Claude for Dummies (SMB Owners)
Source: http://x.com/i/article/2038022950935519232
In the rapidly evolving AI boom, it feels impossible to keep up with new technology.
This March alone, Anthropic averaged more than one major feature launch per week.
The running joke in AI spaces is that people need to be unemployed to avoid falling behind.
SMB owners are in the unique situation of being decision makers in their organization but not having the time nor skillset to implement AI in a useful way. X is littered with "how to" articles, but these target audiences of indie hackers and soloprenuers without the same level of responsibility or risk limits that come with running a business.
In this article, I will share how I implemented a multi-agent architecture in my construction business across sales, operations, finance, and my personal life. At the end, I will provide steps to copy/paste my system to skip months of catch-up work and start with an AI assistant bespoke to your business with minimal risk or added investment.
The entire AI world continues to evolve, so this is not perfect nor does it push the boundaries of what is possible. But it will get you to a place where you can keep up with relatively little effort and update your own system when new needs arise.
Why My System Will Help Specific Dummies
In my past life, I worked in technology and systems in two diametrically opposed organizations. First, I was a Systems Project Manager at McMaster-Carr, a century-old industrial supplier with a tech stack that spanned building their own data centers onsite to being a leader in machine-learning implementation to delivery online results instantly across 1.7M+ SKUs. Then I was a Technical Program Manager at SnapLab, working on AR wearable technology alongside engineers pushing the limits in areas like computer vision and 3D rendering. I share this because keeping up with the AI wave has required me to lean on my past career experience in a way the typical SMB owner cannot.
Still, I would consider myself "non-technical" (aka "Dummy"), which has allowed me to build my AI system in a way that is also achievable for other owners.
Cooper Demolition is a site preparation specialty subcontractor with operations in a handful of states. We are among the most AI-resistant trades out there. Many of my fellow SMB owners feel they are in a similar position, but still worry that they are missing the big AI wave.
My system below is tailored to any owner who might feel intimidated but still wants to explore how AI can improve their business. If a demolition contractor can use this, so can you.
Summary of the "Peggy" Multi-Agent Architecture The Hardware
Peggy runs on a Mac Mini in my office that never sleeps, executing tasks on a schedule with no screen and no clicks required. I access it remotely from anywhere using Tailscale, which makes my laptop and the Mac Mini behave as if they're on the same desk. All of Peggy's memory lives on Airtable, accessed by API calls, so every part of the system reads from the same shared brain.
If you have a spare laptop or computer around in the office, you should be able to run this without extra investment. The $600 Mac Mini is a bit of overkill, so your call.
The Tech Stack
My setup prioritizes using existing databases and subscriptions to minimize the implementation effort. It is intended to use a combination of Claude, an email client, a database provider, and a remote access program. Most businesses already pay for their email and database service, so Claude becomes the main net-new investment.
Cooper Demolition uses:
- Claude Max (you can probably use Pro at first) - $100/mo
- MS 365 (Outlook) - Previously existing cost
- Airtable - Previously existing cost
- Tailscale - Free tier
To be clear, there are more elegant ways to do this if you are a tech-first organization starting from scratch. This setup gets you off the ground with little risk and minimal changes to your existing data infrastructure. This works with gmail, Notion, OneDrive, etc.
The Persona
Every agent in the system operates under the same identity (Peggy Olson, Executive Assistant at Cooper Demolition) and shares the same rulebook, CRM, memory file, and to-do list. This shared context is what makes the system feel coherent rather than a pile of disconnected automations. This also means that any external correspondence follows the same tone and fine-tuning I've done over time. Think of this as a conductor who directs the agents smartly and cleans up their outputs before they are used. Peggy has her own company email, and my team interacts with her by emailing like they would with any other teammate.
The Agents Responsibilities are segregated across multiple agents for two main reasons: parallel task performance and risk mitigation. Peggy can launch multiple agents at once for efficiency. Each agent only has access to their specialty area files to provide some guardrails on how they interact with data as they execute the tasks. This ability to "sandbox" is a major reason for widespread Claude adoption when compared to options like OpenClaw.
Finance Agent handles the money side: pulling live job data from our project database, generating the weekly Work in Progress report for our accountant, maintaining the 13-week cash flow model (planning AP and AR activities), and running AR aging and reconciliation on a fixed schedule. It all lands in the right inboxes before the workweek starts, with no action from me. Eventually, I will implement AR collections to run through this agent with direct follow up to my clients.
Operations Agent manages the inbox and keeps things from falling through the cracks: scanning email every 30 minutes, triaging what needs my attention, maintaining an up-to-date to-do list, and sending me a morning briefing before I wake up so I know exactly where things stand. This includes compliance tracking, project submittals, change order execution, and project budget tracking. This agent directly interacts with my project managers to provide support and keep them on track.
Sales Agent is where we track bids, pipeline, and client relationships: keeping the CRM current with auto-captured contacts, monitoring deal status, and making sure nothing on the revenue side goes quiet without a flag. While this agent is currently limited to handling the top of our sales funnel, it will eventually manage everything from initial outreach to full estimation support/submission.
Personal EA handles the things that don't fit neatly into a business category: travel documents, loyalty programs, important dates, and relationship context in my CRM. To be honest, I have not leaned on this agent much so far but there is a lot of low-hanging fruit here.
How I Interact With Peggy
With Claude's release of Dispatch, I primarily interact with Peggy through a single chat session on my phone or laptop. Dispatch has full visibility into everything Peggy does, so I don't have to provide reminders and can easily jump between topics. Having a single point of contact with the system makes this all feel more human, which is where the accessibility unlock can happen for SMB owners. I, along with other employees, can also email Peggy at any time if that makes more sense.
Because of my non-technical focus, Peggy leverages Claude Cowork scheduled tasks to run things in the background. More technical users would probably use Claude Code, but every Anthropic release indicates a future convergence of capabilities between the two products. For this reason, I feel comfortable using Cowork due to the approachable user interface.
Next Step For You, Dummy
If you think getting set up with a similar architecture is useful and worth a couple hours to get caught up, drop me a follow and reply to this article with "I'm a dummy". Within 24 hours, I will DM you with a file that will behave like an installation wizard and set itself up in a self-tailored way.
Keep in mind that the agents I am using may not be what you need, and that's totally okay.
This setup will cater to your business needs depending on how you interact with it. You will have custom agents fine-tuned to your business.
The process will look like:
- Download the .txt file
- Drag/drop into Claude Cowork on your Desktop App (ideally on the "always on" machine you plan to use)
- Prompt "setup assistant" in the Cowork chat session
- Answer 25 questions in the Claude Cowork Chat (15 min)
- Claude Cowork will build itself. This will include company-specific details, multiple agents, data connections, and suggestions on your hardware setup. You will need to provide clarifying details when prompted, but if you ask questions it will guide you through setting everything up.
- You will talk to your assistant to provide feedback and fine-tune according to your preferences. A feedback tracker is already part of the setup, but the more you provide the better.
If people find this useful, I will keep sharing updates that folks can paste into their system as new capabilities arise.