Two years ago, “AI for your business” meant a chatbot bolted onto your homepage. Last year it meant ChatGPT in a browser tab. This year it means something different: an agent that actually does work — reads your inbox, drafts the reply, files the invoice, schedules the follow-up.
The shift is real. The hype is loud. The question for an operator is narrower: which of this is ready to run in my business this quarter, and which still needs an engineer in the room?
Three names worth knowing.
Productized
Claude for Small Business
Anthropic · launched May 14
15 ready-to-run skills (invoice chasing, lead triage, month-end close, contract review, content strategy, and more) that plug into the tools small operators already use — QuickBooks, HubSpot, Canva, DocuSign, Google Workspace.
Framework
Hermes Agent
Nous Research · MIT, open source
A self-improving agent framework that creates skills from experience, remembers across sessions, supports 40+ tools, and runs on anything from a $5 VPS to a GPU cluster. Pitched as “the only agent with a built-in learning loop.”
Framework
OpenClaw
openclaw.ai · MIT, open source
A local-first AI assistant that runs on your machine, connects to WhatsApp / Telegram / Discord, manages email and calendar, browses the web, controls smart devices, and builds custom skills on demand. Tagline: “The AI that actually does things.”
There are dozens of others — Lindy, Relevance, Cognosys, custom builds on LangGraph or the Anthropic SDK. The three above are illustrative, not exhaustive. The point: there is no single “buy this” answer in 2026. There's a stack of choices, and the right one depends on what you're trying to automate.
What the demos don't show.
An “AI employee” is still a system, not a person.
It needs a job description, access to the right tools, a way to escalate when it’s confused, and someone to review what it did. Skip any of those and you’ve hired someone with no role and no manager.
Frameworks are not the same as products.
Hermes Agent and OpenClaw are powerful — and they’re open source. That means they’re free to start and expensive to operate without somebody who runs servers, debugs Python, and updates dependencies when the model upstream changes.
The agent doesn’t fix bad data.
If your CRM has 800 duplicate contacts and your inbox rules haven’t been touched in three years, the agent inherits all of that. The unsexy work is cleaning the inputs before the agent ever runs.
Four roles worth hiring first.
The inbox triager
Reads every inbound message across email and your contact form, classifies it (new lead, existing customer, vendor, spam), pulls the relevant context from your CRM, and either drafts a reply or routes it to a human. Operator wins back ~5 hours a week.
The invoice chaser
Watches QuickBooks for unpaid invoices, sends a polite reminder at day 7, a firmer one at day 14, a phone-call task to a human at day 30. Logs every touch back to the customer record. Operator gets paid faster without becoming the bad guy.
The proposal drafter
When a qualified lead comes in, the agent assembles a draft scope, prices it against your standard rate card, drops it into your template, and shares it for human review. The salesperson reviews and sends — not writes.
The follow-up automator
Post-job, sends a review request to the right platform, queues a maintenance reminder for the appropriate interval, and flags the customer for a win-back campaign if they go quiet. Builds the recurring revenue your CRM was supposed to.
We set up the agent.
You run the business.
We've been building the systems behind operator websites for years — lead routing, CRM hookups, content engines, search visibility. AI agents are the next layer, and they sit on top of all of that. We don't build the foundation models. We don't pretend to. What we do is the part that's actually hard for a small operator: picking the right agent for the job, wiring it into your stack, and being on call when something breaks.
We package this as a retainer service with three tiers, depending on how many roles you're hiring for and how custom the build needs to be.
If you're still reading.
The honest take: in 2026, a small operator can buy two or three meaningful hours back from their week with an AI agent. Not ten. Not a replacement for a hire. But real, recoverable hours — invoice work, inbox triage, follow-up that used to slip.
The operators who win are the ones who pick one role, wire it in, run it for 60 days, and only then add the next one. Not the ones who try to stand up six agents in a quarter and end up babysitting all of them.