Which office jobs can an AI assistant take on? Our count of 80 jobs in ten trades: 60 it can prepare, 20 it cannot

We listed 80 office jobs across ten trades, eight for each, and judged that an AI assistant can prepare 60 of them for a person to approve. The other 20 stay with a person: 12 because they are a decision, a piece of advice or a sign-off, and 8 because someone has to be on the phone, in the room or doing the work by hand.

The count comes from the job lists behind the ten admin scorecards on this site, as they stood on 1 October 2026. It is our judgement, job by job, not a survey of offices. This page shows the working so that you can check it or disagree.

The count by trade

Every list has eight jobs: six marked as prepared for approval and two marked as kept by a person. The hours are example weeks we wrote for each scorecard. They are not measured, and each scorecard lets you replace them with your own.

"Could be handed over" is the time on the prepared jobs with a quarter of each kept back for checking the draft, as a share of the example week. For all ten: 202 of the 316.5 example hours sit on prepared jobs, three quarters of 202 is 151.5, and 151.5 divided by 316.5 is 48%.

Read the 60 of 80 with care. Every list was written with eight jobs, so the split of six and two describes these lists. It is not a finding that three quarters of all office work can be handed over. The hours say more: where one kept job fills much of the week, such as the phone at a dental front desk or advice in a law firm, the share falls to about 40%.

What the 60 prepared jobs have in common

Sorted by the words in the job name:

• 10 are chasing, one in every trade: unpaid invoices, contractors, clients' records, suppliers, deposits, timesheets, documents from the other side, referrals and providers.
• 11 are replies to emails, enquiries, marketplace messages or reviews.
• The other 39 are mostly documents built from a record the firm already holds (quotes, invoices, notices, demand packs, CVs in the house format, report drafts) and lists of dates showing what is due, for whom and by when.

Three things are true of all 60. The work arrives in writing or sits in a record. There is a pattern for what a good one looks like. The output is a draft, a list or a pack that a person reads before it goes anywhere.

The 20 jobs that are never handed over, and why

Decisions stay because someone is answerable for them. The notes beside each job say so in a few words: the price is the owner's judgement, the sign-off is the practice's responsibility, the advice belongs to the adviser. In the three regulated trades (law, dental and financial advice) the scorecard also says who reviews every client-facing draft before it is sent.

Calls and visits stay for a plainer reason. The count is of an assistant that prepares written work for approval, and a live call cannot be approved before it is said. Phone-answering services exist; we did not count them. For shortlisting, see the ICO's rules on AI CV screening.

How we counted, so you can repeat it

1. Each scorecard lists eight jobs. Each job carries one of two marks, prepared for a person to approve or stays with a person, and a one-line note saying what is prepared or why it is kept.
2. We counted the marks. Ten lists of eight is 80 jobs; 60 are marked prepared and 20 kept.
3. We sorted the 20 kept jobs into four kinds by the reason in the note. One job fits two kinds: the trades firm's "site visits and deciding the price". We counted it once, as a decision.
4. We sorted the 60 prepared jobs by words in the job name: "chasing" for the first group; "emails", "enquiries", "messages" or "replies" for the second.
5. We added the example hours for each list and worked the percentages as shown under the table.

Every job appears on its scorecard page with its mark and note, so steps 2 to 5 can be redone from the ten pages linked above.

Using the count in your own office

Write down your own eight or ten jobs and ask four questions of each:

1. Does the work arrive in writing, or sit in a record you hold?
2. Could you describe a good one on a single page?
3. Would a wrong one be caught by a reader before it leaves?
4. Does it end in a decision someone must answer for?

Yes to the first three and no to the fourth puts a job in the prepared column. Yes to the fourth keeps it with a person.

The scorecards also ask three readiness questions: whether the way you do these jobs is written down, whether the work arrives by email, web form or an export, and whether one named person would approve drafts every working day.

Two pages go with this one: what an AI agent is, for tools that act as well as draft, and what happens when AI gets it wrong, for the check a draft should pass.

Overlyft is an AI agency that sets up AI office assistants for these ten trades, so we have an interest in this count: they are set up in the firm's own office for a fixed price agreed after an audit, the firm pays Anthropic directly for its own Claude plan, and a person approves anything that goes to a customer. This page is general information, not regulatory advice.