How To Add AI To Your Business Without Breaking Everything That Already Works

It's not a magic robot. It's an LLM call on a while loop. Mike Vanamburg has built AI agents for companies across industries — from cannabis dispensaries to bookkeeping firms to his own business — and the lesson he keeps learning is the same every time: start with the human, not the technology. In this episode, Mike and Paul break down the real blueprint for integrating AI into a business, why fully autonomous agents always fail, what the difference between a workflow and an agent actually is, and what work might look like in ten years for the people who get this right. **In this episode:** - **\[00:00]** Why most people are approaching AI adoption completely backwards - **\[00:55]** Start here before building anything: map the human workflow first, then add AI - **\[02:08]** How customer journey mapping and process flow audits expose what you didn't know you were doing - **\[04:11]** Should you use AI for the customer or for internal operations first? Mike's answer - **\[04:30]** How one company replaced a multi-month competitor research process with a two-minute AI agent - **\[05:55]** Why three clicks is now too much friction — and what chat-based UI is replacing it with - **\[07:01]** Big companies vs. small companies: how AI implementation looks completely different at each size - **\[08:44]** Are AI tools raising output standards or just giving people more time back? What Mike is actually seeing - **\[09:45]** The three types of clients: step away more, do 10X more, or just meet the new standard of excellence - **\[10:27]** The most unexpected challenge with AI implementation: it's always the people - **\[11:54]** Why you have to ask clarifying questions before you build a single agent - **\[13:09]** Why every fully autonomous AI agent Mike built ended badly — every single time - **\[13:36]** The human-in-the-loop requirement: LLMs lie, degrade, and mispredict without oversight - **\[15:03]** Why the data behind your AI matters as much as the AI itself - **\[16:12]** The new problem nobody talks about: information isn't scarce anymore — attention is - **\[17:44]** What format should your AI output be in? PDF, short-form, feed? Why this question matters more than you think - **\[18:13]** Why Paul now has to schedule time to review what used to take 20 hours - **\[19:13]** What's actually happening behind the AI curtain: workflows vs. agents explained simply - **\[21:37]** Open source vs. proprietary models: when to use ChatGPT, Claude, Gemma, or a custom harness - **\[24:03]** Why Mike builds every custom harness to work with multiple models — not just one - **\[24:36]** Deep dive: how AI agents are being built specifically for the cannabis industry - **\[26:12]** The dispensary compliance problem — and how an AI agent handles metric tracking and regulation - **\[27:52]** From bud tender backup to inventory purchasing intelligence: what a dispensary AI stack looks like - **\[29:03]** What does work look like in 10 years? Mike's honest answer - **\[31:30]** Why AI can only copy — and why studying your field anyway is still the right move - **\[32:01]** Why software developers still haven't been replaced after years of predictions