By Sam Roux
Ernest Hemingway has a fantastic quote, which is admittedly referenced far too often by writers, myself not being an exception: "The first draft of anything is shit."
Ours certainly was.
The first version of HyperLocal, our AI-powered messaging assistant, launched with real conviction and middling results. Adoption was slow. The ICP was narrow. The setup process felt like assembling flat-pack furniture with half the instructions missing. What we believed was a clever product turned out to be a capable talker with nothing useful to say once the small talk ran dry.
To be fair to ourselves, the core insight was sound: SMS is one of the most direct channels a business has to a customer, and AI could make that channel genuinely intelligent. We just built the wrong product around a right idea.
That gap is the entire story of HyperLocal 2.
A Chatbot That Cannot Act Is Just a Chatbot
The original HyperLocal operated on a generous assumption: if the assistant could hold a coherent conversation over text, businesses would find their own reasons to love it. That turned out to be optimistic in the way that most first-draft assumptions are.
SMS is familiar. Conversational interfaces are intuitive. But when a system cannot touch the records, workflows, or operational state a business runs on, the conversation hits a wall fast. Early customer conversations made this uncomfortable clear. Airbnb hosts (our initial target market) were not philosophically opposed to automation. They were opposed to automation that floated above their business without connecting to it. They asked a question we should have been asking ourselves from day one: what does this actually do, besides reply?
HyperLocal 2 is our answer, delivered the only honest way: by building a product that actually does things.
From Polite Companion to Infrastructure
The single most consequential change in HyperLocal 2 is where it lives in the stack. The original product sat on top of the business like a friendly receptionist who had never been given access to the filing cabinet. The new version is wired directly into CRM state.
That shift sounds architectural because it is. When conversations are linked to contacts and deals, and when the assistant can read and write against those records in real time, the system stops being a chat interface and starts being an operational tool. The practical difference is significant.
HyperLocal 2 can now perform permissioned CRM actions when the operator enables them:
- Creating or updating contacts when a conversation warrants it
- Managing calendar events (scheduling, rescheduling, cancelling) when calendar capabilities are active
- Persisting conversations as part of the operational record, rather than letting them vanish into a chat log nobody reads
This is not a feature upgrade. It is a product redefinition. The old system was conversational software. The new system is messaging infrastructure that happens to be conversational.
Solving the Right Problem Is Harder Than It Looks
Another uncomfortable truth from the first launch: Airbnb hosts were simply not the right market. The use case was too narrow, the automation anxiety was real, and the value proposition required more imagination than most buyers are willing to extend.
HyperLocal 2 is built around a sharper, more durable problem: tenant-governed messaging operations over SMS, embedded in a CRM environment. It is less poetic than "AI concierge," but it describes something businesses will actually pay for. We are not selling the dream of a smart assistant anymore. We are building infrastructure for businesses that need routed messaging, structured record linkage, controlled automation, curated knowledge, and a clear audit trail.
Novelty fades. Operational value compounds. We are in the second business now.
Helpful Within Limits, and That Is the Point
One of the most important architectural additions in HyperLocal 2 is what we call "permissioned agency," a phrase doing some real work.
We did not want a system that freelances. We wanted a system that is genuinely useful within explicit, operator-defined boundaries. Agent profiles now carry capability flags that determine exactly what the runtime is permitted to do. Contact management enabled? The system can act on contact-related requests. Calendar management off? It will not pretend otherwise. It redirects and escalates cleanly instead.
This matters for two distinct reasons. When an operator wants automation, the assistant becomes meaningfully more capable. When they do not, it becomes meaningfully safer. Both outcomes are intentional. The first release had neither with any reliability. HyperLocal 2 was designed around this discipline from the start.
Intelligence That Actually Comes From Somewhere
A quieter weakness in the original product was that its "intelligence" was largely borrowed. It was in reality loosely prompted, weakly grounded in the actual business context. It sounded informed. It was not always actually informed.
HyperLocal 2 addresses this with operator-controlled knowledge sources. Tenants can now register, sync, and assign curated knowledge directly to their agents. At runtime, responses are grounded in that material rather than assembled from whatever the model happens to know about the world at large.
The improvement in answer quality is obvious. What is less obvious, but more important, is the change to the trust model. Operators know exactly where the context comes from. They can update it, audit it, and control which agents have access to which sources. That is a fundamentally different relationship with the system than simply hoping the model is quietly doing the right thing.
An Admin Experience That Does Not Require a Developer
The setup process for the original HyperLocal was, to put it charitably, technical. Even when the concept landed in a sales conversation, the path from "yes, I want this" to "this is running in my business" resembled a developer integration more than a product deployment.
HyperLocal 2 is organized around clear operational primitives that non-technical operators can actually reason about:
- Agent profiles
- Routes
- Knowledge sources
- Conversations
- Permissions and policy controls
The Messaging workspace is becoming a genuine control plane: a place where routes are configured, knowledge sources managed, webhooks inspected, and permissions clearly understood. The first version had none of that legibility. The second version is designed to be administered, not merely installed.
Failing Well Is a Feature
Not every improvement in HyperLocal 2 makes for a compelling demo. This one certainly does not. But it may be the most professionally important one.
The system now has intentional failure behavior. If AI completion is unavailable, the assistant falls back to a bounded, business-safe response rather than producing something incoherent in the middle of a customer interaction. If a requested action falls outside the agent's permissions, the system refuses cleanly and routes toward escalation. No bluffing. No silent failures. No hallucination dressed up as product behavior.
The first release asked users to trust that the AI would handle things. HyperLocal 2 earns trust differently: by being predictable when things go wrong, not just impressive when they go right.
The Smarter System Is the More Constrained One
This is the part that surprised us most in building the second version, and it is worth saying plainly.
We started this work believing that intelligence would come from the quality of the conversation: the fluency of the model, the elegance of the prompts, the cleverness of the interface. We now believe the opposite is closer to the truth. The real intelligence of the system comes from the quality of its constraints: what records it can see, what actions it is permitted to take, what knowledge it draws on, what route it operates within, and how it behaves when certainty runs out.
That is the philosophical shift underlying every concrete change in HyperLocal 2.
It is not just a more articulate product. It is a more grounded, more operational, more governable one, and a great deal more honest about what it can and cannot do.
The first draft was shit. This one is closer to something we are proud of.