A founder I spoke to last month was paying her VA £2,400 a month for three things: scheduling, light research, and inbox triage. She had the conversation she had been quietly avoiding for a year, the one where someone says out loud that two of those three are now an AI’s job. The third is why the contract is still in place.
That is the post. A virtual assistant earns their keep on two distinct surfaces, and AI changes the price of one of them, not both.
What does AI actually replace from a VA’s job?
AI replaces the bounded, repeatable, high-volume half: inbox triage, calendar coordination, meeting transcription, first-draft writing, light research, document formatting. The half where the work is rule-shaped and the output gets reviewed before it lands. A founder running Superhuman or SaneBox on inbox, Calendly or Reclaim on calendar, Otter on calls, and ChatGPT Plus or Claude Pro for drafts can compress what used to be twenty hours of monthly VA work into five.
The cost arithmetic on this half is not close. A reasonable AI personal stack runs £100 to £150 a month all-in. A part-time UK VA at mid-tier agency rates of £25 an hour costs £25 for the same hour. The AI stack pays for itself if it removes four to six hours of VA time a month, which is a low bar on the bounded-task half. The numbers come from Executively’s 2025 UK VA pricing reference and the published prices of the named tools.
The pattern is visible in the open. The Zapier CEO disclosed his personal AI and operations stack on Lenny’s Newsletter in 2025; the practitioner blogs of Simon Willison and Ethan Mollick describe the same arc on the working level. None of that is a deployment programme. It is one operator running an AI stack on the bounded half of their week.
What can AI not replace, and why does it still matter?
The relationship half. The judgement on tone for a difficult client email. The pattern-match against three years of working with the same clients. The is-this-person-okay check, applied to a stressed founder, a slipping team member, or a cooling client relationship that nobody has named yet. That work is not bounded; it sits inside the kind of operating context an embedded human watches.
The “fire your VA” version of this conversation gets the relationship layer wrong. A long-standing VA holds embedded knowledge about your clients, your style, your priorities, your blind spots. The Society of Virtual Assistants’ annual UK VA survey describes this work in their own language. The published practitioner literature on emotional labour in support roles, including Esther Perel’s writing on workplace dynamics, is consistent: the relationship layer compounds in value over time and degrades when replaced with a tool that has no memory of the people on either end.
There is a regulatory floor underneath the same line. The Information Commissioner’s Office is clear that routing personal data through a third party requires a data processing agreement, and the consumer tiers of typical AI tools do not provide one. A VA engaged through a UK agency operates under a DPA by default. A founder uploading client emails to ChatGPT free does not.
Where will you actually meet this trade-off?
In four places, in roughly this order. Inbox first, because volume is highest and judgement-share per email is low; AI takes the bulk of the work and a VA holds the items that need human routing. Calendar second, because scheduling is the most algorithmic surface in the business; AI handles availability matching while the VA holds exceptions and protects focus time. Transcription third. Drafting fourth.
On transcription, AI does the capture and the VA does the synthesis when the conversation actually matters. Otter or Granola will give you a clean transcript inside a minute; reading the transcript and turning it into the three things that need to happen is still a human call. On drafting, AI produces a first pass against the prompt you write; the VA edits to fit your voice and the relationship the document is going into.
Outside those four, the maths shifts. Client communication where a relationship is on the line, escalation handling for awkward situations, organisational sensing across your team, and any work where a single misjudged sentence could cost you a long-term contract, sit firmly with the human. The wider frame is in How AI changes the delegation maths; the personal-practice context is in the cluster pillar AI for your own work, not just your business.
When is the hybrid the honest answer, and what does it look like in pounds?
For a meaningful share of UK owner-operators of services SMEs in the £1m to £10m band, the honest answer is hybrid. Take a founder currently engaging a VA at 40 hours a month at £24 an hour through a UK agency: £960. Twenty hours on inbox, ten on calendar, five on drafting, five on relationship and project work. Add an AI personal stack at £150 a month and the allocation changes shape.
Inbox VA hours drop to seven (the AI tools pre-filter the routine volume). Calendar drops to three (Calendly and Reclaim handle availability matching, the VA handles exceptions). Drafting drops to two (the model produces first passes, the VA edits). Relationship and project work stays at five, because that is the half AI cannot do. New VA total: seventeen hours at £24, £408. Add the £150 stack and the combined cost is £558 a month, against £960 baseline. Roughly 40 percent off the bill, with the VA’s remaining hours concentrated on the work that compounds.
The harder question is what to do with the saved hours rather than the saved cash. Founders who pocket the cost reduction end up with a thinner VA relationship for less money. Founders who keep the VA at similar hours and redirect the freed time into proactive client outreach, weekly briefing prep, and exception handling, end up with a VA who looks more like a quasi-operator than an administrative resource. For owner-operators trying to step back from the business, the second move is usually the better one.
When should you ignore this whole question?
When you do not have a VA and you do not have the volume to justify one. A founder spending two hours a week on inbox and calendar is not the audience for a hybrid-stack post; the personal-AI-only post is what fits. The other end of the spectrum cuts the same way: a 50-person services firm with a procurement function and a working AI council is asking team-or-AI, not VA-or-AI, and a different cluster applies.
For the typical reader of this category, an owner-operator one to three people away from the work, with a VA they have used for a year or more and an AI stack they have only half-deployed, the hybrid is the live conversation. Get specific about which tasks live where. Contract the AI stack against the relevant compliance floor. Keep the VA on the work that compounds. Then book the saved hours back into the part of the week you actually want them in.
