AI Can Now Log In and Do Your Marketing. Here's What Changes.

This week I set up conversion tracking on two websites and built two Google Ads search campaigns. I didn't do most of the clicking. An AI agent did, inside my own browser, while I made the calls and checked its work.

That's the shift nobody is explaining to small business owners yet. AI agents for marketing don't just write copy anymore. They can open the tools you already pay for and do the work a marketing coordinator would do. Here's what that actually means for you, and what they still get wrong.

What are AI agents for marketing, exactly?

A chatbot answers. An agent acts. That's the whole difference.

When you use ChatGPT or Claude as a chatbot, you ask for something, it writes something, and you copy it somewhere. When you use one as an agent, it opens a browser, goes to the website, and clicks, types and fills in forms itself. You watch it work, step in when it needs you, and approve anything that matters.

Two versions are worth knowing about right now:

• Claude in Chrome. Anthropic's browser extension lets Claude read, click and navigate websites in your own Chrome browser. Anthropic's help article says it's available on all paid plans (Pro, Max, Team and Enterprise) as of this writing, and it says plainly that it "is still risky." I appreciate the honesty.
• ChatGPT Work. OpenAI retired its original ChatGPT agent mode, and its Atlas browser has been retired too. The browser work now runs through ChatGPT Work, which uses a cloud browser that can read pages, click buttons and fill in forms. OpenAI's help page lists it on paid plans, excluding Free and Go, as of this writing.

Product names in this space change every few months. OpenAI's first agent mode has already come and gone. Don't get attached to a brand. What matters is the capability: software that can operate other software.

What can an agent actually do inside your marketing tools?

Here's how I'd sum it up for the tools most small businesses use, along with the part that still has to come from a person.

Google Ads and Tag Manager are the two I ran end to end this week. I wrote up both: conversion tracking with AI and building a local campaign with AI. The rest of the table is the same kind of work: log in, click through screens, fill in fields, save.

The agent is slower than a skilled person on some of it, especially drag-and-drop editors. That matters less than you'd think. It doesn't get bored, and I'm doing something else while it works.

What changes when production gets cheap?

For twenty years, the expensive part of marketing was production. Somebody had to write the posts, build the campaigns, set up the tags and click through every screen. You paid for those hours, whether you hired a person or an agency.

That part is getting cheap and fast. The two campaigns I built this week took an afternoon instead of days. The bottleneck moved.

Think about what that means in practice. An agent can build a Google Ads campaign in an hour. It can't tell you whether you should be running Google Ads at all, whether anyone answers the phone when the ads work, or whether $10 a day buys anything in your market. Those are judgment calls, and they're most of what makes marketing work.

That's also why the old agency model is in trouble. It was built to sell production hours through layers of account managers. When production costs almost nothing, the layers have nothing left to sell.

What do AI agents still get wrong?

Plenty. Anyone telling you these tools run your marketing on autopilot hasn't watched one work for a full afternoon.

• Misclicks. During my campaign build, the agent clicked the same checkbox twice, which left a setting off that was supposed to be on. It reported the step as done. I caught it in review.
• Logins and security checks. Agents get stuck on CAPTCHAs, two-step verification codes and "is this really you?" screens. You have to step in.
• Confident wrong answers. An agent will sometimes say a task is finished when it isn't, or describe a setting it never actually changed.
• Drift. Over a long task it can wander from what you asked, like adding keywords you never approved because they "seemed relevant."
• Getting tricked. Hidden text on a web page can try to steer an agent into doing something you didn't ask for. This is called prompt injection. OpenAI's own agent help page says its safeguards "don't eliminate all risks."

None of this makes agents useless. It makes them like a fast new hire in their first week: capable, quick, and in need of someone checking the work before it goes out the door.

How do you let an AI agent into your accounts safely?

This part is simple, and it's not optional.

• Sign in yourself. Never paste a password, API key or card number into an AI chat. You log in, and the agent works in the browser session you already opened.
• Use a separate browser profile. Keep your personal email, banking and anything unrelated out of the window the agent can see.
• Give people their own access. Add your marketer as a user in Google Ads, Business Profile and Meta with their own login, instead of sharing the owner account. You can remove them anytime.
• Require approval for anything final. Anything that sends, publishes, pays or deletes waits for a human. Say it in every instruction.
• Watch the first few runs. You learn fast where it gets confused.
• Check the result, not the report. "Done" from an agent means it thinks it's done. Open the account and look.

Should you run the agents yourself?

You can. The tools are on regular paid plans, and nothing above requires a developer. But be honest about what you're signing up for.

Running agents well means knowing what Google Ads, Tag Manager and your Business Profile should look like when they're set up right. If you don't know that, you can't catch the agent's mistakes, and you pay for them in wasted ad spend or broken tracking. For most owners, an hour spent running the business is worth more than an hour spent supervising software.

What a week looks like when agents do the production

Here's how I run it for clients now.

Monday: a short meeting with the owner. What's happening in the business, what's working, what moves up the list.

During the week: I brief the agents task by task, with the right access and a clear picture of what done looks like. They handle the production: campaign changes, posts, page updates, tracking.

Before anything goes live: I check it. Every piece. What's wrong goes back. What publishes, sends or spends waits for approval.

Next Monday: what got done, what the numbers say, and what's next.

That's the job now. Less typing and clicking. More deciding and checking.

What this means for your business

If you've been putting off marketing because you couldn't afford the production, that excuse is mostly gone. Production is cheap now.

What you need is someone who knows what to build, gives the agents the right instructions and access, and catches the mistakes before your customers see them. If you want to see what that looks like on your own accounts, the first month is $99.