The opportunity in AI for virtual assistants is not “do more admin for less money”. It is to become the person who knows what should be automated, what needs a human check and what must never cross a client's privacy boundary.
That is a better service. It is also harder to replace than someone selling hours of copy-and-paste work.
What AI should and should not do in VA work
- AI sorts information into a known system. You decide what is urgent.
- AI drafts from an approved template. You send it and represent the client.
- AI extracts actions from a transcript. You confirm owners and deadlines.
- AI turns a screen recording into an SOP draft. You test whether the process works.
- AI compares data across a reporting period. You explain what changed and why.
The difference is responsibility. Software can produce an output. Your client hired you to make sure the output is useful and safe.
Workflow 1: inbox triage with approval kept in place
Begin by sorting. Keep sending behind a human decision.
Create categories with the client: urgent today, client decision needed, routine reply, delegated task, invoice or receipt, and reference only. Use real examples to define each category.
Then let AI:
- classify new messages
- suggest a reply from an approved template
- extract any task and due date stated in the email
- flag uncertainty instead of guessing
Keep every external send behind approval until the client explicitly changes that rule. An inbox assistant should reduce scanning. It should not become an unsupervised version of your client.
Workflow 2: meetings to actions without losing the detail
Meeting summaries sound useful until nobody knows who owns the next step.
Use the transcript to create a structured action list. Every item needs the action, owner, due date, source quote and current status.
Tell the model to write “not stated” when the owner or date is missing. Never let it tidy ambiguity into a fake commitment.
After you review the table, move approved actions into the client's project tool. My guide to using AI meeting notes properly covers seven useful outputs beyond a generic summary.
Workflow 3: turn screen recordings into SOP drafts
This is a strong packaged service for virtual assistants.
Ask the client to record the process once while they do it. Use the transcript and screenshots to draft:
- purpose and expected result
- access required
- numbered steps
- decision points
- common errors
- what to do when the normal path fails
- owner and review date
Then run the SOP yourself. Every missing click becomes a correction. The finished document is valuable because it has been tested, not because AI wrote it quickly.
Workflow 4: create a weekly client report people will read
Most weekly reports list completed tasks. A useful report helps the client decide.
Collect data from the tools already in use, then have AI draft four sections:
- completed this week
- blocked and waiting on whom
- numbers that changed
- decisions needed next week
Add your interpretation before sending. If enquiries dropped, the client needs more than a red arrow. She needs the context you noticed while doing the work.
Workflow 5: research without handing over a pile of links
AI can speed up supplier comparisons, podcast research, competitor scans and travel options. Your value is in turning research into a decision-ready brief.
Require:
- the original link for each claim
- the date the information was checked
- a clear comparison criterion
- conflicts or missing information
- a short recommendation labelled as your recommendation, not the model's
Open the final sources. Prices, availability and product features change. A confident paragraph is not verification.
The small stack worth learning
- Claude or ChatGPT: drafting, extraction and document work
- Fathom or the client's recorder: meeting transcripts with consent
- Notion or Google Drive: SOPs and approved knowledge
- Zapier or Make: simple moves between common apps
- n8n: multi-step workflows, branching and more control
You do not need to master all five at once. Learn the general assistant already allowed in the client's business, then build one repeatable service around it.
How to package AI-assisted VA work
Do not sell “AI virtual assistance”. Sell the finished job.
- Inbox triage setup: categories, draft rules, escalation guide and one week of tuning
- Meeting action system: transcript, reviewed action table and task-tool handoff
- SOP sprint: five recorded processes turned into tested documents
- Weekly reporting system: approved metrics, draft template and review routine
State where human review happens. That makes the service clearer and builds trust.
Your first build: one meeting, one action table
Take a meeting transcript you are authorised to use. Run the action-table prompt. Check every owner, date and source quote. Move only verified actions into the project tool.
Once that works three times, automate the handoff. Not before.
For the wiring, read four business automations you can build this week. If you want support while you build, see how to choose the best AI community for women or visit Wright Mode.