I Replaced My VA With Claude API—Saved $3K/Month

I paid a virtual assistant €650 a month for two years. She was good. She was also the single biggest bottleneck in my business every time she took a holiday, got sick, or simply had a bad week. In February 2024, I started quietly replacing her tasks with Claude API automations. By June 2024, I had let her go entirely. That decision saved me €7,800 in the first year and added roughly 15 billable hours back into my week. This is the full story of how that happened — including the parts that didn’t work.

Why I Started Looking for a VA Replacement

Running a one-person real estate consulting business in Madeira means wearing every hat. I’m the salesperson, the marketer, the analyst, and — before 2024 — the editor who reviewed everything my VA sent me before it went out the door.

That last part is the one nobody talks about. Even with a VA, I was still spending 90 minutes a day reviewing her work, rewriting her property descriptions, correcting her client emails, and explaining context she never quite had enough of. She wasn’t bad at her job. The problem was structural: no matter how good your VA is, they are not inside your head. They don’t know that a particular buyer from Germany has already seen three listings and is losing patience. They don’t know the tone I use with long-term local contacts versus new international leads.

I started testing the Claude API in early 2024 not because I’d read some article telling me to. I started because I was exhausted from the constant context-switching. A friend who runs an e-commerce store mentioned he’d wired Claude into Make.com for customer support replies and cut his response time from 6 hours to 11 minutes. That conversation stayed with me for about a week before I opened an Anthropic account and started experimenting.

The Exact Tasks I Automated With the Claude API

Before I get into the how, here’s the full list of what my VA used to do and what I moved to Claude API automation. Some transitions were seamless. Two were complete failures. I’ll be clear about which is which.

The successful automations:

• Property description drafts — I fill in a structured form (property type, size, location, key features, target buyer profile), and a Make.com scenario sends that data to the Claude API, which returns a Portuguese and English description in my house style. I review and publish. Used for 47 listings since I set this up.
• Lead follow-up email sequences — When a new inquiry comes into my CRM, a trigger fires a Make.com flow that sends lead data to Claude. It returns a personalized first reply and a 3-email follow-up draft. I approve before sending, but I’m not writing from scratch anymore.
• Monthly market analysis reports — I export raw data from my local market sources, run it through a Python script that formats it, and pass it to Claude with a fixed prompt. It returns a structured report I send to my newsletter list. My VA used to spend 4 hours on this. It now takes me 35 minutes total, including my own review pass.
• Social media content calendar — I give Claude a topic, a target audience, and five key facts. It returns 8 posts in different formats (carousel hooks, single-image captions, short videos scripts). I used to spend half a Saturday on this every two weeks. Now it’s a 20-minute task.

The failed automations (and I mean genuinely failed):

• Client phone call summaries — I tried using a transcription tool to feed call notes into Claude for auto-generated action summaries. The transcription quality in European Portuguese was inconsistent enough that Claude was working with garbage input and producing unreliable outputs. I dropped this after 3 weeks.
• Legal document pre-drafting — I experimented with having Claude draft preliminary clauses for buyer agreements. A real estate attorney friend reviewed three of the outputs and flagged two issues that would have been problematic under Portuguese property law. This one I killed immediately. Claude is not a lawyer, and in real estate, that boundary matters.

How I Actually Set This Up (No CS Degree Required)

The technical stack I use is not complicated. I’m not a developer. I can read Python well enough to modify existing scripts and I’m comfortable in Make.com, but I’ve never built software professionally.

Here’s the core setup:

• Make.com — Handles the triggers and the workflow logic. When something happens in my CRM or my inbox, Make catches it.
• Claude API (claude-3-5-sonnet) — The model I use for almost everything. I tested Haiku for cost savings on high-volume tasks, but the output quality difference was noticeable enough that I stayed on Sonnet for client-facing work.
• Airtable — Where I store my prompt templates. Each automation has a base prompt that lives in Airtable, so I can update the wording without rebuilding the Make scenario.
• Google Docs — Final outputs land here for my review before anything goes to a client.

My total API spend runs between €60 and €90 a month depending on volume. Compare that to €650 for a VA. The math is not subtle.

My Real-World Experience: The Month I Had 19 Active Listings

October 2024 was the most listing-heavy month I’ve had in my 12 years in this business. A development project came through that added 14 new units to my portfolio on top of the 5 I was already actively marketing. In a normal year, that would have meant calling my VA, explaining every property in detail, waiting 3–4 days for drafts, spending another day editing, and still feeling like half the descriptions didn’t quite capture what made each unit worth seeing.

Instead, I sat down for two hours and filled out my structured property intake form for all 19 listings. I’d built this form in Airtable specifically for this workflow — it captures property type, area in square meters, orientation, key features, proximity to local landmarks in Madeira, price range, and the buyer profile I’m targeting (investment buyer, lifestyle relocator, golden visa applicant, etc.). Once those forms were complete, I triggered the Make scenario and walked away.

Two hours later, I had 19 property descriptions — each in both Portuguese and English — waiting in a Google Docs folder. Each ran between 180 and 230 words. The tone matched my house style because I’d spent considerable time in September refining the system prompt to include specific phrasing I use, sentence rhythm patterns, and things I explicitly do not want (superlatives like “stunning” and “breathtaking,” vague phrases like “nestled in,” anything that sounds like it was written by a hotel brochure).

I reviewed all 19 in 55 minutes. I edited 6 of them — mostly small factual details I’d not captured precisely in the intake form, and two cases where Claude had emphasized the wrong selling point for the buyer profile. The other 13 went live with only light copy edits for punctuation or local terminology.

In the old workflow, 19 descriptions with my VA would have taken a combined 6–7 hours of her time, at least 3 days of back-and-forth, and around 2 hours of my own editing time. This time: 2 hours of intake forms, then 55 minutes of review. Under 3 hours total, start to published. I saved at minimum 6 hours on that single task in that single month.

The follow-up sequence automation was running in parallel. Of those 19 listings, 11 generated new inquiries within the first 10 days. Every one of those inquiries got a Claude-generated first reply that referenced the specific property they’d asked about and included three qualifying questions I always ask early in the buyer relationship. I reviewed each before sending — that part I haven’t automated away, and I don’t intend to. But the drafts were solid enough that my average editing time per email was 4 minutes.

The month I used to dread became one of the most operationally smooth months I’ve had. Revenue from that batch of listings ended up being my highest single quarter since 2019.

Honest Limitations: What Claude API Automation Cannot Replace

Anyone telling you this approach is frictionless is selling something. Here’s what still doesn’t work well, after more than a year of testing.

Relationship management is still fully manual. Claude can draft an email, but it cannot read the subtext of a client relationship. A buyer I’ve been working with for 8 months who suddenly starts replying in shorter messages — that’s a signal I pick up on and respond to. No automation catches that.

The setup time is real. I spent probably 30 hours total across February–April 2024 building, testing, and refining these workflows. If you’re in a rush or not comfortable with no-code tools, the initial investment is significant. There’s no plug-and-play solution here — you are building something custom.

Prompt maintenance is ongoing work. The system prompts that drive my automations need revisiting every 2–3 months. Markets change, my voice evolves, new property types require new framing. The automation doesn’t update itself. I have a recurring task every quarter to audit my prompts and check output quality hasn’t drifted.

Claude API pricing is usage-based, not fixed. In a high-volume month, my costs can hit €110–€120. Still well below a VA’s cost, but if you’re not monitoring usage, it can creep. I use Make’s built-in logging to track API calls per workflow.

Cost and Time Comparison: VA vs. Claude API Automation

Category	Virtual Assistant	Claude API Automation
Monthly cost	€650	€60–€120 (API + Make.com)
Property descriptions (batch of 10)	3–4 days turnaround	2–3 hours (intake + review)
My review time per task	90+ minutes/day	20–40 minutes/day
Availability	Business hours, holidays excluded	24/7
Setup time	Onboarding: 2–3 weeks	Initial build: ~30 hours
Consistency of output	Variable (depends on her day)	High (prompt-controlled)
Legal/sensitive tasks	Could handle with guidance	Not safe — do not automate

What I Would Do Differently If Starting in 2026

If I were building this from scratch today, I’d make three changes.

First, I’d spend more time on the intake forms before touching the API at all. The quality of what Claude returns is almost entirely determined by the quality of structured data you send it. Garbage in, garbage out is aggressively true here. My intake form went through four redesigns before the outputs became consistently usable.

Second, I’d start with one workflow and run it in parallel with my VA for at least six weeks before cutting anything. I was impatient and switched over property descriptions unilaterally in month two. It worked out, but there were two weeks where output quality was inconsistent and I had to manually rewrite more than I expected. A proper parallel-run period would have let me catch prompt issues before they affected client-facing work.

Third, I’d have kept the VA on a reduced retainer for relationship-heavy tasks rather than a clean break. In retrospect, there were things she did — managing my calendar, coordinating property visits, chasing paperwork from notaries — that I now handle myself because I haven’t found a good automation path for them. Keeping her 10 hours a month for those tasks might have been the smarter call.

Is This the Right Move for Every Solopreneur?

Honestly, no. This works well for solopreneurs who produce a high volume of written content as part of their core deliverables — real estate descriptions, reports, email sequences, social content. If your VA is primarily handling calendar management, travel bookings, or anything that requires phone calls and human judgment calls, this approach won’t replace much of that.

It also requires a baseline comfort with tools like Make.com and a willingness to invest real time upfront. The ROI is there — my first-year savings were €7,800 in direct VA costs, plus I estimate I recovered 600+ hours of review and editing time — but it’s not passive. You’re building a system, not buying a subscription.

For solopreneurs in content-heavy service businesses who can invest 20–30 hours in setup and are willing to maintain their prompts quarterly, replacing a VA with Claude API automation is one of the highest-ROI operational decisions I’ve made since starting this business in 2012.

Practical Summary and Next Steps

Here’s what actually matters from everything above:

• Claude API automation replaced my VA for written content tasks — property descriptions, email sequences, market reports, social content
• My monthly cost dropped from €650 to €60–€120
• I recovered an estimated 15 hours a week across review time and task handoff overhead
• The setup took ~30 hours across three months — not a weekend project
• Legal drafting and relationship management stayed human — no exceptions
• The biggest mistake I’d avoid is skipping the parallel-run phase before going all-in

If you’re a solopreneur considering this, start with one workflow. Pick the most repetitive written task you currently delegate. Build the intake form first — before you touch the API. Run it alongside your existing process for 4–6 weeks. Measure the output quality honestly, not optimistically. Then decide whether to expand.

The Anthropic API documentation is solid and the Make.com + Claude connector is well-documented. You don’t need to hire a developer to get started. You need patience, a structured approach, and a clear-eyed view of what automation can and cannot do.

If you want to see the exact Make.com scenario structure I use for property descriptions, I’ve outlined the full workflow in my solopreneur tools guide. Start there, then come back with specific questions — I read every comment.