The average office worker receives 121 emails a day. Their phone buzzes with 46 push notifications. Their Slack has unread threads from three days ago. And nearly half of all those push notification recipients will opt out entirely if they get more than five messages a week.
Every communication channel we've built in the last twenty years is drowning in its own success. Email worked — so we flooded it. Push notifications worked — so we spammed them. Slack worked — so we turned it into a second inbox nobody asked for.
But here's what nobody has figured out yet: how to let someone else — your coach, your org, your advisory chair — send you a message that shows up inside your ChatGPT session.
Not an email about ChatGPT. Not a Slack link to a prompt. A message from another person, waiting for you the next time you open your AI and ask a question.
The Attention Context Hierarchy
Not all communication channels are created equal. The difference isn't delivery speed or open rates — it's the attention state of the person when the message arrives.
Email hits a queue. The recipient is triaging, not thinking. McKinsey's research shows knowledge workers spend 28% of their workweek just managing email — reading, sorting, responding to messages where only about 24% are actually business-critical. That's nearly a third of your week spent on a channel where three-quarters of the content doesn't matter.
Push notifications are interruptions. They catch someone mid-scroll, mid-task, mid-conversation. The average smartphone user sees 46 of them daily. Open rates drop below 5% when you send more than three a day. The channel is structurally adversarial — you're competing against the thing the person was already doing.
Slack and Teams sit somewhere in between. Medium-attention. The user is aware of the channel but drowning in notification fatigue. Most messages get a glance and a mental "I'll get to that later" that never materializes.
Now consider the AI assistant. When someone opens Claude or ChatGPT, they're in a fundamentally different cognitive state. They're actively thinking. They're working through a problem, asking questions, processing information. The average session on these tools runs 14–18 minutes of sustained, focused interaction.
That's not a queue. That's not an interruption. That's the highest-attention context available in modern knowledge work.
And right now, no expert, no organization, no platform has a way to reach someone inside that context. You can email them. You can push-notify them. You can post on social and pray. But you cannot send a message that arrives inside their AI workflow, at the exact moment they're thinking about the thing you have to say.
The Sender Problem Nobody's Solved
At Skill Refinery, we've been building AI-native knowledge delivery — experts package their IP into skill cards, subscribers access that knowledge through the AI tools they already use via MCP and API integrations. The subscriber never visits a dashboard. They never log into a portal. They just ask their AI a question, and the expert's knowledge shows up in the response.
That architecture created an interesting problem: experts had no way to reach their subscribers.
An expert with 500 subscribers and a new framework to share had exactly one option — send an email that would land in an inbox alongside 120 other messages. Or post on social media and hope the algorithm cooperated.
The insight was simple: if a subscriber's AI tool is already calling our API every time they ask a question, we have a delivery channel. Not push. Not email. A message queue that surfaces the moment someone opens their AI and engages with anything Skill Refinery-related.
Think of it this way — it's like being able to surface a message inside someone's search results, but only when they're actively searching for something in your domain.
How It Actually Works
The technical pattern is straightforward, and that's part of why it's powerful.
MCP (Model Context Protocol) and custom GPT Actions are pull-based. The subscriber's AI tool initiates a request. There's no server-push mechanism — you can't send a notification to someone's Claude session the way you'd send a push notification to their phone.
So we built an inbox pattern. When an expert, org admin, or advisory group chair broadcasts a message, it goes into a message queue. When the subscriber's AI tool makes any Skill Refinery tool call — asking a question, pulling a framework, checking a skill card — our middleware checks for pending messages and appends them to the response.
From the subscriber's perspective, it feels like getting a notification the moment they open their AI. They asked about quarterly planning, got their answer, and then: "New message from [Expert Name]: Updated Q2 planning framework now available — includes revised KPI benchmarks based on March market data."
The knowledge answer always comes first. The message enhances the interaction. It never hijacks it.
The Design Decisions That Are the Product
Here's where most people building communication infrastructure get it wrong. They optimize for sender reach and ignore recipient experience. That's how every channel in history has been ruined.
We made specific design choices that are the product, not afterthoughts.
Messages are capped. A subscriber never sees more than two messages per tool call, regardless of how many senders are active. Someone with ten expert subscriptions doesn't get hit with ten messages at once. There's also a daily ceiling per subscriber — the system protects the experience even if every sender broadcasts simultaneously.
Auto-dismiss after repeated delivery. If a subscriber has seen the same message across three separate sessions and hasn't engaged, they're not going to. The message disappears. This prevents the stale-message pile-up that makes people abandon every notification system ever built.
Subscriber control lives inside the AI. No dashboard. No settings page. No separate login. You tell your AI "stop showing me messages from [sender]" or "turn off broadcast messages" and it happens. The AI is the inbox. The AI is the control panel. The subscriber never leaves their workflow.
Knowledge first, always. The subscriber asked a question. Answer it. Then surface the message. This is the opposite of how advertising works — interrupt first, deliver value second. In this system, value comes first. The message is additive.
Three Sender Personas, One Channel
The messaging layer serves three distinct use cases, each with different rate limits and message types.
Experts with subscriber audiences. A leadership coach with 200 subscribers can announce a new framework, share a timely insight, or nudge subscribers toward content they haven't engaged with. Limited sends per day — because the moment an expert starts broadcasting daily "check out my podcast" messages, the channel is dead.
Enterprise org admins. A company with 500 staff members using Skill Refinery for internal knowledge can push protocol updates, compliance changes, and operational announcements directly through the AI tools those employees already use daily. Some messages — safety protocols, compliance requirements — can be flagged as non-dismissable until expiration. When an equipment manufacturer pushes a safety protocol update, it needs to reach every technician, not sit unread in an email thread.
Peer advisory group chairs. A chair running a monthly advisory session can send prep materials, follow-up action items, and session reminders that reach members the next time they engage with their AI. No more "did everyone see the email?" No more prep materials that nobody reads before the meeting.
The Honest Limitations
This isn't email. A subscriber who doesn't engage with their AI tool for two weeks doesn't see the message for two weeks. Delivery depends on the subscriber's usage pattern, not the sender's schedule.
We also can't guarantee the AI displayed the message. We can confirm a message was included in a tool response. We can't confirm Claude or ChatGPT rendered it prominently — or that the subscriber read it. "Delivered to AI session" is the metric, not "read." Setting that expectation clearly with senders is critical. Overclaiming delivery is how you lose trust.
And the existential risk is real: spam kills this product. The thing that makes in-AI messaging valuable — high-attention context, low noise — is the same thing that makes it fragile. One bad actor broadcasting daily promotional content and the subscriber opts out of everything. The rate limiting isn't a technical constraint. It's the entire product.
Why This Couldn't Exist Before
This infrastructure is architecturally difficult to replicate outside of MCP and API-based knowledge delivery systems.
Course platforms can't do it — they don't have a tool call layer inside the subscriber's AI. LMS platforms can't do it — they operate inside their own interface, not inside Claude or ChatGPT. Traditional SaaS can't do it — they'd need MCP server integration as a foundational architecture, not a bolt-on feature.
Skill Refinery can do it because the entire platform was built AI-native from day one. The MCP server isn't an integration — it's the product. Every subscriber interaction already flows through tool calls, which means every interaction is a potential delivery point.
The pattern is technically straightforward for anyone building MCP servers — but being first in a communication channel creates compounding advantages that are hard to unseat. Network effects, subscriber trust, and sender adoption all compound with time. The infrastructure is simple. The positioning is not.
What Comes Next
The message queue is Phase 1. It's the infrastructure layer that enables everything after it.
Phase 2 is contextual messaging — messages that reference what the subscriber is currently working on. "You're building a quarterly plan — here's a framework update that's relevant right now."
Phase 3 is proactive coaching nudges. "You haven't reviewed your quarterly planning framework in 6 weeks. Your planning cycle typically starts around now." That's not a notification. That's an AI-powered coach that knows your patterns and reaches you at the right moment.
Phase 4 is automated behavioral triggers — the system watching for dormant subscribers and generating reconnection nudges, not from a queue, but fresh at the moment of reconnection.
Each phase builds on the same infrastructure. The message queue is the foundation. What you put in the queue gets more intelligent over time.
Who This Is For
If you're reading this and thinking "I want that for my audience" — good. That's why I'm writing it.
If you're an expert, coach, or consultant who's built an audience and wants to reach them where they're actually paying attention — not another email they'll archive, not another social post the algorithm buries — Skill Refinery is where you set up your storefront. Your knowledge gets delivered through your subscribers' AI tools, and soon, your messages will too.
If you run an organization where staff are already using AI tools daily and you're tired of compliance updates, protocol changes, and operational announcements getting buried in email — this is the delivery channel that meets your people where they're already working. Book time with me and I'll walk you through what enterprise access looks like.
If you chair a peer advisory group and you're sick of sending prep materials that nobody reads before the session — imagine those materials surfacing the next time a member opens Claude and asks anything related to your group's focus. Let's talk about what that looks like for your group.
I'm Matt Cretzman. I build AI agent systems through Stormbreaker Digital and products like Skill Refinery, TextEvidence, and LeadStorm AI. If any of the above sounds like your situation, don't sit on it — grab 15 minutes on my calendar and let's figure out the fit.