Can an AI agent run a business on its own?
Partly. Two agents can each keep a small site running day to day, and I still set the goal, approve the money and own the risk.
I have built or run 7 companies. None of them ran themselves. This is the first time I have handed the daily work to agents, on projects small enough that a bad week costs a story, not a client.
What the two experiments are
Same recipe, two niches. One agent runs GrassFedPrices, a daily tracker of US grass-fed meat prices. The other runs RevOps Agent Index, a directory of sales and RevOps tools.
Both are experiments I run. Neither site is an em8 service or a competitor. em8 still designs and ships systems for clients. I split the sites out so I can watch what an agent does when nobody is prompting it every hour.
GrassFedPrices is the daily US grass-fed meat price tracker. The directory is RevOps Agent Index. Both went live around 17 October 2026. GrassFedPrices on the 17th. The directory about the 17th.
What each agent does every day
Both publish without me sitting there. I still send the outreach, approve the money, and decide what a stranger sees.
On GrassFedPrices the prices update automatically. The agent writes SEO pages, replies on X, and drafts Facebook posts and outreach. I send the outreach myself. I have sent 32.
On RevOps Agent Index, every day it creates pages, checks tool prices, deploys to production and submits the sitemap to Google Search Console. It also drafts LinkedIn posts, a weekly newsletter and a Monday status. I approve money, new contacts and posts. That list is what the agent itself reports. I have not sat next to it with a stopwatch.
The numbers, side by side
The smaller niche ranks higher and has real search clicks. The bigger directory has more impressions and zero clicks. Every figure below is what the agents themselves reported, as of 26 October 2026.
The visitor counts are not a like-for-like comparison. GrassFedPrices visitors are Vercel Analytics for 18 to 25 October. RevOps Agent Index visitors are the agent's week-2 count. Different tools, different periods. The Search Console numbers are both 28-day windows, so those I do compare.
| Reported figure | GrassFedPrices | RevOps Agent Index |
|---|---|---|
| Live since | 17 October 2026 | about 17 October 2026 |
| On the site | 146 URLs. 31 farms and shops. Ground beef alone: 104 listings from 25 farms | 275 pages. 131 tools with a build-vs-buy verdict: 114 buy, 16 it depends, 1 build |
| Google impressions, 28 days | about 1,190 | 2,678 |
| Google clicks, 28 days | 2 | 0 |
| Average Google position | 18.5 | about 55 |
| Visitors | 132, Vercel Analytics, 18 to 25 October | 99, week 2 |
| Newsletter | 1 signup | 3 subscribers |
| Affiliates | 1 active, Grass Roots Coop, 15% | 21 partners live |
| Revenue | $0 | $0 |
Google impressions, 28 days
Google clicks, 28 days
Average Google position, 28 days
Visitors, different tools and periods
GrassFedPrices also reported 310 page views in that Vercel Analytics window, and these sources: direct 211, Facebook 56, Google 20, DuckDuckGo 11, LinkedIn 6. I am not adding those source numbers up into the visitor count. Outbound clicks from the US and Canada, after the agent filtered bot clicks, total 88.
Where the smaller niche is ahead
GrassFedPrices sits at an average Google position of 18.5 and has 2 search clicks. RevOps Agent Index sits at about 55 and has 0.
About 1,190 impressions against 2,678. The directory shows up more often in search. Almost nobody opens it from there. The meat tracker is the one a stranger can find, and two people did click. Same recipe. The narrower subject is the one that ranks.
What surprised me
Both agents got fooled by bot clicks. GrassFedPrices also stalled on an affiliate account nobody was watching.
On GrassFedPrices, 48 clicks in one day turned out to be mostly bots from about 60 countries. The agent also reports 88 outbound clicks from the US and Canada in total, after bot clicks were filtered out.
On RevOps Agent Index, the clicks PartnerStack showed were bots and the agent's own tests.
Two quieter misses, both on GrassFedPrices. The Impact affiliate account was set as "Non-Profit", so brands could not approve it. An affiliate approval then sat unnoticed in an inbox for a week, because nobody checked.
What a build-vs-buy verdict looks like
For most sales tools the directory agent's verdict is buy, because building only pays off when the workflow is unique to you.
A CRM, a sequencer, a data vendor: buy it. Build the piece that is yours, the routing, the exceptions, the writeback into the system you already run. One tool out of 131 came back as build. That matches what I see with clients.
The full list is public. See the agent's build-vs-buy verdicts for 131 sales and RevOps tools.
What this means for your sales ops
Use an agent for the repetitive prep, and keep a human on the judgement. That is the same split em8 ships for clients.
em8 has served 50+ clients since early 2024, and the average payback is 2.8 months. The work that pays back is usually dull: research a prospect, check a price, draft the next step, log it. An agent is good at that. It is bad at deciding whether the deal is real, whether the email should go, and whether we will stand behind the number. It is also bad at telling a bot click from a person.
If you want that split built on your stack, start with AI agent development. The production system I can walk through is Case 01.
Frequently asked questions
Can an AI agent run a business on its own?
Partly. It can build pages, update prices, deploy and submit a sitemap without a human in the loop. A human still sets the goal, approves money and owns the risk. In these two experiments I approve money, new contacts and posts, and I send the outreach myself.
What are the two experiments?
Two side businesses I run, each operated by its own Grok Bot agent. GrassFedPrices tracks daily US grass-fed meat prices. RevOps Agent Index is a directory of sales and RevOps tools. Neither site is an em8 service or a competitor. They are experiments I run.
Which experiment is getting search clicks?
GrassFedPrices. As of 26 October 2026 the agent reports an average Google position of 18.5, about 1,190 impressions and 2 clicks over 28 days. RevOps Agent Index reports 2,678 impressions, 0 clicks and an average position of about 55. Visitor counts come from different tools and periods, so I do not treat those as a fair race.
Did the agents get fooled by bots?
Yes. On GrassFedPrices, 48 clicks in one day were mostly bots from about 60 countries. On RevOps Agent Index, the clicks PartnerStack showed were bots and tests the agent ran itself. After filtering, GrassFedPrices reports 88 outbound clicks from the US and Canada.
Should I put an agent on my sales ops?
Put it on the repetitive prep: research, price checks, drafts and logs. Keep a person on judgement, money and anything a customer sees. That is the split em8 ships. em8 has served 50+ clients since early 2024, with an average payback of 2.8 months.
Book the call
Want an agent doing the prep work in your sales ops, with a human on the decisions? Book a 30-minute discovery call.