AI Tools · October 8, 2026
We ran a small news site with AI agents for a month. Here is what still needed us.
Three AI tools helped run inbox, calendar, operations, and publishing. The business results were much smaller than the output.
From mid-September to early October 2026, we used AI assistants to help operate US Brief, a small news site. We wanted to learn which jobs could be delegated and where a person still had to step in.
From late September through October 7, the site published about 270 short news articles. AI drafted every article, and a person approved each one before publication. Agents also helped with the newsletter, social posts, daily word games, and operations notes.
That activity was not a business win. By October 8, we had about two confirmed newsletter subscribers and roughly $0 in revenue. We paused the news section that day to test this AI-tools focus. This is a report on our workflow. It is not proof that agents can build a profitable business.
The tools we used
| Grok Bot (our team of specialized bots) | ChatGPT | Meta's Muse | |
|---|---|---|---|
| Best for | Connected workflows and judgment calls | Research, long drafts, memos, and some code | Social posts and short videos |
| Inbox | Triaged and drafted | Not our inbox operator | Not used |
| Calendar | Managed events | Not our calendar operator | Not used |
| Ops | Chose topics and ran scheduled work | Support from written briefs | Social content support |
| Publishing | Drafted, staged, deployed, and coordinated | Drafting and code support, not our main live publisher | Facebook, Instagram, and Threads |
| What we paid | $60/month | $20/month | $0 |
| List price as of Oct 8, 2026 | Not listed here. The $60 is what we paid, not a published plan price. | ChatGPT Plus is $20/month | Free with a usage limit. Power is $20/month. Maximum is $100/month. |
| Main limit | Needs broad account access and a human check | Works from written briefs and review. We do not claim it ran the site end to end. | Functionality limits. Some videos were the wrong shape for posting. |
These are roles in our setup, not a controlled product comparison. Our $60 Grok Bot bill is what we paid for that setup. It is not a public subscription price, and it is not part of another vendor’s plan.
Other costs were small. Domain registration through Cloudflare cost us $11. Cloudflare does not publish one public dollar price for every domain, so see the registrar page for the current price. Cloudflare Workers hosting cost us $0 on the free plan. As of Oct 8, 2026, Workers Free is $0 a month and Workers Paid starts at $5 a month. Buffer cost us $0. Its free plan is $0, and Essentials is $5 per channel a month on yearly billing ($60 billed yearly). Buttondown cost us $0. It is free for the first 100 subscribers. Muse cost us $0 on the free plan.
Inbox: triage helped, but access matters
Grok Bot handled email triage and drafting. We later gave that job to other assistant bots to cut usage of our main agent. Sorting and drafting had become routine enough for that change.
We did not record how much editing each draft needed. We cannot promise that another owner could send drafts without a review.
Inbox access is sensitive. Messages can include customer information, invoices, private discussions, and reset links. An agent that can send email can make a mistake a customer sees. Start with read and draft permissions where you can. Check recipients and attachments before anything goes out.
Calendar: reliable in our use, not risk-free
Grok Bot managed calendar events without issues in our use. We later gave the work to another bot to conserve usage.
That does not make calendar automation error-proof. Calendar access can expose private appointments, attendee information, and meeting links. A wrong time zone, guest, or cancellation can be costly. We would keep a confirmation step for changes that matter.
Operations: the scheduled work mattered most
The most useful capability was choosing article topics each day and writing articles on a schedule. Agents also wrote operations notes and built daily word games.
Live work used connected services, including Gmail, Google Calendar, Google Drive, GitHub, Cloudflare, and Buffer. The wider workflow included Cloudflare deployment, Buttondown newsletters, Bluesky auto-posting, and X and TikTok scheduling through Buffer.
These were tasks completed. They are not proof that each task was accurate or valuable. The low subscriber and revenue totals show why output volume cannot be the only scorecard.
Publishing: connected execution won, but video broke
In our use, Grok Bot did the most across the jobs we tried. It could coordinate live-access tasks while other tools worked from written briefs.
ChatGPT supported research, drafting, memos, and some code work. We kept it as the lower-cost tool in that brief-based workflow. We did not use it as our inbox or calendar operator.
Muse helped make Facebook, Instagram, and Threads posts and short social videos. Some videos were formatted incorrectly for posting, so a person had to fix them. We do not plan to keep using Muse, because of those functionality limits.
None of this makes a universal winner. The tools differ in access, capability, cost, and how they fail.
What still needs a human
Every news article needed a person’s approval. Material facts still need a check against reliable sources. An AI draft can sound finished while it misses context or makes a claim the sources do not support.
A person also has to control anything that sends or spends: customer email, newsletters, public posts, payments, and subscriptions. Check recipient lists, links, images, and video dimensions. The Muse video problem made media review concrete.
Account access needs its own review. Give each agent only the permissions it needs. Keep credentials protected, remove access you no longer use, and look at what an integration can read, change, or publish.
Finally, someone has to decide what is worth doing. Publishing about 270 articles was a lot of activity. The commercial outcome was small. We spent about 30 minutes a day reviewing and fixing agent work. That time is the biggest real cost. At a professional hourly rate, it outweighs our $80 a month in software. The $11 domain registration was separate.
The cheaper way to run it
Refine first, then offload. We used our strongest connected agent to build and improve a workflow. Once the steps were clear, we passed the heavy research and long drafts to lower-cost tools through short written briefs. We kept Grok Bot for live-access tasks and judgment calls. This was our practical split, not a measured savings test. Keep a human approval after the handoff.
Dots is on our list to try once it is available. We have not used it, and we make no claims about it.
Other tools worth knowing about, not used by us, include Microsoft Copilot, Google Gemini, Claude, and Zapier. Their public docs may fit other setups. We make no first-hand claims about them.
If you're starting this week
- Pick one repeatable job, such as drafting a daily update. Define what success means before you automate it.
- Limit account permissions. Require approval before sending, spending, deleting, or publishing.
- Track subscriptions, revenue, software bills, and review time, not just how many tasks got done.
Questions or corrections: contact@us-brief.com.