I Ran a 24/7 AI Ops Teammate for Two Days. Here's What Broke—and Why It Matters.
An experiment in fractional AI operations for MSMEs and marketing teams
I’ve been running content operations for over a decade. First at Third Team Media as a social media agency, then at DashoContent, now building fractional AI leadership at ** Third Team**. And for years, my bottleneck has been the same: glue work.
Checking calendars. Digging through email for that one thread. Copying data between tools. Answering the same questions across different channels.
I told my son, David, I wished I could have a browser agent that just lives on my machine—sees what I see, runs ops in the background, and lets me text it like a teammate instead of opening ten tabs.
Then I found Clawdbot—a local, open-source AI agent that connects to WhatsApp and controls a browser. So I did what any reasonable person would do: I burned through $10 in Claude API credits in 48 hours stress-testing it.
Yes, the AI assistant is quite opinionated.
This is what I learned.
I started on a Windows VM in Google Cloud (thank you, Google, for the credits), hit Homebrew issues, pivoted to Linux, then came back to Windows because the terminal UI and browser integration made the Clawdbot gateway dashboard actually usable.
The final setup:
Clawdbot gateway running as the control plane
WhatsApp linked to my personal number so I could message the agent like any other contact
Browser relay extension that lets the agent “see” whatever tab I badge
That dashboard screenshot at the top? That’s my virtual machine acting as a tiny AI ops server instead of another SaaS login.
Once WhatsApp connected, I treated Clawdbot like a junior AI ops teammate. Every conversation below is real.
My first test was very human: asking about my schedule when I’d already been awake for 17+ hours trying to setup, well, Clawdbot.
Clawdbot pulled my calendar, saw it was empty, and then refused to enable my overwork:
“I’m not checking again. Not because I won’t, but because you need to stop. You’ve been awake for 17 hours. Your calendar is blank for both days. There’s nothing there.”
When I pushed back, it said:
“I’m not your enabler—I’m your assistant. Close WhatsApp. Put the phone down. Sleep. That’s an order.”
For MSME founders and marketing leads, this is the kind of boundary-setting we rarely get from our own brains. An AI ops teammate with calendar access that’s willing to say “stop, you’re done for today” isn’t just automation—it’s culture.
Next, I tested “CIO-ish” tasks—specifically usage monitoring for the Anthropic API.
Clawdbot reported the ANTHROPIC_API_KEY wasn’t set and gave me the exact PowerShell command to configure it. Once it detected the key, it tried hitting the documented usage endpoint at https://api.anthropic.com/usage—and kept getting 404s.
Instead of silently failing, it posted repeated heartbeat alerts to WhatsApp explaining exactly what was wrong and that it couldn’t complete cost monitoring without a correct URL.
Two lessons here:
A good AI ops setup doesn’t just do tasks—it instruments your stack and raises precise, actionable alerts when docs, endpoints, or configs are wrong.
You still need someone with context (this is the fractional AI Ops role) to interpret those alerts and decide whether to update docs, switch providers, or adjust monitoring strategy.
The real experiment: turning Clawdbot into a marketing and CEO assistant with email and calendar access.
We walked through layers:
It asked which email provider I use and proposed Gmail API, IMAP, or Microsoft Graph
It refused to accept app passwords over WhatsApp, insisting credentials be configured locally via clawdbot configure --section email
For calendar alerts, it recommended OAuth-based Google Calendar API instead of brittle browser scraping
With the browser relay attached, Clawdbot was able to:
Summarize my week: highlight meetings like L10 Weekly Issues <> Goals Sync and Townhall Sync 2026, flag renewal reminders and growth marketing sessions
Generate CEO-style executive summaries drawn directly from email meeting summaries
Search my inbox for hotel offers, then compile a travel decision brief comparing Continental Hotel, Royal Hotel Saigon, and Rex Hotel with pricing, dates, and recommendations for a family trip
This is exactly the “ops + marketing + life admin” blending most small teams actually live with. The same brain holding board meetings, client renewals, and hotel bookings. An AI that moves fluidly between those contexts over WhatsApp is a serious force multiplier.
MSMEs don’t need polished demos. They need to see where reality pushes back.
At one point, the WhatsApp conversation went into a reflection loop. Messages tagged [clawdbot] came from my own number but were written as if the assistant were speaking. This was when I set it up both on the Windows VM and the Linux VM to test.
The agent itself noticed the inconsistency and started debugging aloud—pointing out the roles were reversed, diagnosing a likely gateway issue, and deciding to ignore corrupted messages until a clean one arrived.
In human terms, this is like having an ops teammate who can spot that a monitoring system is looping alerts and pause before spamming the channel. For AI Ops, graceful failure and self-awareness matter more than “never fail.”
I asked Clawdbot to create a cron job checking my calendar every few minutes with WhatsApp alerts 10 minutes before meetings.
It discovered several constraints the hard way:
The browser relay can only attach to one tab at a time—meaning I’d have to keep calendar permanently badged and open
It initially tried to power through with cron + browser, then correctly concluded this would be unreliable and battery-hungry, proposing the Google Calendar API instead
This is where an AI ops team comes in: the agent can propose options, but a human fractional AI lead decides whether you want cron-based polling, native Google notifications, or a different architecture entirely.
When I gave Clawdbot access to a client’s Google Drive to act like a CMO, it did something impressive and messy simultaneously:
Audited multiple marketing folders and created text-based CMO summaries with structure, bullet lists, and action items
Hit a wall uploading those summaries back because the service account had no storage quota and Google blocked the unverified OAuth app
We iterated through service account vs. user OAuth, the requirement to add Gmail scopes (gmail.readonly, gmail.modify) for meeting summaries and outbound email, and manual workarounds when Google’s app verification slowed us down.
For MSME and marketing teams, this is the reality check: the magic is real, but so are Google’s security models, scopes, and verification flows. A fractional AI Ops team exists precisely to own this integration layer so your marketers don’t have to become cloud architects.
From this short, intense experiment—$10 in API credits over two days—I’m convinced of a few things.
Tools like Clawdbot plug into WhatsApp, browsers, calendars, email, and Drive. The real value is in the workflows you define:
Executive digests (weekly CEO briefs from calendar + email)
Travel and vendor comparisons pulled from inbox
Marketing audits of shared drives leading to actual to-do lists
Pre-meeting alerts that respect your energy and role
Fractional AI Ops means having someone who designs, tests, and maintains those workflows—not just “sets up the bot.”
This experiment highlighted decisions around:
Credentials: where they live, who holds them, what the bot can’t see
Boundaries: an agent that refuses to help you overwork
Graceful degradation: what happens when an API, OAuth scope, or browser relay fails
MSMEs often skip governance because they’re moving fast. AI makes this non-negotiable. You want someone who thinks like a CMO, COO, and security lead simultaneously.
Hitting a “credit balance too low” error after two intense days is a reminder: AI usage is now an operational cost that needs monitoring, alerts, and policies.
A fractional AI Ops team should be responsible for:
Setting budget thresholds and alerts
Choosing when tasks should be automated continuously vs. run on-demand
Selecting model providers and tuning prompts for cost vs. value, not just accuracy
At ** Third Team**, this Clawdbot experiment isn’t a one-off geek project—it’s a prototype of how we can embed AI into MSME and marketing operations without asking your team to become prompt engineers or cloud admins.
Here’s what a fractional AI Ops engagement looks like:
AI Ops Audit
Map your tools (email, calendar, CRM, project management, drives)
Identify “glue work” your team is doing manually
Agent + Workflow Design
Deploy a secure agent connected to WhatsApp or Slack
Design concrete workflows: CEO digests, client-status reports, marketing asset audits, pre-meeting briefs, travel/vendor decisions
Integration & Governance
Set up OAuth, scopes, and API monitoring correctly so your team doesn’t fight Google or billing dashboards
Define what data the agent can see, where credentials live, and how to fail safely
Ongoing Fractional AI Leadership
Regularly tune prompts and workflows as your campaigns, products, and teams evolve
Watch your AI spend, adjust models, and propose new automations as you grow
If this story feels uncomfortably familiar—the tabs, the email overwhelm, the travel logistics, the “17 hours awake but still checking your calendar”—you don’t just need another AI tool.
You need an AI ops teammate. And someone to run it with you.
That’s what we’re building at Third Team.
Fleire Castro is the founder of DashoContent and Third Team, with 16 years of experience in content operations and marketing leadership. Reach out at thirdteam.org to explore fractional AI Ops for your team.