How to use Claude as a Virtual Assistant (and why clients expect it)

If you’re a VA, EA, or PA you’ll have noticed the conversation around AI has changed. It’s no longer, “do you use AI?” it’s “how do you use AI, and what’s your skill level?” Clients are looking for Assistants who are AI-fluent, not just using it here and there.

We recently ran one of our internal Lunch & Learn sessions on exactly this. How the Upsource team uses Claude (Anthropic’s AI assistant) day-to-day across client accounts. And we thought we’d share it as a resource for fellow freelance and agency affiliated Assistants.

Why Claude?

ChatGPT became the default because it was first to reach the wider public consciousness. Most people default to ChatGPT because it’s the name they’ve heard of. And it’s good, enough. But you’ll probably have been on the receiving end of some iteration of the “Claude is the winner” conversation. 

In the last six months, Claude and Anthropic have been everywhere, new product releases, making a stand against the US government, and a starring ad at the Superbowl promising “no ads, ever”. But Upsource has moved towards it for more practical reasons: it handles long documents well, reasons carefully through complex requests, as questions when it needs clarification and is less prone to confidently citing facts that aren’t actually true.

It also has some very useful structural features and a good understanding of them will help you master working with AI.

• Projects work like folders, a separate space for each client, workstream, or topic, so Claude only draws on the context relevant to that specific area.
• Skills are pre-built capabilities, creating Word docs, Excel spreadsheets, PowerPoint decks without any setup beyond switching them on.
• Connectors link Claude to tools you already use: Gmail, Slack, Google Drive, ClickUp, calendars. Once connected, Claude can read from, and act on, those tools directly.
• Memory lets Claude retain preferences, terminology, and context over time, so you’re not re-explaining the basics every time you use it.

What Assistants (and everyone!) aren’t talking about enough: account security

This is especially for freelance VAs juggling multiple clients

Use a separate Claude account per client
Treat it the same way you’d treat separate browser profiles. Don’t let one client’s information end up in a conversation intended for another. If a client is paying for their own Claude seat (increasingly common, and something worth proposing if you’re doing significant AI-assisted work for them), work within their account so their data stays contained there.

Never use real client names in your own personal Claude
If you’re using it to think through client problems, talking through a “tricky client situation” is fine, but anonymise it. Use a codename (Client A, Client ABC) rather than a real name, the same instinct you’d apply to GDPR. Claude will start picking up on patterns over time if you talk about the same anonymised situation repeatedly, so if it starts referring to someone in particular, reset the thread.

Be deliberate about what you share, and be careful with connectors
If Claude is connected to a client’s CRM, inbox, or support tools, give it what the specific task actually requires, not blanket access to everything. Claude will typically ask for permission before it connects to something or sends data anywhere. Make sure you understand what those permissions really mean before clicking through it automatically.

Never let AI send anything on your behalf
Ask it to draft, not send. This applies to emails, messages, anything client-facing. We can all tell when we’re receiving something written by a “robot”, and it smacks of carelessness.

Always tell your client when something was AI-assisted.
We always recommend the transparent, honest approach – and your clients will appreciate it. Further than that, this transparency protects you: if something’s slightly off, your client understands why, and knows to sense-check everything rather than sign it off without the proper attention.

Worth knowing, too: on personal/consumer Claude accounts (Free, Pro, Max), conversations can be used to train future models. It’s an opt-out setting in Privacy Settings, not opt-in, and if you leave it switched on your data can be retained for up to five years rather than the standard 30 days. 

Claude for Work (Team and Enterprise plans) isn’t used for training at all, by default. This is a positive of working inside a client’s paid account rather than your own personal one for anything client-related. It’s not just tidier, it’s a different data policy. “Strong protections” on the commercial side are no substitute for good judgement. Treat every prompt the way you’d treat any other written communication containing client information.

Where Claude shines

Here’s how Upsource has found Claude to be really useful in our VA work:

• Operations: sorting and structuring data, building workflows, helping design automations
• Financial admin: with explicit permission, working through VAT queries or tax calculations
• Contracts: stress-testing a contract before you sign it, or before you send one out, by asking Claude to look at it from the other party’s perspective
• Troubleshooting: working out why a piece of tech has broken, before you escalate it

None of this replaces human judgement, which is what Assistants are really hired for. Every output still needs a person to check it, and polish it, humanise it, before it goes to a client, or to the public.

One of our favourite automations that will actually save you time

A scheduled daily brief.

Set Claude to run every weekday morning, reading from your connected inbox, CRM, calendar, and task manager, and produce a short summary: who and what needs attention, why, how urgent, and anything left over from yesterday’s list. It takes about fifteen minutes to set up initially, and it means every morning opening your laptop to a ready-made starting point, absolutely no effort required.

And you can scale that logic up. Some teams build an internal knowledge base that Claude reads from before answering anything about a specific client or business, so brand voice, process decisions, client history, which means new information compounds instead of getting lost, and new starters can get up to speed in days rather than weeks.

A few honest caveats

Claude isn’t infallible, here’s where to keep an eye out for missteps.

• It doesn’t reliably know when it’s wrong, and will state things confidently either way. Better to sense check, and come to it with an idea of what the right answer is, rather than blind.
• Left untrained, it will be your own personal hype man, whether you and your work deserve it or not. If you want honest pushback, you need to explicitly ask it to stress-test your work, not just review it.
• Long conversations can lose context. Ask it to remember specific context explicitly where it’s important, or start a fresh conversation for a new topic.

AI isn’t replacing VAs, it’s changing how we work

AI hasn’t reduced the value of a good Assistant, but it’s changing what we are working on and where we are really demonstrating our value. Less time on easily automated repetitive admin, more time on judgement calls, discretion, and relationship management. The fundamental human skills that AI can’t replicate.

As Upsource teamie Bo, put it in the Lunch and Learn:

“There will always need to be a driver. If you’re equipped with the knowledge of how to use Claude, whether it’s Claude or GPT or whatever, if you become the expert in it, you’re not going to lose your job, because clients will always need someone to drive the AI.”

The Virtual Assistants who lead the way and leaning into AI are the ones clients are actively looking for right now.

If you’re a Virtual Assistant who wants to work the way Upsource does, with excellent training and ongoing support, we’d love to hear from you. Check out our careers page here.