I Replaced My VA With Claude Agents: 60% Faster

I fired my virtual assistant in March 2026. Not because she did bad work — she was genuinely good. I let her go because I built a Claude agent workflow that does 80% of what she did, runs 24 hours a day, costs me €47 a month instead of €650, and never asks me to repeat myself. That decision felt uncomfortable for about a week. Then my next invoice cycle hit and I stopped second-guessing it.

I want to be precise about what I mean by “agent workflow” before this turns into another vague AI article. I’m not talking about typing prompts into Claude.ai and copying the output. I’m talking about a configured system — Claude as the reasoning core, connected to my actual business data, running repeatable tasks with minimal manual input from me. That’s a different thing entirely, and it took me about six weeks to get right.

Why I Had a VA in the First Place

Running a one-person real estate consultancy in Madeira since 2012 means I wear every hat. Listings, client emails, market reports, follow-up sequences, social media, property descriptions in three languages — that’s the full job description. By 2022 I had outsourced roughly 12 hours of weekly admin to a remote VA based in Eastern Europe. She handled first-draft email replies, formatted my property descriptions, scheduled social posts, and compiled my weekly inquiry summaries.

The arrangement worked. But it had friction. Briefing her on new listing types took time. When I changed my tone for a specific client segment, I had to re-explain everything. Turnaround was 24 hours minimum. And at €650 per month for part-time hours, it was my second-largest business expense after my CRM subscription.

When Claude 3 Opus landed and I started testing it seriously in late 2023, I noticed something: given the right context, it matched her output quality on drafting tasks. Not immediately. But once I learned how to write proper system prompts and feed it the right business context, the gap closed fast.

What I Actually Built: The Workflow Explained Step by Step

The phrase “Claude agent workflow” gets thrown around loosely. Here’s exactly what mine looks like in practice, as of mid-2026.

The Core Setup

I use Claude via the Anthropic API, connected to Make.com as the automation layer. My business data — listing details, client profiles, email templates, tone guidelines — lives in a Notion database that Make.com reads from. When a trigger fires (new inquiry email, new listing added, weekly report scheduled), Make.com pulls the relevant data, constructs a prompt, sends it to Claude, and routes the output to wherever it needs to go: my Gmail drafts, my CMS, my WhatsApp Business account, or a Google Doc.

The API cost averages €31–47 per month depending on volume. Add Make.com’s Core plan at €9/month and I’m at roughly €56 total. Against €650 for the VA, the math is not subtle.

The Five Tasks Claude Now Handles

1. Property description drafts. When I add a new listing to my Notion database with raw details (location, size, price, features, target buyer type), Make.com triggers a Claude prompt that generates a description in English, Portuguese, and German. My VA used to do this in about 2 hours per batch. Claude produces all three versions in under 4 minutes. I edit for maybe 10 minutes. Last month I processed 9 listings this way.

2. Inquiry email replies. New inquiries come into a dedicated Gmail label. Make.com monitors that label, extracts the sender’s question, cross-references my listing database, and sends a Claude-generated draft to my Gmail Drafts folder. I review, adjust if needed, and send. About 70% of drafts go out with minimal edits. This handles roughly 35–40 emails per month.

3. Weekly market summary reports. Every Monday, a Make.com scenario pulls recent transaction data I’ve logged, plus notes from my Notion, and prompts Claude to draft a 400-word market update I send to my subscriber list. My VA used to spend 90 minutes on this. The draft is ready before I finish my first coffee.

4. Lead follow-up sequences. When a lead goes cold (no response in 7 days), the workflow generates a follow-up email personalized with the specific property they inquired about. Not a generic nudge — an actual contextual message that references their situation. Conversion on these has stayed consistent with what my VA produced.

5. Social media captions. New listing goes into Notion, Make.com fires, Claude produces Instagram and LinkedIn captions in my voice (the system prompt contains extensive examples of my writing style). I post manually — I haven’t automated that step and I’m not sure I want to.

My Real-World Experience Running This for 8 Months

The most concrete example I can give you: February 2026 was a heavy month. I had 11 new listings come in within three weeks — a combination of a small development I’d been brought onto and a few individual sellers who came through referrals. Normally that volume would have overwhelmed my VA’s part-time hours and my own schedule. I would have been chasing her for drafts while also handling client calls.

Instead, I processed all 11 listings through the Claude workflow. Property descriptions in three languages: done. Announcement emails to my subscriber segments: done. Social captions queued up: done. Total time I personally spent on content production for those 11 listings was approximately 2 hours and 20 minutes — mostly reviewing and light editing. My pre-workflow estimate for the same output, split between me and my VA, would have been 9 to 11 hours combined.

That same month I handled 47 inquiry emails through the drafting workflow. I sent 38 of them with edits of under 2 minutes each. The other 9 needed more work — complex legal questions, specific pricing negotiations, one client who was difficult to read and needed a more careful tone. Those I wrote from scratch. The workflow doesn’t pretend to handle everything, and that’s fine.

What surprised me most was not the time saved — I expected that. What I didn’t expect was the consistency. My VA was good, but her output varied. Some weeks her property descriptions were sharp. Other weeks they felt rushed. Claude’s quality is flat. It doesn’t have an off day. Given the same input quality, I get the same output quality. For someone running a solo operation where brand voice is everything, that consistency is genuinely valuable.

I also noticed my turnaround time for new inquiries dropped from an average of 19 hours (waiting for my VA’s shift) to under 3 hours. In a competitive market like Madeira, where international buyers are often comparing multiple consultants simultaneously, responding faster closes more deals. I can’t attribute revenue directly to this change, but anecdotally, my contact-to-call conversion rate improved in Q1 2026.

The setup took about 6 weeks to get right. The first two weeks I was still running both the VA and the workflow in parallel, comparing outputs. Weeks three and four I refined the system prompts — particularly the ones that govern tone and language register for different buyer profiles (local Portuguese buyers versus Northern European expats read very differently and need different messaging). By week six I was confident enough to let the VA go.

Where Claude Agent Workflows Fall Short

I want to be honest here because this is where most AI articles go soft.

It cannot handle genuinely novel situations. When a client sends a complicated email with an emotional undertone — say, someone going through a divorce and navigating property ownership stress — Claude’s draft is technically correct but often emotionally flat. I catch these and rewrite them personally. My VA, who had worked with me long enough to know my client base, handled these better instinctively.

The setup time investment is real. If you’re not comfortable with Make.com, API keys, and prompt engineering, the 6-week setup timeline I described is probably optimistic for you. A non-technical freelancer would likely need 3–4 months or would need to hire someone to build the initial workflow. That upfront cost matters.

Portuguese output quality lags English. This is a known Claude limitation. My English descriptions are excellent. My Portuguese ones need more editing — probably 25 minutes per batch instead of 10. German sits somewhere in between. For a Madeira-based business where local Portuguese content matters, this is a real limitation, not a minor footnote.

Context window management is manual work. For long client histories, I have to manually curate what context I feed into each prompt. There’s no memory that persists across my workflows automatically — I have to build that data layer in Notion and keep it clean. That’s ongoing maintenance work that didn’t exist with a human VA who simply remembered things.

Cost Comparison: VA vs Claude Agent Workflow

Item	VA (Previous)	Claude Workflow (Current)
Monthly cost	€650	~€56 (API + Make.com)
Property descriptions (9–11 listings)	~6 hours combined	~2 hours (review only)
Email inquiry turnaround	~19 hours average	~3 hours average
Availability	Part-time hours, one timezone	24/7, instant
Languages handled well	English, Portuguese	English, German (Portuguese needs editing)
Emotional nuance in difficult emails	Good (experienced)	Weak — I write those manually
Setup time to start	1–2 days onboarding	~6 weeks to build properly

Who Should Actually Do This (and Who Shouldn’t)

This workflow makes sense if you’re a solo operator or small freelancer with repetitive, structured output tasks — drafting, summarizing, formatting, following up. If your business runs on predictable content types that follow a template logic, Claude can learn those templates and execute them faster than any human assistant.

It doesn’t make sense if your work is heavily relational, judgment-heavy, or involves tasks that don’t repeat in a structured way. If your VA spends most of her time on client relationship management, research that requires live internet access, or tasks that need real-world coordination, the workflow won’t cover those gaps cleanly.

I’d also say this clearly: if you’re not prepared to invest time upfront — building the Notion data structure, writing and testing system prompts, configuring Make.com scenarios — you’ll end up with a half-broken workflow that frustrates you and makes you miss your VA immediately. The payoff is real, but it’s not instant.

My Rating and What It’s Based On

Claude agent workflow for solo freelancer operations: 4.2/5. It replaced a €650/month expense with a €56/month system that handles 80% of the same tasks at comparable or better quality — that result, across 8 months of daily use in a real estate business, justifies a high score, with the deduction for the weak Portuguese output and the real setup investment required.

Practical Summary and Next Steps

Here’s what I’d tell someone who’s considering this switch in 2026:

• Start by auditing your VA’s tasks. List everything she does. Identify which tasks are structured and repetitive versus which require genuine human judgment.
• Build the Claude workflow for one task only first. I started with property descriptions. Prove it works there before expanding.
• Use Make.com as the automation layer — it’s the most flexible option I’ve tested for connecting Claude’s API to real business data sources.
• Write your system prompts with real examples from your actual work. Generic prompts give generic output. Feed Claude your best past emails, your best property descriptions, your actual tone guidelines.
• Keep a manual override habit. The 20% of tasks that don’t fit the workflow still need you. Build time for those rather than assuming the workflow covers everything.

The decision to let my VA go wasn’t really about saving money — though €594/month in savings is hard to ignore. It was about eliminating the coordination overhead that came with having a human dependency in a one-person business. Now my output pipeline runs whether I’m on a property visit in Funchal or sitting at my desk. That operational independence has changed how I work more than any individual feature of Claude itself.

If you’re running a freelance or solo operation and you haven’t seriously tested a structured Claude workflow yet, the gap between what you’re spending on admin support and what this costs is worth one honest afternoon of your time to investigate. Start with the Anthropic API docs and Make.com’s Claude integration — both are well-documented and you don’t need to write a single line of code to get a basic version running.

Have you replaced a VA or assistant with an AI workflow in your own freelance business? I’m genuinely curious what’s worked and what hasn’t — drop a comment below or reach out directly.