If employees use AI but the admin queue keeps growing, examine the whole workflow. A quick draft will not resolve missing information, repeated data entry or an approval that waits three days.
Someone is writing emails faster. Someone else has a favourite prompt for proposals. Yet customers still chase responses and the owner still clears a queue each Friday.
That is a useful signal. The team may have improved individual tasks while the way work moves between people has stayed the same.
Microsoft’s 2026 Work Trend Index describes a gap between employees’ AI capability and the organisational systems supporting it. Its research uses Microsoft product signals and a survey of AI users across ten countries. It does not establish what a particular Australian business will save.
You can investigate your own business without commissioning a large review. Start with five working days and one recurring job.
Follow a job from arrival to completion
Choose a defined process such as responding to a quote request, opening a client file or preparing a weekly report. Avoid a category as broad as “administration”.
Write down its start and finish:
Start: a new quote request reaches the shared inbox.
Finish: a checked quote is sent and recorded against the enquiry.
Then sketch the actual steps. Include the informal ones: asking the owner which price applies, finding the latest template, copying details into a spreadsheet and chasing the person who knows the answer.
The Lean Enterprise Institute’s explanation of value-stream mapping distinguishes the time spent on individual operations from the time something takes to travel through the complete process. That distinction is useful for office work too. Record both time actively spent and time waiting.
Use a five-day observation sheet
Ask the people doing the work to record a manageable sample. Explain that you are testing the process, not ranking employees by speed. Use case references rather than customer names in the review sheet.
| Field | What to record |
|---|---|
| Case reference | A non-identifying job ID |
| Arrival and completion | When the job entered and left the process |
| Active work | Minutes spent reading, preparing, entering and checking |
| Waiting | What the job was waiting for and for how long |
| Missing input | The specific detail someone had to chase |
| Correction | What had to be changed and why |
| AI use | Which step used an approved AI tool |
| Next owner | Who was responsible after each handoff |
Do not upload the observation sheet to an AI service unless the data and service are approved for that use. A shared spreadsheet is enough to begin.
On day one, agree the definitions and record a few cases together. On days two to four, let staff add observations during normal work. On day five, review the cases as a team. Include difficult cases rather than selecting only the smooth ones.
For a low-volume process, extend the observation period. Five days is a practical starting point, not a statistically reliable sample for every business.
Separate four different problems
Classify each delay before discussing a solution.
Missing information. Staff cannot finish because the request lacks a site address, required document or clear scope. Improve intake and validation.
Repeated handling. Someone types the same approved details into three systems. Check whether the systems can share those fields safely.
Judgement. A person needs to interpret an unusual case or approve a commitment. Give them a complete review pack and an agreed response window.
Unclear responsibility. Everyone can see the job, so everyone assumes someone else owns it. Assign an owner and backup at the point of arrival.
AI may help extract details or prepare a review pack. It cannot decide your approval authority or make an absent customer provide information.
Worked example: a faster draft, the same late quote
Consider a fictional maintenance business. Its quote process takes 35 minutes of staff handling and two working days of waiting. AI reduces drafting from 15 minutes to five.
The team has released ten minutes per quote. However, the quote still waits because a supervisor must confirm access requirements and the owner approves every price.
A useful next experiment would be to collect access details at intake and give the owner two scheduled review windows. Standard cases could follow a documented approval policy if the owner authorises it. Unusual work would still be escalated.
The result must be measured. It would be misleading to describe ten minutes released as a wage saving if staffing costs have not changed. It may instead create capacity for more enquiries or better follow-up.
Change one handoff before adding more software
For the most common problem, agree:
- The missing input or repeated action to remove.
- The employee responsible for the next step.
- The approved source, template or rule they will use.
- The cases that require a person’s decision.
- The measure that would show improvement.
Test the change on comparable work. Keep the previous process available while the team learns. An improvement that only works when its designer is present is not ready for routine use.
After a fortnight, compare handling time, elapsed time, corrections and first-pass acceptance. Check whether the work became harder for the next person. Faster intake is not an improvement if it doubles the review queue.
Make the useful change repeatable
Name one workflow owner. Give them responsibility for the approved instructions, source documents, exception list and review of results. Involve the employees who know where the work becomes awkward.
The immediate output should be one better way of completing a recurring job. It might need AI, a simple form, a clearer approval rule or a small integration.
If you have not selected a process yet, use the first AI workflow guide. Find more setup patterns in the AI workflow guides. If you already have a process but cannot see where the time goes, an AI assessment can map the work with your team and identify the next useful change.
