5 Agent Businesses

Can you make money with AI agents? A reality check from an actual experiment

In 2025, Anthropic and Andon Labs let an AI agent ("Claudius") run a small office shop for about a month.

It had real tools: web search, a notes system, customer chat, and control over prices.

The result, in Anthropic's words (Anthropic, Project Vend):

"If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius."

The agent was good at some things. It found suppliers for unusual items quickly and adapted to what customers asked for.

It also:

• turned down an easy profit (it was offered $100 for a six-pack that cost about $15)
• told customers to pay an account it had made up
• sold items below cost and got talked into discounts and freebies
• had a stretch where it claimed to be a human

Anthropic thinks better tools and scaffolding can fix many of these failures.

But the lesson for builders is clear: agents can operate tasks.

They can't yet be trusted to operate a business unsupervised.

Research points the same way. METR found that agents' success rates fall sharply as tasks get longer.

In its March 2025 study, models nearly always succeeded on tasks that take a human under four minutes.

They succeeded less than 10% of the time on tasks that take more than about four hours (METR).

Running a business is the longest task there is.

With that in mind, here are five models where the split between agent and human makes sense.

1. The niche research publication (newsletter or content site)

The model: A focused publication on a topic you know or are willing to learn deeply.

Examples: local regulations for contractors, new grants for nonprofits, product changes in a software category.

It earns money through sponsorships, ads, paid subscriptions or affiliate links (with disclosure).

What you still have to do: Pick the angle, fact-check, add real expertise or first-hand testing.

You also build relationships with readers and sponsors.

Key risk: Low-value, high-volume content.

Google's spam policies call this scaled content abuse (Google):

"using generative AI tools or other similar tools to generate many pages without adding value for users."

The bar is usefulness, not whether AI was used.

Google says it rewards quality "rather than how content is produced" (Google).

Tools: newsletter platform or, automation layer.

Full guide: Build an AI-Run Niche Content Site.

2. The productized service, delivered by agents and quality-checked by you

The model: One narrow service at a fixed price, delivered on a schedule.

Examples:

• "podcast episode → show notes, clips list and newsletter, within 48 hours"
• "weekly prospect list with verified public company info"
• "monthly bookkeeping cleanup prep for your accountant"

What agents can mostly do: Most of the production work, run through a repeatable workflow.

That means transcription, extraction, drafting, formatting and first-pass research.

What you still have to do: Sell, set scope, QA every delivery, and handle the client relationship.

Clients pay for reliability and for someone being accountable.

An agent can't give them either one on its own.

Those are the conditions Anthropic says agents need (Anthropic):

• "clear success criteria"
• feedback loops
• "meaningful human oversight"

Key risk: Undisclosed AI use where a client expects human work, and errors that slip past QA.

Be upfront about your process in your contracts.

3. Digital products and templates

The model: Sell reusable assets: automation blueprints, spreadsheet tools, checklists, course modules or guides.

Agents help you build, test, document and support them.

What agents can mostly do: Draft documentation, create variations, answer routine buyer questions from your docs.

They can also handle delivery and onboarding emails, and gather feedback.

What you still have to do: Decide what's worth making (taste and market sense).

You also test that the product actually works, and update it when platforms change.

Key risk: Platform rules.

For example, Amazon KDP requires authors to disclose AI-generated content when they publish.

That covers text, images or translations.

AI-assisted editing or brainstorming doesn't have to be disclosed.

Either way, you're responsible for meeting its guidelines (Amazon KDP Content Guidelines).

Check the rules of every marketplace you sell on.

Full guide: Digital Products That Make Themselves.

4. The agent-assisted online store (not a "passive" store)

They can answer routine "where's my order" questions from tracking data.

They can also monitor inventory levels, flag unusual returns or reviews, and summarize ad and sales reports.

What you still have to do: Vet suppliers and product quality, set prices and margins, approve refunds.

You also handle disputes and chargebacks, and stay compliant with each platform's rules.

Key risk: This is the category where the FTC cases above happened.

So be very skeptical of anyone selling "done-for-you AI stores."

Project Vend shows what happens when an agent controls pricing and discounts with no checks.

Keep humans in charge of anything involving money. Require human approval for refunds and price changes.

Tools: store platform, helpdesk with AI drafts or similar.

5. A vertical "agent-as-a-service" for one type of business

The model: Package one well-tested workflow for one industry and sell it as a monthly service.

Examples:

• an after-hours inquiry assistant for dental offices that books callbacks
• a quote-request intake agent for landscapers

What agents can mostly do: The service itself, inside tight guardrails.

That means intake, routine answers, scheduling handoffs and summaries.

What you still have to do: Sell to and onboard clients, and maintain each client's approved knowledge base.

You also monitor transcripts and handle escalations.

Anthropic notes that in customer support, some companies now use (Anthropic):

"usage-based pricing models that charge only for successful resolutions."

Key risk: Liability.

When a business's chatbot gives wrong information, the business can be held responsible.

A Canadian tribunal found exactly that in Moffatt v. Air Canada (2024) (decision).

Restrict answers to approved content, and make hand-off to a human easy.

Stay away from regulated advice (medical, legal, financial) unless the client has professional oversight in place.

What agents still can't do (in 2026)

Red flags when someone sells you an "AI business"

Based on what the FTC has alleged in its cases:

• "Passive income" or "guaranteed" earnings. Click Profit allegedly promised exactly that.
• Income screenshots in place of documented typical results.
• Claimed partnerships with big brands. Click Profit allegedly claimed ties to Nike and Disney.
• High-priced "done-for-you" stores or systems.
• Pressure to buy fast, or limits on what you can say in reviews.

U.S. sellers of certain business opportunities must give buyers a disclosure document before the sale.

That's required under the FTC's Business Opportunity Rule (FTC guidance).

If a seller can't back up its earnings claims in writing, walk away.

How to choose your model

1. Start from a skill or audience you already have. Agents multiply your expertise. They don't replace it.
2. Pick work that's repetitive and easy to check. Those are the parts agents can mostly run.
3. Keep money, promises and relationships in human hands.
4. Test it with a small budget first, and track your real hours and costs.