Notes from building an AI Practice.
The agent platform landscape of 2026
Five categories, one missing winner. The 2026 agent stack has split into hyperscalers, frameworks, workflow tools, durable engines, and vertical copilots — and the MSP layer is still unclaimed.
The best templates are the ones you didn't design
Every template marketplace starts with designers in a room guessing. There's a different starting point — don't design templates. Extract them.
A movement is named by a person, not a company
Tribes don't form around logos. The voice has to come from a person, not a brand. Person first, conviction second, product third.
Intelligence is cheap. Brains are not.
The MSP labor model is collapsing because the floor keeps dropping. What stays expensive is judgment, and AI makes judgment more valuable, not less.
Most of what you call "AI" costs almost nothing
Teams blame their AI bill on the model. The bill is a routing problem dressed up as a model-cost problem. Classify operations by the intelligence they actually need.
The retrieval problem nobody's talking about
A 1M-token window is impressive for thirty seconds. The needle in the haystack is a chain of needles, and most can only be identified after you've found the others on the chain.
Intelligence and context are multiplicative, not additive
Most teams treat reasoning and context as separate investments and get linear returns. The two axes compound. Bet on the product, not the sum.
Execution at the speed of trust
The 4.5% gap between 95% accuracy and 99.5% is not rounding error. It is the entire difference between a tool and an operator. Trust is the unit of execution.
The AI in your RMM should be a router, not an interpreter
Most AI-in-RMM features ship the model as interpreter. The model should be a router that picks a tool. The boundary belongs in the tool's code, not the prompt.
Prompt injection is a category mistake
Half the security industry is trying to sanitize LLM input. The bug is upstream — it's a type system problem dressed up as a content problem.
An email attack only works if a human is in the loop
Every common email attack requires a human to parse raw data they weren't qualified to parse. Take the human out and most of the attack surface stops existing.
Why is email treated different than any other data?
Email is text. Stop wrapping it in custom protocols and proprietary stores. Ingest it as data, link it to a customer, and the middle layers stop earning their keep.
There are no apps. There is only data and intent.
The desktop metaphor is forty years old. The "files" are database rows pretending to be paper. When the human stops being the reader, the translators stop being the product.
Ninety percent of computing exists to make things human-accessible
Most of the tower of modern computing is a translation layer between deterministic machines and unreliable wetware. When the interface stops being a human, most of it evaporates.