk, and Instinct’s early stumble makes the lesson concrete.
There’s a nice footnote here. Instinct’s founder, Noah Shinn, wrote the well-known Reflexion paper, agents that improve by reflecting on their own feedback and storing the lesson in memory for next time. I can’t confirm Instinct runs that exact technique, but its “learn your preferences, do better next time” behavior clearly rhymes with that work. The memory isn’t a side feature; it’s the whole personality.
What’s confirmed, and what I’m guessing
Because it’s a closed product, keeping these straight matters:
| Claim | Confirmed (public) | Unknown / my guess |
|---|---|---|
| No app, text or call it | Yes, from its own site | |
| Drives a real browser + phone like a human | Yes, stated and reported | |
| Reads codes from your email to finish tasks | Yes, reported | |
| Indexes + retains your data, revised its terms | Yes, well documented | |
| Founder wrote the Reflexion paper | Yes | |
| The exact pipeline above | My design, not theirs | |
| Which AI model it runs | Not public | |
| The storage / infra details | Not public |
What I take from it
Two things. First, the winning interface for an assistant might be no interface, a phone number you already know how to use, backed by an agent that operates the rest of the world on your behalf. That’s a bigger idea than any app.
Second, the hard part was never the chat. It’s everything the pipeline above has to get right: operating real software without breaking, handling the login codes and phone calls, and above all, gating the moment it spends your money or acts in your name, and keeping a memory you can actually see and delete. Build an assistant that’s fast and capable but can’t be trusted with the “yes,” and you’ve built a liability. Build one that does the boring 80% and stops for you on the 20% that matters, and you’ve built the thing everyone’s racing toward. Instinct’s early privacy stumble is the whole industry’s lesson in miniature: capability is easy now, trust is the product.