HubSpot Wanted $80,000 a Year. He Built an AI Agent Instead.

Software used to sell you seats. Now it can just do the job.

A $80,000 invoice, and a different idea

Late last month, Lewis Carhart, co-founder of the compliance startup Comp AI, got a quote from HubSpot: roughly $80,000 a year, for a sales team of fewer than ten people. He did not sign it. Instead, according to his own account on X, he spent half a day building a replacement.

Then he did something most founders would not. He open sourced it. Free, MIT licensed, the code sitting in public on GitHub for anyone to read, copy, or run themselves.

That reaction is not really about a free CRM. It is about what people saw when they opened the code. The README says it plainly: most CRMs are a database with a form in front of it, and the AI ones just bolt a chat box onto the side of that form. This one is built the other way round. The agent is not a feature bolted onto the CRM. The CRM is where the agent keeps its notes.

It runs on its own schedule, against its own queue of work. It researches a new contact, checks what is already known, decides what still needs confirming, and writes down only what it can back up with evidence, never a guess dressed up as a fact. Close the laptop, and it keeps going.

What is an agentic CRM, and why does it matter for a small business?

An agentic CRM is customer software where an AI agent, not a human, does the ongoing work of researching contacts, updating records, and flagging what needs a decision, while a person simply reviews and steers. It matters to a small business because it replaces hours of manual data entry and lead research with a system that keeps working after everyone has logged off, at a fraction of the cost of a seat-based subscription.

That is the shift worth sitting with. For fifteen years, growing a small business meant buying more software, then paying someone to keep the software updated. A new lead came in, someone typed it into the CRM. A deal moved forward, someone updated the stage. A contact changed jobs, nobody noticed for six months. The software was never the bottleneck. The keeping-it-updated was.

An agent removes that layer. Not by being smarter than the software. By being the thing that does the typing, the checking, and the noticing, on a schedule, without being asked.

You are not behind. You are missing one word: agent.

Most small business owners already use AI. A chat window, a question, an answer, back to work. That is useful. It is also the ceiling of what a chatbot can do, because a chatbot only moves when you move it.

An agent is a different animal, and the distinction is not marketing. It comes down to one thing: what happens when you close the laptop.

A skill is a job, written down once and saved so it can be reused: a name, the task, the context it needs, the tools it is allowed to touch, and what it should produce. An agent is that skill, put to work and set loose to run on a schedule. You hand it an outcome one time. It works out how to get there on its own, every time after that.

That is exactly the design behind the Comp AI CRM, just built with engineering hours instead of a form. Eighteen separate tools, four written skills, one schedule that decides nothing except which piece of overdue work to pick up next. No confidence scores, because a model asked to grade its own certainty will grade generously. Evidence goes in the record. Guesses become a note for a human to check.

How do I build an AI agent for my small business without hiring a developer?

You do not need Comp AI's stack, and you do not need to write a line of code. Claude's desktop tool, Cowork, does the same job for a solo owner or a five-person team, through a plain-English setup. A recent walkthrough from the YouTube channel Marketing Forward shows the exact steps, using a made-up marketing job as the example. The steps map to any recurring task in any business.

1. Connect the tool where your business data already lives. Google Drive, Notion, Slack, HubSpot, whatever holds the documents or records the agent needs to read. This is a one-time setup, in the connectors menu, and every skill you build afterward can reach it.
2. Describe the job in five parts, in plain English. Name it something specific. State the job in one line. Give it the context it needs, such as a brand document or pricing sheet. Name the tools it is allowed to use. Describe the output you want and where it should land. Claude asks a couple of clarifying questions, then writes and saves the skill itself. No code, no file to build by hand.
3. Turn the skill into an agent with a schedule. Type a schedule command, tell it which skill to run and how often: hourly, daily, weekly, or only on demand. Pick an off-peak window, an evening or a weekend, when the system is quieter and your usage stretches further.
4. Let it run without your laptop being open. Scheduled agents run on the provider's own servers, not your machine. Close the lid on Sunday night, and the finished briefing, report, or update is waiting Monday morning. The one exception: anything that only lives on your computer needs the desktop app open to reach it, so keep source material in a connected drive, not a local folder.
5. Check it, then trust it with more. Review the first few runs from the scheduled tasks page before you stop watching. Let it run unsupervised for low-stakes work like research, summaries, and drafts. Keep a human in the loop for anything hard to undo, like sending an email, posting publicly, or deleting a file.

Build it yourself, or borrow one off the shelf?

You do not have to write every skill from nothing. Claude ships ready-made packs, called plugins, for common jobs: a competitive brief, an SEO audit, an email sequence, a performance report, each pre-built and installed in one click, each usable the same two ways, right now in a chat or scheduled to run itself. The Comp AI CRM works the same way in spirit, just for a specific job: an open source, ready-built agent for the one task, contact and company research, that eats the most unpaid hours in a small sales team's week.

The rule of thumb is simple. Borrow a ready-made skill when the job is a standard one you want solved today. Build your own when the job is specific to exactly how your team works, and no off-the-shelf version quite fits. Most businesses end up doing both: two or three borrowed skills covering the obvious jobs, one or two custom ones covering the thing that makes your business your business.

What you have after this

A week in, you have one working agent: a real job, off your desk, running on a schedule you set once. A month in, you likely have three or four, quietly covering the research, the reporting, and the follow-up that used to eat your Friday afternoons. Six months in, you have something closer to a small, tireless back office: a set of standing agents that work while you sleep, that never ask for a raise, and that cost nothing close to $80,000 a year.

Same business. Same team. A completely different week.

If you want help figuring out which job in your business is worth automating first, and building it properly rather than as a one-off experiment, that is exactly what the Claude skills work we covered recently is about, and it is the first thing we map out in the 14-Day AI Readiness Sprint.