Why we don't replace VAs with AI (and probably never will)
An Australian VA agency's contrarian take. AI is brilliant at outputs and bad at judgement. Here are the seven tasks where AI consistently loses to a competent VA, with the data behind why.
I get asked this every fortnight, often by founders who already use Claude or ChatGPT daily. Why are you still placing humans?
The honest answer is that we have run the experiment several times. Every six months we audit every task our VAs do and ask: can AI do this end-to-end, without supervision, at the same quality?
The list of “yes, with supervision” gets longer every audit. The list of “yes, end-to-end, no supervision” has not moved much. Here is what we have learned about where AI keeps losing.
What AI is genuinely brilliant at
Before the contrarian part, the credit where it is due. AI is genuinely transformative at:
- Draft outputs. First draft of an email, a SOP, a meeting agenda, a customer reply. Faster than a human, usually 80 per cent of the way there.
- Pattern matching on text. Summarising a meeting transcript, extracting action items from a Loom video, pulling fields out of a PDF invoice.
- Bounded research. Comparing 12 options against a known criteria list, building a feature matrix, producing a market-rate range.
- Translation between forms. Turning bullet points into prose, turning prose into bullet points, restructuring a document.
A VA without AI in 2026 is like a designer without Figma in 2018. They can still work, but they are slower, and the floor of their output quality is lower.
We train every VA on Claude and ChatGPT in their first week of onboarding. That is non-negotiable. The AI-augmented VA is faster than a human-only VA at every workflow we measure.
That is the multiplier. Here is where the multiplier ends.
The seven losses
1. Escalation
A customer emails. They are unhappy. They have used the word “lawyer”. They are also a long-standing client whose annual spend matters.
The right move is judgement under uncertainty. Maybe the right reply is “we hear you, let us call you today”. Maybe it is “this is what our terms say”. Maybe it is “let us refund you and never hear about this again”. Maybe it is “we need to escalate this to our managing director”.
I have not yet seen an AI make that call well, because the inputs that decide it are not in the email thread. They are in last quarter’s relationship history, today’s company cash position, our owner’s personal tolerance for refunds, and the rough probability that this customer is bluffing.
A human VA who has worked with you for six months knows which of those levers to pull, or, more often, knows when to flag it to you rather than pull any.
2. Ownership
A task that requires owning an outcome over weeks, not minutes.
“Our suppliers are slow at responding. Get us back on track” is not a prompt. It is a multi-week project that requires choosing the right channel for each supplier, calling instead of emailing when the email fails, escalating to our team when the supplier ghosts, following up at the right cadence so they do not forget us but do not feel harassed.
AI can draft any individual message. AI cannot run the project. The state of “we are halfway through, three of the eight suppliers are sorted, two are dead leads, one needs a phone call I cannot make from here” is not something AI carries forward across sessions. A human VA who owns the project does carry it.
3. Novel context
The fastest way to break Claude in production is to give it a task that depends on context you forgot to share.
Our VAs absorb that context over time. They know that Tuesday is when the founder is at the studio so do not schedule calls. They know that this customer’s contract has a clause we negotiated verbally and never wrote down. They know that the last time we tried this kind of campaign we got complaint emails about exactly this wording.
AI starts every conversation with zero context. You can stuff context windows, build memory layers, use retrieval. It is still rebuilding from scratch every session. A six-month VA brings six months of accumulated unwritten knowledge to every task.
4. Ambiguous priorities
“What’s most important right now?” is the question I ask my team three times a week. It is also the question I cannot reliably get useful answers to from an AI.
Priority is a function of: what is on fire today, what the founder values this quarter, what we promised a client three weeks ago and forgot, what our cash position requires, what the team’s energy can sustain. The inputs are mostly tacit. The answer changes hourly.
A VA who has been embedded in the business for three months can answer “what should I work on?” usefully. An AI cannot, unless you re-load all the context into the prompt, at which point you have done the priority work yourself.
5. Sensitive communication
Telling a long-term contractor that we are not renewing them. Letting a client know that we cannot make their deadline. Pushing back on a partner’s pricing without burning the relationship.
These communications are 5 per cent the words and 95 per cent the choice of words. Tone, tact, the exact level of warmth that maintains the relationship while still saying the hard thing. AI can draft these. Every time we have shipped an AI draft unedited, the relationship has been worse for it.
We use AI to draft. Then a human reads it, finds the three sentences that read off, and rewrites those. The rewrite is the value-add. The draft was just the cost saving.
6. The soft no
Saying “no” politely is one of the underrated skills of small business operations.
“Can you do this favour for me?” “Can we get a discount?” “Can you call me right now?” The right answer is often no. But “no” in writing reads harsh, and “no” delivered badly costs you the relationship.
A skilled VA learns your soft-no language over time. The way you decline without sounding like you are declining. The bridge phrases you use to redirect to what you can do. AI defaults to either too-direct (“Unfortunately we cannot accommodate this request”) or too-saccharine (“We so appreciate you reaching out and absolutely understand…”). Both feel wrong. The middle requires a human ear.
7. Follow-through
The single hardest thing to automate is following up on things humans have not done.
“I asked the supplier three weeks ago, they said they’d come back, they haven’t, what now?” That requires deciding to call, looking up the right person to call, calling at the right time of day for their timezone, leaving the right voicemail, deciding what to do next if voicemail fails, remembering to try again the next day.
It is a chain of small judgement calls, each easy alone, impossible end-to-end without persistence. AI does not persist across sessions. Humans do.
What this looks like in practice
A working example from the last quarter, lightly fictionalised.
Task: chase 47 small overdue invoices for a client. Total AR balance about $38k AUD. Mix of clients who forgot, clients who disputed, clients who are genuinely struggling.
AI alone: drafts 47 chase emails in 4 minutes. Decent template, customer name merged correctly. Sends them. 11 customers reply (some pay, some object, some confused). 36 silence. AI cannot decide what to do next.
VA alone: writes 47 chase emails in 90 minutes, slightly less polished but more personal. Same response rate (about 25 per cent). VA then makes 8 phone calls to the silent ones over the next week, identifies 3 that are genuinely struggling and need a payment plan, escalates 1 to the founder for a decision, recovers $34k of the $38k.
VA + AI: VA uses Claude to draft the 47 emails in 6 minutes, edits the tone in another 20 minutes, sends them. Same response rate. VA then makes the 8 calls and runs the payment plans. Total time: 4 hours vs the VA-alone’s 10 hours. Same recovery.
The recovery is what the client paid for. The recovery required the calls, the payment plans, the escalation. AI did not do any of those, and would not have without a human directing it.
The next five years
I keep being told that next year’s Claude will do all of this.
Maybe. The trendline on AI capability is real, and I am not going to pretend it is flat. But the specific failure modes above are not “AI is not smart enough yet”. They are “the inputs the task needs are not in the prompt”. That gap is not closed by a bigger model. It is closed by an AI that has worked inside your business for six months, knows your customers, has the relationships, makes the calls.
That thing already exists. We call them virtual assistants.
We will keep using AI to multiply them. We are not in a hurry to replace them.
What to take from this
If you are evaluating whether your business needs a VA in 2026: the bar has not moved much. The 15-20 hours a week of judgement-heavy admin still requires a human. AI makes that human 1.5-2x faster, which is the wedge that makes our pricing work.
If you want to try the AI-multiplier pattern for yourself before hiring, our prompt library for AU VAs is the lift-off point.
If you are ready to talk about an actual placement, book a discovery call. We will tell you which of your tasks AI can take, and which need a VA. Honest answers, even when they cost us a sale.
Tools mentioned in this post
- Claude (Anthropic)
- ChatGPT (OpenAI)
- Perplexity