Splitting Work Between AI and a Part-Time Assistant

The split that works is not "AI does the boring parts, the assistant does the rest". It is decided by one question asked task by task: does this task have a verifiable right answer, and who is cheaper at verifying it? AI is fast and needs checking. A part-time assistant is slower and mostly does not. Once you sort work by verification cost rather than by difficulty, the boundary stops being a judgement call and starts being obvious.

This matters because most small teams get it backwards. They give the assistant the repetitive work, because that feels like what an assistant is for, and give AI the interesting work, because that feels like what AI is for. Then they spend their own evenings checking the interesting work.

Sort every recurring task into four boxes

Take your actual list of recurring work and put each item into one of these.

Three of those four boxes are uncontroversial. The interesting one is top-right: work AI does well but where checking the output is costly. That is where the assistant earns the most, and it is not by doing the task. It is by owning the check.

A worked example. Extracting line items from fifty supplier invoices into a spreadsheet is something AI does well. Verifying it is expensive, because a wrong number looks exactly like a right number. So AI extracts, the assistant reconciles totals against the invoice PDFs, and the assistant owns the number that goes to your accountant. The assistant is doing twenty minutes of checking instead of three hours of typing, and the error rate is lower than either would achieve alone.

The tasks that should never go to AI first

Not because AI cannot do them, but because the failure is asymmetric and lands on a person.

• Anything a customer receives without a human reading it. Not a hard rule about quality, a hard rule about liability. One wrong refund amount in an automated email costs more than the time saved on a hundred.

• Anything involving a relationship that is currently strained. A late-paying client, an unhappy customer, a supplier you are renegotiating with. The drafting is fine. The sending is not.

• Anything where the right answer depends on context nobody wrote down. Your assistant knows that this particular customer always orders on account and hates phone calls. That knowledge is not in any system, and it is the whole reason the assistant is valuable.

That last one is worth dwelling on, because it points at what actually changes when you add AI to a small team. The assistant's value shifts away from execution and toward context. The things they know that are not written down become the scarce input.

Give the assistant the AI, not the leftovers

The mistake that wastes the most money is running the assistant and the AI as separate channels, both reporting to you. You become the integration layer, which is the most expensive possible use of your time.

The alternative is to hand the assistant the tools directly. They run the extraction, the drafting, the summarising. They own the output. You see the finished thing.

This requires two things from you.

A written standard for what "done" means. Not a style guide, a checklist. "Every invoice line reconciles to the PDF total. Any invoice that does not reconcile goes in the queries tab with a note." Ten lines. Without it, the assistant does not know what they are allowed to accept from the model, and defaults to either checking nothing or checking everything.

A budget and a limit. Give them their own account rather than sharing yours. It keeps your context clean, makes spend legible, and means offboarding is one click. If you have never thought about what happens when someone leaves, offboarding an employee from AI tools is a fifteen-minute read that saves an awkward month.

What to write down before the first week

Two documents. Both short.

The first is a context file: the things about your business that AI cannot infer and a new assistant would not guess. Your customers' names and quirks, your pricing rules and their exceptions, the tone you use with each type of client, the three things you never say in writing. This file is what makes AI output usable rather than generic, and it is the asset that compounds. Our guide to giving AI context about your business covers what to put in it.

The second is a one-page escalation rule: what comes to you, and what does not. Be specific about amounts and about named customers. "Anything over 500 and anything from the three accounts on this list" beats "use your judgement", because it is checkable and because it lets the assistant act without asking.

A realistic week

For a solo business owner with an assistant at ten hours a week, a shape that works:

• Monday, 2 hours. Assistant runs the weekly admin batch: invoices extracted and reconciled, inbox triaged into three folders, calendar conflicts flagged. AI does the first pass on all three, assistant owns the output.

• Midweek, 4 hours. Customer-facing work. Assistant drafts with AI, sends the routine ones, queues anything on the escalation list for you. You review a queue, not an inbox.

• Friday, 2 hours. The thing you keep not doing: following up quotes, updating the website, chasing a late payment. AI drafts, assistant sends.

• Remaining 2 hours. Buffer, and updating the context file with whatever came up. Protect this. It is the part that makes next month cheaper.

Note what is not in there: the assistant spending hours retyping things. If that is still happening after month one, the split is wrong.

When the answer is "neither"

Sometimes the honest answer is that the task should stop existing. A weekly report nobody reads does not need automating or delegating. Before you split work between AI and a person, delete the third of it that only survives because it was always done.

The same applies in reverse when the volume grows. If the assistant's checking work becomes a full-time job, the process needs fixing rather than staffing. Which tasks to automate first is the framework for that call, and getting your team to actually use AI covers the adoption side, which is where most of these arrangements quietly fail.

FAQ

Should I tell my assistant they are using AI to do their job?

Yes, explicitly, and frame it as tooling rather than as a test. The arrangement only works if they own the output, and people do not own output they were tricked into producing.

Does this mean I need fewer assistant hours?

Usually not fewer hours, different hours. Most small businesses find the same hours cover roughly twice the work, and the constraint moves from admin capacity to your own decision-making capacity.

What if my assistant is not comfortable with AI tools?

Start with one task, one tool, one written checklist, and let them see the output improve. Comfort follows a few reps of the tool being useful, not a training session.

Who is responsible when the AI gets something wrong and it goes out?

You are. That is not a technicality, it is the reason the escalation rule exists. Who is responsible when AI makes a mistake covers how this plays out with customers and contracts.