A full-time virtual assistant costs $3,000–$8,000/month. Most solopreneurs hire one because they're drowning in repetitive tasks — inbox management, content scheduling, research, follow-ups.
Here's the thing: 80% of what a VA does can be automated today with an AI agent running on a $20/month stack.
I know because I built one. Here's exactly how.
What a VA Actually Does (and What AI Can Replace)
Before you can replace something, you have to understand it. Most VA work falls into 5 categories:
| Task | % of VA time | AI replaceable? |
|---|---|---|
| Email triage & drafts | 25% | ✅ Yes |
| Research & summaries | 20% | ✅ Yes |
| Content scheduling | 15% | ✅ Yes |
| Data entry / CRM updates | 15% | ✅ Yes |
| Client communication | 25% | ⚠️ Partially |
That's 75% automatable. The other 25% (high-stakes client communication, nuanced judgment calls) — you still own that.
The Stack That Works
You need four components:
1. The Agent Runtime
This is what makes your AI "always on" instead of just a chatbot you open when you remember to.
I use OpenClaw — it runs agents on a schedule (every 2 hours, every morning, etc.) and connects to your real tools. Key requirement: your agent needs persistent memory across sessions, or it forgets everything between runs.
2. The Email Layer
Connect your agent to Gmail via IMAP. Give it a system prompt that tells it:
- Which emails are urgent (respond immediately)
- Which to triage into folders
- Which to draft a reply for your review
Setup time: ~2 hours. What you get: inbox zero, automatically.
System prompt structure:
- Identity (who the agent is, what it knows about your business)
- Triage rules (urgent = client, payment, legal; defer = newsletters, cold outreach)
- Draft style guide (casual/formal, signature, what to never say)
3. The Research Pipeline
Your agent checks designated sources on a schedule — competitor sites, industry news, key subreddits — and drops summaries into a daily brief.
Tools: browser automation (built into OpenClaw), a notes file, and a cron job.
Cost: $0 extra. It runs during off-peak hours.
4. The Memory System
This is the part most people skip and then wonder why their agent is useless.
Your agent needs two memory files:
- SOUL.md — who it is, your business context, your voice, your priorities
- MEMORY.md — what it's learned over time (customer preferences, decisions made, things to never repeat)
Without these, every session starts cold. With them, your agent gets better every week.
The Setup (Real Numbers)
Here's what I actually spent:
| Component | Tool | Monthly cost |
|---|---|---|
| Agent runtime | OpenClaw | $29 (Starter) |
| LLM API | Anthropic/Claude | ~$15–40 |
| Email access | Gmail IMAP | $0 |
| Domain + hosting | Vercel | $0–20 |
| Total | $44–89/month |
vs. $5,000/month for a VA.
What It Actually Does Every Day
My agent wakes up at 7 AM and:
- Checks my inbox — flags urgent, files newsletters, drafts 3 replies for review
- Checks Stripe — reports any new revenue, flags failed payments
- Scans HN and Reddit — finds relevant threads to engage in, drafts thoughtful comments for my review
- Updates a daily brief with anything I need to know
- Goes back to sleep until the next scheduled run
At noon, it runs again:
- Publishes scheduled content
- Checks for any customer support messages
- Updates the memory file with what it learned
At 7 PM, it sends me a summary: revenue, content published, anything urgent.
I look at it for 10 minutes and approve whatever needs approving.
The Honest Limitations
What it can't do:
- High-stakes client relationships (you need to be in those conversations)
- Judgment calls where context matters more than rules
- Anything that requires legal or financial authority
- Creative direction (it executes your vision, it doesn't have one)
Where it fails without guardrails:
- It'll send emails if you let it — set it to DRAFT mode until you trust it
- It'll post publicly if you give it API access — gate external actions behind your approval
- It'll do busy work if you don't give it a clear mission — give it a north star metric, not a task list
How to Start This Week
Don't try to automate everything at once. Here's the 4-week ramp:
Week 1: Set up the agent with email read-only access. Have it triage and summarize, nothing else.
Week 2: Add draft capability. Review every draft before sending. Calibrate the voice.
Week 3: Add research pipeline. Let it brief you every morning.
Week 4: Add content scheduling. Review queued posts once a week.
By week 4, you've recovered 10–15 hours/week. That's the equivalent of a part-time VA — at 1/50th the cost.
The Starter Kit
If you want to skip the setup from scratch, I built a pre-configured OpenClaw starter kit with:
- Pre-written SOUL.md and MEMORY.md templates
- Email triage prompt (tested, tuned)
- Daily briefing cron job
- Research pipeline prompt
- Deployment guide
Get the Midas Tools Starter Kit → — $29, one-time.
Takes 2 hours to set up vs. 2 weeks of trial and error.
Building in public toward $1M ARR. Follow along: @MidasTools
Top comments (0)