Every year someone announces the death of email. Every year, four billion people send another trillion messages. Not because email is convenient — it rarely is — but because it is universal. Every device, every operating system, every age group, every country. No installation required, no permission slip, no account to create. You already have it.
That universality is exactly what makes email the most interesting deployment surface for AI agents in 2026.
The App Dream, and Why It Failed
For fifteen years the dominant model for launching a new service was: build an app, persuade people to download it, persuade them to create an account, persuade them to return. Each of those steps is a funnel that leaks. The average smartphone user downloads zero new apps per month. The average B2B SaaS product sees fewer than 20% of registered users active in any given week.
Apps made sense when the smartphone was new and the novelty justified the friction. That era is over. The marginal cost of asking a user to install something is now very high — attention is scarce, home screens are full, and people have learned to be suspicious of new apps requesting permissions.
Meanwhile, email open rates for well-targeted newsletters sit between 40% and 60%. People check email before they open any app. The inbox is, behaviorally, the most-visited interface in digital life.
The inbox is not a legacy channel. It is the universal interface that survived every wave of disruption because nothing else has matched its combination of reach, reliability, and ambient availability.
What Changed: LLMs as the Missing Layer
Email as a service delivery channel is not new. What is new is the ability to run sophisticated, stateful, contextually aware conversations through it without human agents on the other end.
Large language models can now read and write email that is indistinguishable in quality from human correspondence — and do it at latency low enough to feel like a live conversation. They can maintain context across many exchanges, follow complex instructions embedded in replies, adapt tone and style, and escalate to humans when genuinely needed.
This is the missing layer. Email already had the reach. LLMs supply the intelligence. The combination produces something genuinely new: a conversational service that reaches everyone, requires no installation, and can handle nuanced, open-ended requests.
The Anatomy of an Email-First Service
What does such a service look like in practice? The pattern I have settled on after building Silex and Aleik has three components:
Outbound initiation. The service sends the first message — a research digest, a weekly briefing, a prompt. This is not spam; it is the service doing its job. The subscriber signed up to receive it. The message is the product.
Reply-driven direction. The subscriber responds in plain language. No forms, no dropdowns, no interface to learn. They write what they want the way they would write it to a colleague. The LLM reads the reply, understands intent in context, and acts accordingly.
Delivery back to the inbox. The result — an article, a report, an analysis, a booking confirmation — arrives as an email reply. The subscriber stays in the one interface they already use constantly.
The entire interaction is asynchronous, ambient, and frictionless. It runs on infrastructure that has been reliable for forty years.
On Reliability and Trust
One underappreciated advantage of email-first architecture is the audit trail. Every exchange is logged in both parties' inboxes. There is no black box. When an AI agent does something unexpected, the user has a complete record of the conversation that led to it. This transparency is valuable in itself, and it becomes even more valuable as AI agents take on consequential tasks.
The services I find most promising are those where the stakes are moderate — high enough that getting a good result matters, low enough that an imperfect result is not catastrophic. Research writing is a good example. Financial analysis is another. Coaching and advisory services are a third. These are domains where the combination of LLM capability and human oversight via email approval workflows produces reliably good outcomes.
What This Means for Builders
If you are building an AI-powered service and you are starting from scratch, I would encourage you to ask: does this need to be an app? Does it need a web interface at all? Or can the entire workflow happen through email?
The constraints of email — asynchronous, text-first, no rich UI — are also its strengths. They force you to make the service useful enough that people will seek it out in their inbox rather than requiring you to fight for home screen real estate. They push you toward services that are ambient rather than demanding.
The inbox has survived every disruption since 1971. It will survive whatever comes next. Building there is not a compromise. It is a choice to meet people where they already are.