For years, getting your software to cooperate meant integrations. Somebody wrote code that let QuickBooks talk to the CRM, or you paid for a connector, or you just had a person copy things over by hand.
That's no longer the only option. AI agents for small business can now open a web browser, sign into the tools you already pay for, and click through them the way your office coordinator does. They read the screen, type in the boxes, press the buttons.
I've been all-in on AI since ChatGPT launched in November 2022, and this is the biggest practical shift I've seen. Here's what changed, where it still falls short, and what I'd do about it if I owned a 10-person shop.
What changed: AI can now operate a browser
An "agent" is just an AI that takes actions instead of only answering questions. The new part is that the actions happen inside real software.
Two examples you can use today:
- Claude in Chrome. Anthropic's browser extension lets Claude >read, click and navigate websites, including filling out forms the way a person would. As of this writing it's available on Claude's paid plans and only works in Google Chrome.
- ChatGPT Work. OpenAI's agent mode for longer tasks can use a >cloud browser that clicks buttons and fills in forms on a separate computer. When it hits a login, you type your credentials into a secure form the model can't see.
Plans and names change fast, so check the current details before you buy anything.
This week I had an agent working inside Google Ads and Google Tag Manager on my own sites. It clicked through the same screens I would have. I made the decisions and checked the work. The clicking was not my job anymore.
Do you still need APIs to connect your software?
An API is a door one program leaves open so another program can send and receive data without a person. It's still the best way to move data that flows constantly.
But not every tool has a good API, and plenty of work is not "move this field to that field." It's "look at this, decide, then update three things." That's where an agent in the browser fits.
| Fixed workflow (API, n8n) | Agent in the browser | |
|---|---|---|
| Best for | Same steps, many times a day | Varied steps, weekly or one-off |
| Speed | Seconds | Minutes, sometimes slow |
| Reliability | Very high once built | Good, but needs review |
| Needs an API? | Yes | No, just a login |
| Judgment | Only what you program in | Can read and decide, then ask |
My rule of thumb: if it happens 50 times a day the exact same way, build a pipe. If it happens weekly, takes judgment, or lives in a tool with no API, an agent can probably do it while a person reviews.
What can an AI agent actually do in your existing tools?
Say you run a six-person commercial cleaning company. You already pay for QuickBooks Online, a CRM, Google Workspace and a website builder. Here is the kind of work an agent can take on in each, with you or a director approving the risky parts.
- Accounting software. Pull the accounts receivable aging report and summarize who is past 30 days. Compare this month's supply spending to last month's. Prep a list of invoices for a human to send.
- CRM. Find duplicate contacts and propose merges. Update deal stages from a list of signed proposals. Log call notes someone dictated into a voice memo.
- Google Workspace. Build the weekly crew schedule sheet from a template. File signed contracts into the right Drive folder. Draft replies to routine inbox questions for review.
- Website. Update holiday hours. Swap a photo. Draft a new service page and leave it unpublished.
- Vendor portals. Check prices on your top supplies across three supplier sites. Download monthly statements into one folder.
None of that is glamorous. That's the point. It's the clicking your best people do between the work that actually needs them.
Where AI agents still fall short
I'd be lying if I said this just works. Here's what I run into:
- Logins and two-factor codes. Agents get stuck at sign-in screens. The right answer is that you sign in, not the agent. Never paste a password into an AI chat.
- CAPTCHAs. Those "prove you're human" puzzles stop agents cold, and Anthropic's own safety guide says Claude is >prohibited from bypassing them.
- Misclicks. Agents sometimes click the wrong box or toggle something twice. Fine if someone checks. Expensive if nobody does.
- Speed. A person who knows QuickBooks cold is still faster on a quick task. The agent wins on the boring, repetitive stretch.
- Hidden instructions. A web page or email can contain text written to trick an agent. I cover how to handle that in what I never let an agent do without a human check.
- Anything risky. Sending to customers, paying, publishing and deleting should all wait for a human yes. Both Anthropic and OpenAI build in confirmation steps for high-impact actions, but you still own the result.
What should small business owners do about AI agents now?
You don't need to buy anything this week. You need a list.
- Write your clicking list. Every recurring task where someone logs into a tool and clicks through screens. Include who does it and roughly how long it takes.
- Sort it. High-volume and identical goes in the "pipe" column. Weekly, judgment-heavy or one-off goes in the "agent" column.
- Fix access first. Give the agent its own user seat with the lowest permissions that still get the job done. You sign in. You can turn it off any time.
- Start read-only. Reports, summaries and price checks first. Nothing that changes data until you trust the output.
- Decide who directs. Someone has to write the instructions, check every result and approve anything that goes out. That person should not be you if your time is better spent selling or running jobs.
If you want to try one yourself, start with the delegation prompt. It's the difference between an agent that wanders and one that finishes.
How I'd approach this
I treat agents as the production line and myself as the person running it. High-volume work gets a proper workflow in n8n. Everything else on your clicking list goes to an agent working in your tools, in priority order, with me checking every piece before it goes anywhere. You keep ownership of every account and file.
Frequently asked questions
Can an AI agent really use QuickBooks or my CRM?
Yes, through the browser. Tools like Claude in Chrome and ChatGPT's Work mode can read a page, click buttons and fill in forms on sites you are signed into. They are slower than a person who knows the software well and they make mistakes, so anything that sends, pays, publishes or deletes should wait for a human to approve it.
Do I still need integrations and APIs?
For high-volume, repeat work, yes. A fixed workflow that moves data through an API is faster, cheaper per run and more reliable than an agent clicking through screens. Agents fill the gaps: tools with no API, weekly tasks that need judgment, and one-off setup jobs nobody wants to pay a developer for.
Is it safe to give an AI agent my passwords?
Don't give it passwords at all. Sign in yourself, let the agent work in that signed-in browser, and approve anything risky. Better yet, give the agent its own user seat with limited permissions so it can only touch what the task needs.
Should I run the agents myself?
You can, and it's worth trying once so you understand it. But your hour is usually worth more running the business, and the learning curve and mistakes are real. Most owners are better off with one person who directs the agents, checks the output and reports back weekly.
Bring me your clicking list
Write down the ten tasks your team clicks through every week. That's the starting list. My agents work it in priority order, I check every piece, and we meet weekly. It's $499 a month flat, and the first month is $99.
