Your first week using AI for admin work

Most small businesses that try AI for admin work start with whatever task is most annoying that day, paste something into a chat window, get a half-useful answer, and stop. A week later nobody can say whether it saved any time.

This plan fixes that by doing less. You pick three recurring tasks, time them the way you do them now, write one prompt for each, check every output by hand, and time them again. By Friday you have three saved prompts and a number for each task: minutes before, minutes after, including the checking. That number tells you what to keep.

You need a chat assistant that accepts file uploads (Claude, ChatGPT or similar) on a business or team plan whose data terms you have read, a timer, and about 30 minutes a day on top of the work itself.

Pick tasks the AI can do and you can check

A good first task has four traits:

• It repeats. Weekly or monthly. A one-off task gives you one data point and no reason to save the prompt.
• The input already exists as a file or text. An export, an email thread, a PDF, a set of notes. If you have to type the input from memory, the AI saves you little.
• You can tell a right answer from a wrong one in a few minutes. The checking step has to be faster than doing the work, or the math never works out.
• A mistake gets caught before it costs money. Drafts that a person reviews before sending are fine. Anything that moves money, changes a system of record, or goes straight to a customer is not a week-one task.

Good first candidates: summarising a weekly sales or job report for the owner, drafting replies to routine supplier or customer emails, turning meeting notes into action lists, reformatting an export into the layout your manager wants, writing first drafts of job postings or policy updates, and categorising a list of expenses for someone to review.

Leave for later: anything that posts entries in your accounting system, runs payroll, pays vendors, or sends messages without a person reading them first.

The task-picking table

List every recurring admin task you can think of (aim for ten), then fill in this table. Score each column 0 to 2 and add them up. Pick the three highest totals, breaking ties in favour of the task that takes longest today.

Scoring guide: 2 means yes without qualification, 1 means partly, 0 means no. The vendor payment run above scores 0 on "low cost if wrong" because a wrong payment leaves the building; it stays off the week-one list even though it repeats.

The day-by-day plan

Monday: time the work as it is

Do each of your three tasks the normal way and time it from the moment you open the first file to the moment the output is ready to hand over. Write the time down. If a task is monthly and isn't due this week, time the last instance from memory and mark it as an estimate, then replace it with a real time next month.

Save a copy of Monday's input and your finished output for each task. You will use them on Wednesday to judge the AI's version.

Timing log:

Task:
Date:
Start time:          End time:          Minutes:
Input files used:
Output produced:
Anything that slowed you down:

Tuesday: write one prompt per task

Every prompt follows the same five parts. Writing them out takes about 15 minutes per task the first time.

CONTEXT
I am the [role] at [company], a [what the business does] with [size].
This task is [what it is] and the result goes to [who] for [what they do with it].

INPUT
I have attached [file name / pasted text]. It contains [what the columns or sections are].

OUTPUT
Produce [format: a table / an email draft / a bullet summary] with [exact sections or columns].
Keep it under [length]. Use [tone].

RULES
- Use only the figures in the attached file. Do not estimate or fill gaps.
- If something needed is missing, write "MISSING:" and what is missing.
- [Any house rules: date format, currency, names we never abbreviate, etc.]

FLAG FOR ME
At the end, list anything that looked unusual or that you were unsure about,
with the row or line it came from.

The RULES and FLAG FOR ME sections do most of the work. They turn a confident-sounding draft into one that tells you where to look.

Save each prompt in a shared document named after the task, with today's date. That document is the asset you are building this week.

Wednesday: run it on Monday's input and compare

Give the AI Monday's input and your prompt. Put its output next to the version you made on Monday and go through it line by line.

Checking routine:

• Every number in the output appears in the input, or is a calculation you can repeat in a minute
• Totals match the input's totals
• Nothing in the output was invented: names, dates, amounts, commitments
• Every item you included on Monday is present, or its absence is explained
• The flags at the end point to real issues, and you agree with them
• The format is one you could hand over without reworking

Write down each error and which part of the prompt would have prevented it. Most first-run errors come from missing context (the AI did not know who the reader was) or a missing rule (it rounded figures, or used a different date format).

Thursday: fix the prompt and run it on fresh input

Edit each prompt to cover Wednesday's errors. Then run it on this week's real input, the one you would otherwise do by hand. Time the whole thing: uploading, prompting, reading the output, fixing anything wrong, and getting it ready to hand over. Checking time counts. If you leave it out, the "after" number is fiction.

Friday: measure and decide

Fill in the results table and make one of three calls for each task.

Decision rules:

• Keep if the after time is clearly lower and Thursday's errors were small enough that checking caught them all.
• Fix and retry if the time saving is there but you found an error that checking only caught by luck. Tighten the rules section and run another week.
• Drop if checking took about as long as doing the work, or if the errors were the kind that would reach a customer. Some tasks are a poor fit, and finding that out in a week is a good result.

A worked example

Harbor Tile Supply is a fictional 14-person distributor. Dana, the operations manager, picked the weekly sales summary, supplier email replies and the monthly expense categorisation draft.

On Monday the sales summary took 45 minutes: open the sales export, build a pivot by rep and product line, compare to last week, write five bullets for the owner. Her prompt asked for the same five bullets, a table of week-over-week change by rep, and a flag list.

On Wednesday the AI's version matched her totals but described one rep's numbers as "a strong week" when the increase came from a single large order. She added a rule: "If one order is more than 30% of a rep's weekly total, say so instead of calling it a trend."

On Thursday the summary took 15 minutes including checking. The supplier replies went from 20 minutes to 12. The expense categorisation went from 90 minutes to 70, with three miscategorised lines caught in review; she kept it on "fix and retry" and added the company's category list to the prompt. Across a month that came to roughly two and a half hours back, with each figure measured rather than guessed.

Rules that keep the checking honest

• The person who used to do the task does the checking. They know what wrong looks like.
• Never check by reading the output alone. Check against the input file.
• When the output is right ten weeks in a row, keep checking anyway; reduce it to totals and flagged items, but don't skip it.
• Keep the prompt document current. When you fix an error by hand, update the prompt the same day so the fix sticks.

Week two

Keep your three tasks running and add one more from the table. Share the prompt document with one colleague and have them run a prompt on their own input; if they get a usable result without asking you anything, the prompt is good enough to become a team standard.

Do it with Merit AC: the AI Skills Academy course walks through prompt writing and output checking for business tasks one lesson at a time.