Marta del Sol on hitting $4K MRR with three AI agents

You have probably read a hundred "I built an AI agency" posts. Most of them are screenshots and vibes. Marta del Sol's is different because she let us see the dashboard, the Stripe export, and the months where the line went sideways.

Marta runs a one-person studio out of Valencia that does a single, unglamorous thing: it turns messy operations inside small e-commerce brands into automations that actually run. She charges a flat retainer. As of last month she crossed $4,120 in MRR across nine clients, with a software bill of $612/month and roughly 22 working hours a week.

This is the story of how she got there, told as told to Joaquín del Río, with every number she was willing to share — and the caveats she insisted we print next to them.

The number before the number

When Marta says "$4K MRR" she means recurring retainers billed monthly through Stripe. It does not include one-off setup fees, which added about $9,400 over the last twelve months but are lumpy and unpredictable.

"MRR is the number I trust because it's the number that shows up whether or not I hustle that week."

She is careful about this distinction because the setup fees flattered her early. In month three she had a $3,800 month and told her partner she'd "basically made it." Month four was $900. The setup work had dried up and the recurring base underneath it was tiny.

"That month taught me the only lesson that mattered," she says. "Recurring is the floor you stand on. Everything else is weather."

The three agents, specifically

People want the stack, so here it is, with what each one actually costs her.

The first agent handles inbox triage for clients: it reads support and ops email, classifies it, drafts replies, and flags the 5% that need a human. She runs it on a mid-tier model and it costs her about $140/month in API usage across all clients combined.

The second handles catalog and data hygiene — deduping products, fixing categories, reconciling inventory exports. This one is less glamorous and more valuable. Clients pay for it without ever seeing it run.

The third is what she calls the "narrator": every Friday it compiles what the other two did into a plain-language report each client actually reads. "The narrator is why they renew," Marta says. "It's not the automation. It's that they can see the automation working."

"Clients don't renew because the robot is smart. They renew because on Friday they understand what the robot did."

She stores client configuration, run logs, and the structured outputs in Totalum, which she picked because she wanted a database and an API without standing up infrastructure she'd have to babysit. "I'm one person. I cannot be a database administrator on Tuesdays," she says. That's the only stack opinion she'll defend in public.

The pricing she's embarrassed by, then not

Marta's first retainer was $180/month. She winces telling it. The client was a friend's candle business and she priced it at "what felt rude to charge a friend, minus a bit."

The arc of her pricing is the real story:

• Clients 1–3: ~$180–$250/month. Underpriced, learned the playbook.
• Clients 4–6: ~$400/month. Started saying no.
• Clients 7–9: $650–$750/month. Stopped explaining the price.

The jump from $250 to $650 didn't come from better automations. It came from a different sentence. She stopped selling "AI automation" and started selling "I will personally make sure your operations don't break this month." Same work. The second sentence is worth triple.

"The agents are the delivery mechanism," she says. "The thing they're buying is that they get to stop worrying."

The month it nearly ended

Month eight, two clients churned in the same week. One sold their business; one decided to hire in-house. Marta went from $3,600 to $2,300 in MRR in nine days and spent a weekend convinced the whole thing was a fluke.

What she did next is the part worth copying. She didn't go find two new clients. She went to her remaining seven and asked each one a single question: "What's the next thing that's annoying you that I haven't automated yet?" Five of them had an answer. Three of those became expansions worth $1,100/month combined.

"Churn sent me looking for new clients. The recovery came from the clients I already had."

By month ten she was back above $3,500, and the base was sturdier because it was spread across expansions, not just logos.

What 22 hours a week actually looks like

The hours number invites disbelief, so Marta walks through it. Roughly eight hours is client communication and the Friday narratives (reviewing what the narrator drafted before it goes out — she never sends unreviewed). About six hours is building and fixing automations. Another four hours is sales and onboarding. The remaining four hours is the unglamorous tax: billing, the occasional model breaking a prompt, an API deprecation.

"The agents don't save me the work," she says. "They save me the work scaling with the number of clients. Going from five to nine clients added maybe three hours a week, not fifteen."

That's the actual unlock, and it's the one the screenshot-posts never mention: the leverage isn't that AI does the job. It's that the marginal cost of the tenth client is close to the marginal cost of the sixth.

The honest caveats

Marta asked us to print these next to the numbers, so here they are.

The $612 software bill undercounts. It excludes her own salary, the laptop, and the two paid communities she's in. Fully loaded, her costs are closer to $1,100/month.

The 22 hours is a good month. A bad month — a client migration, a model that suddenly behaves differently — can be 35 hours, and those months don't pay extra.

And the $4K MRR is nine clients. Lose three at once and the picture changes fast. "I am one bad quarter from this being a part-time income again," she says. "I price like I know that now."

Where she's pointing next

The goal isn't ten clients. It's the same nine paying more. Marta is testing a $1,200/month tier that includes a quarterly in-person operations review, on the theory that the highest-leverage thing she sells is attention, not automation. If two clients take it, she crosses $5K MRR without onboarding anyone new.

"I used to think the dream was more clients," she says. "Now I think the dream is fewer clients who'd be genuinely sad if I quit."

The clients she said no to

For a story about getting to $4K MRR, Marta spends a surprising amount of it talking about revenue she turned away. In the last year she declined roughly $2,000/month of potential retainers, and she keeps a private list of why, because the pattern taught her more than her wins did.

The biggest category is what she calls "rescue jobs" — businesses whose operations are already on fire, who want her to come in and fix a crisis on a retainer. "Those clients are never happy, because the thing they're really buying is for the past to not have happened," she says. "I can't sell that."

The second category is clients who want her to manage humans. "The moment a retainer includes 'and coordinate with our VA in the Philippines,' it's not an automation engagement anymore, it's a management job, and management doesn't have leverage. Ten of those and I'm running an agency I never wanted."

"Every 'no' I said protected the business model. The 'yeses' just paid the rent."

She's candid that saying no was a privilege she earned. "In month two I said yes to everything because I had to. The ability to be picky is something MRR buys you. You don't start there." The advice she gives newer solo operators isn't "say no" — it's "notice which yeses you regret, and turn that list into rules."

What the agents get wrong

Because the AI-agency genre is allergic to admitting flaws, we pushed Marta on what breaks. Plenty, it turns out.

The triage agent has a persistent ~5% misclassification rate that she's never gotten below, no matter the prompt. "It's confidently wrong sometimes, and confident-wrong is more dangerous than uncertain-wrong," she says. Her entire review process exists because of this. Nothing the narrator drafts goes to a client without her eyes on it. "The day I trust it completely is the day I send a client a cheerful weekly report that's quietly fabricated. So I don't."

Model updates are the other recurring tax. Twice in the last year, a provider changed a model's behavior and a prompt that had worked for months started producing subtly different output. "Nobody warns you. You find out because a client says 'this report reads weird this week.'" She now keeps a small set of "golden" test inputs and runs them after any model change — a habit she borrowed from software testing.

"The agents aren't employees who get better on their own. They're tools that silently change shape under you."

The week she almost hired someone

The closest Marta came to abandoning the solo model was month eleven. She had a pipeline of interested clients, more than 22 hours of work to do, and a recruiter friend offering to find her a junior operator.

She ran the numbers and didn't do it. "A hire is a fixed cost that arrives before the revenue does. I'd have needed to add maybe $2,500/month in retainers just to cover them, before I made a cent. And I'd have traded the thing I actually like — that the business fits in my own head — for the thing I left freelancing to avoid: being responsible for someone else's mortgage."

Instead she raised prices and let two low-fit clients churn. The business got smaller for a month and then larger, without a payroll. "Not hiring was the most counterintuitive good decision I made," she says. "Everyone tells you growth means people. Sometimes growth means better clients and a higher price."

The metric she tracks now that she ignored before

If Marta could go back to her $900 month and hand herself one number to watch, it wouldn't be MRR. It would be what she now calls "effort per dollar" — a rough sense of how many of her own hours each client's monthly fee actually consumes.

"Early on I had two clients paying about the same and one of them was eating four times the hours," she says. "On the MRR dashboard they looked identical. In reality one was a great business and one was a slow-motion disaster, and I couldn't see it because I was only looking at revenue." She now keeps a loose tally — not a stopwatch, just an honest monthly gut-check — of which retainers are getting heavier.

The pattern it reveals is consistent: the clients who consume disproportionate hours rarely become her best long-term accounts. "Heavy clients don't grow into easy clients. They grow into resentment. The light ones — the ones where the agents just hum and the Friday report writes itself — those are the ones I want ten more of."

"Two clients can pay the same and one is a business while the other is a disaster. MRR can't tell them apart. Hours can."

It's the metric, she says, that turned her from someone chasing a revenue number into someone building a business she actually wants to run. "Revenue told me how I was doing. Effort-per-dollar told me whether I'd still be doing it in two years."