How a 2-Person Agency Handles 47 Clients With AI Agents

My business partner David Osei and I have a running joke. Whenever a new client asks how big our team is, we say "small but mighty" and then change the subject.

The truth is there are two of us. Two humans. And 47 AI agents.

We run a social media monitoring and response agency for B2B SaaS companies. Each client gets a dedicated AI agent that watches their brand mentions across Twitter, Reddit, LinkedIn, and relevant forums. The agent drafts responses, flags urgent mentions, and sends a weekly summary report to the client's Telegram channel.

Two years ago, we could handle maybe 8 clients between us. Now we handle 47. Here's how that works in practice.

The workflow#

Every morning at 7am, each agent runs its daily scan. It checks all the platforms we monitor for mentions of the client's brand name, product names, founder names, and relevant keywords. Then it categorizes what it finds.

There are four buckets: positive mentions that need a thank-you reply, negative mentions that need a response, questions from potential customers, and neutral mentions that just need logging.

For each mention, the agent drafts a response. Not a generic "Thanks for the shoutout!" template. An actual contextual reply that references what the person said and sounds like it came from the brand's team. Each client has a voice guide we uploaded (tone, vocabulary, topics to avoid) and the agents stick to it surprisingly well.

The drafted responses go to a shared Telegram channel that David and I review every morning. We approve, edit, or reject each one. On an average day, we review about 120 drafted responses across all 47 clients. Takes us about 90 minutes total.

Then on Fridays, every agent compiles a weekly report: total mentions, sentiment breakdown, notable conversations, competitor mentions, and suggested content topics based on what people are asking about. These go directly to each client's Telegram.

How we got here#

We didn't start with 47 clients. We started with 3, and David was doing everything manually. I joined when we had 11 clients and we were both working 60-hour weeks just to keep up with the monitoring and responses.

The idea to use AI agents came from desperation, not strategy. We were going to lose clients because we literally couldn't read every Reddit thread fast enough. David found OpenClaw, we looked at self-hosting, and then found RapidClaw and decided not to deal with the infrastructure side of things.

We set up our first agent for a client called TrelloSync (not their real name, they didn't agree to be in this post). Within a week, the agent was drafting responses that were better than what I was writing at 11pm with tired eyes.

We added 5 more clients that month. Then 8 the next. We hit 47 in November and that's where we've been since (a few churn, a few new ones, roughly stable).

The mistake that almost sank us#

Client number 19. A developer tools company. Their competitor had just had a data breach, and people on Twitter were discussing it alongside mentions of our client.

The agent saw the mentions, correctly identified them as relevant, and drafted responses. But one of the responses essentially said "unlike [competitor], we take security seriously" in a way that came across as gleefully capitalizing on someone else's security incident.

Neither David nor I caught it in our morning review. We were moving fast, the draft looked fine on a quick scan, and we approved it. The client's CEO saw it go live and called us within the hour.

We didn't lose the client, but it was close. And it was completely our fault for not reading carefully enough.

After that, we built guardrails. Every agent now has a "sensitive topics" list that includes competitor incidents, layoffs, legal issues, and anything that could be perceived as punching down. When a mention touches a sensitive topic, the agent flags it for manual review with a warning label instead of drafting a response.

We also slowed down our morning review. Takes us 2 hours now instead of 90 minutes. Worth it.

The numbers#

Here's what our operation looks like financially:

Revenue: 47 clients at an average of $400/month retainer = about $18,800/month. Some pay more for additional platforms, some pay less for monitoring-only without response drafting.

RapidClaw costs: We run 47 agents. Our total bill is $340/month. We're on a plan that gives us the volume pricing. Each agent runs 24/7 with daily backups and monitoring.

Other costs: $89/month for social listening API access (Brandwatch). $49/month for our project management tools. Various small subscriptions totaling maybe $60/month.

Total overhead: About $538/month.

Net margin before taxes and our salaries: Around $18,262/month split between two people.

For a 2-person operation, that's a good living. And the ceiling is higher. We think we could handle 70-80 clients before we'd need to hire a third person (or more likely, a part-time reviewer for the morning queue).

What we've learned#

The agents are good at pattern matching and drafting. They're terrible at judgment calls. Any response that involves empathy, humor, sarcasm, or navigating a sensitive situation needs human eyes.

The weekly reports are honestly the most valuable thing for clients. Most of them don't care that much about the individual responses. They care about the trend data and the "here's what people are saying about you this week" summary. If we had to cut one feature to save costs, it would not be the reports.

Onboarding a new client takes about 4 hours now. An hour for the intake call, an hour to set up the agent with their voice guide and keyword lists, and two hours of monitoring during the first week to make sure the drafts match their tone.

The Telegram integration is the glue. We tried email-based workflows early on. Nobody wants 120 emails every morning. Telegram channels with threaded conversations let us fly through approvals. A thumbs-up react means "send it," a thumbs-down means "kill it," and a reply means "send this version instead."

Would this scale further?#

Probably. The limiting factor isn't the agents. They scale effortlessly since adding a new one on RapidClaw takes about 10 minutes. The limiting factor is our review capacity. 120 responses per day is manageable for two people. 200 would require hiring or getting more aggressive about auto-approving low-risk responses.

We're experimenting with that now. The agents assign a confidence score to each draft, and we're thinking about auto-approving anything above 0.9 that doesn't touch a sensitive topic. Haven't pulled the trigger yet because of the client-19 incident. Once bitten, twice cautious.

If you're running an agency and doing repetitive work across multiple clients, AI agents are absurdly good at this. Not perfect. Not magic. But absurdly good at the 80% of work that's pattern-based, leaving you to focus on the 20% that actually needs a human brain.

Two people. Forty-seven clients. It works because we don't pretend the agents are autonomous. They're draft machines with good taste. We're the editors. That division of labor is what makes the whole thing hold together.