A lot of contractor AI talk is aimed at the wrong problem. It is usually not dispatch optimization or a giant back-office rebuild. It is this: call comes in while the office is busy form hits after hours customer texts half the info nobody follows up fast enough lead leaks First workflow to fix.

👀

This is huge! Now you can connect Claude code and context across agents, across teams, store memories and track projects and so much more.

Notion is quickly becoming the solution to agent memory

Talking to small business owners about AI this week confirmed something: They show up asking what tool to buy. Twenty minutes later they're describing how their business actually runs in a level of detail they probably haven't articulated to anyone in years. AI is forcing operators to map their own businesses again. Strangely, that's the most valuable thing about it right now.

If you're looking for help deploying AI in your business, but aren't quite big enough to hire OpenAI directly, I'm your guy. Feel free to reach out with questions, ideas or if you just want to brainstorm. I'll help you figure out how AI can transform your business and what the best tools for the job are. DMs are open 🤘

The most common reason an AI tool fails inside a small business has almost nothing to do with the model. A good agent setup is something like 20% prompt and 80% context infrastructure — written rules, customer history, formatted examples, definitions of internal terms, edge cases, and patterns about what your team chooses to do when nothing's obvious. You can't tell from a tool's homepage whether it's any good. You find out 30 days in, when the thing has either been fed enough about your business to be sharp or hasn't. Stop evaluating AI tools by their model. Evaluate them by how much your specific business shows up in their outputs after a month.

How a one-person agency runs four active client engagements: Most of the recurring work isn't being done by me anymore. It's being done by a stack of small agents running on a server I own, on a schedule I set. Email triage that routes inbound by project and drafts replies in each project's voice. Infrastructure watcher that catches drift before the client notices. Morning briefing that loads the state of every active project before I sit down. Sales research agent that handles competitive intel before any call. None of these are flashy. Each one saves between 30 and 90 minutes a day. Together, about seven hours a week recovered. That's the entire reason this works.

The "AI is replacing jobs" frame is mostly looking at the wrong layer. Most jobs are about 20% intelligence work and 80% judgment, escalation, and absorbing the irregular stuff that breaks every clean system. AI handles the 20% well. The 80% still needs a person. The bookkeeper isn't doing math. The salesperson isn't pitching. The dispatcher isn't routing. They're absorbing the chaos that would otherwise hit the rest of the team. What AI gives you back is hours. The dispatcher absorbs more chaos. The bookkeeper untangles more weird transactions. The salesperson runs more parallel threads. You're not replacing the team. You're upgrading them by removing the part of the work that wasn't really their job to begin with.

Aaron Levie wrote something a few weeks ago that I keep coming back to in client work: An outsourced task tells you three things — the company has accepted this work can be done outside the team, there's a budget line attached, and the buyer is purchasing an outcome. Replacing a vendor with an agent is a vendor swap. Replacing a person with an agent is a reorg. The first happens quietly. The second carries the political weight of a layoff even when nobody's being laid off. So when a small business asks where to start with AI, the question I ask isn't "what does your team need." It's "what are you already paying someone outside the company to do." That's the wedge, almost every time.

Shipped the first production AI Instance for a client a few weeks ago and I want to write about it without the usual tech-blog framing. The setup: one server, one CLI command, an hour later there's a fully isolated AI environment running under the client's branded subdomain. Their own database, their own agents, their own context layer, their own auth. Nothing shared with anyone else. The reason this matters is that almost everything currently sold to small businesses as "AI" is a wrapper sitting on top of shared infrastructure. Same models, same prompts, same defaults across every customer. There's no real compounding because the agent never gets meaningfully smarter about your business specifically — it isn't really inside your business in any structural sense. What I'm building runs the other way. The compounding lives in the context layer, and the context layer only exists if you control the environment. Agents that get sharper about your operation over time because they live with your data, your documents, your team, and your historical patterns. One server, multiple clients, no team behind it yet. The entire stack runs on me and my own agent setup, which is the part I'll keep writing about as the count goes up.

Every small business owner who calls me about AI shows up wanting a roadmap, a vendor evaluation, and a 90-day plan. Almost always wrong for businesses under 50 people. What works at this scale is a weekly habit. Pick the task that annoyed you most last week. Build the dumbest version that does it. Use it for a week, pay attention to what breaks. Then pick another one the next Monday. Three months in, you have eight to twelve agents running, all built specifically for your business, by people on your team who now know how to build more. Strategy is the thing you write up after that, not the thing you write to plan it.

people keep asking what "context" actually means when working with claude treat it like a new hire write down what your business does. paste in examples of good work. tell it your rules. put all that in a project or cowork folder then start working. give feedback as you go. "more like this, less like that" it gets sharper every session because youre actually training it on your business almost everyone who tells me claude didnt work for them just skipped this part