Notes from a team building the inbox.
Original research, opinions, and case studies. The math behind our pricing, the cost of self-hosting an LLM, the metrics that actually predict retention, and the patterns we see across 4,200 workspaces.
What we learned from helping 300 teams migrate off Intercom and Zendesk
Three quarters in, the patterns are clear. Migration is rarely about the chat. It is about the four other tools sitting around it.
Self-hosting an LLM in 2026: the cost math that made us do it
We run our own AI on our own hardware. Why we did it, what it cost, and at what scale it stops making sense.
The real cost of per-seat pricing (the number is bigger than the bill)
Per-seat pricing taxes hiring. We modeled the second-order effects on five real customer accounts. The hidden cost ranges from 1.4x to 2.7x the bill.
RAG vs fine-tuning for support AI: stop fine-tuning
The most common AI support pattern in 2026 is grounded RAG. Fine-tuning still has a niche. Most teams should pick one and not the other.
Phones are back, and the AI is finally good enough to help
Voice traffic is up across every customer segment in 2026. Three things changed. Here is what we did about it.
The four customer support metrics that actually predict retention
We ran a regression on 14 months of customer outcomes. Two of the standard metrics correlate weakly. Two correlate strongly. Two new ones we now publish.
Intercom vs Message: the real cost at 5, 15, and 40 seats
Headline Intercom pricing is $39 per seat. The real bill is roughly 1.7x that. We modeled it on three team sizes.
AI on phones: what callers actually want (it is not what you think)
We ran post-call surveys on 12,000 AI-handled phone interactions. The findings flipped two of our product assumptions.
Writing help docs people read (and how to know if they do)
The classic 'long, comprehensive' help-doc playbook is wrong. Shorter, denser, opinionated docs win on every metric we measure.
Case study: Document.com cut response time from 11 hours to 6 minutes
Document.com runs an automated state-document generator. Speed is competitive moat. They consolidated four tools into Message and saved $1,400 a month on the way.
Every page on your marketing site is a sales page
The 'pricing page is the only conversion page' mindset misses where most conversions actually happen. Audit your help center as your most-trafficked sales page.
The 5 saved replies every support team needs (and the 20 that are just clutter)
We audited 30 customer workspaces' saved replies. The pattern: every team has 50 saved replies, uses 5 of them. Here are the 5 that matter.
Writing an AI policy for your support team (template included)
Most support teams enable AI without ever writing down what the AI can and cannot do. Three months later, an AI does something dumb and there is no one to blame except yourself.
What changed in customer support in Q1 2026 (the patterns from 4,200 workspaces)
We aggregated anonymized usage data across 4,200 workspaces for Q1. Five patterns that should change how you plan Q2.
One email per month. Worth reading.
New posts, customer data, and the occasional founder-rant. No promotional fluff.
Subscribe by email →