# Start with the job you keep doing twice.

Find one repeated job, count the time it really takes, and test a smaller way to do it. A practical worksheet for owners who already have enough tools.

Published: 2026-10-09 | Updated: 2026-10-09
Author: /init editorial
Method: AI-assisted, source-checked
Canonical: https://init.news/notes/first-business-automation-bottleneck/

## The answer

Choose your first automation from a repeated bottleneck you can measure: a job that consumes time, delays a customer, or causes avoidable rework. Write down the trigger, inputs, steps, and required result before choosing software. Start with a small workflow you can check and reverse, include review and maintenance in the cost, and compare it with your existing process. The right first move may be a better template or a simple rule rather than AI.

## Pick a job you can describe in one sentence

You may be answering enquiries, chasing missing details, preparing work, and checking invoices in the same afternoon. Start by noticing where you repeat a step or lose the thread. A useful candidate sounds like: when an enquiry arrives, collect the missing details and prepare a reply for me to check. That is specific enough to improve.

Keep a short task log for a normal week. Record frequency, hands-on minutes, waiting time, corrections, and what happens when the task is late. Separate your working time from elapsed time: a two-day delay may contain only ten minutes of work. The delay may matter more to the customer than the ten minutes matter to you.

## Use five questions to choose

Question | A promising first candidate
--- | ---
Does it happen often enough? | The task recurs in normal operations, not only during an unusual week.
Are the inputs usable? | The information already exists in a form, document, inbox, or system you control.
Can you define a good result? | Someone can check whether the output is complete and correct.
What happens if it is wrong? | Errors can be caught before a customer commitment, payment, or important record changes.
Will you maintain it? | A named person can check failures and update the process when the business changes.

Then decide whether to remove the step, standardize it, use a rule, or ask AI to help. Copying a fixed field between systems usually calls for a rule. Turning a varied enquiry into a draft may suit AI assistance. An unclear offer or missing delivery policy needs a business decision before either approach will help.

## Borrow the lesson, not someone else's promised result

The 2023 working-paper version of Generative AI at Work studies assistance for 5,179 customer-support agents. The authors report different effects across workers, with stronger gains for less experienced staff. That is evidence from a particular support setting, not a forecast for your business or a reason to automate every task. [1]

The OECD's 2025 report draws on a late-2024 survey of more than 5,000 SMEs in seven countries. It records reported benefits alongside concerns about suitability, information handling, and skills. These are survey responses, not a controlled test of your proposed workflow. The practical implication we draw is to check fit and staff effort before expanding. [2]

## Count the work that remains

Use the following worksheet with your own measurements. Every figure below is illustrative, including the hourly value and software cost. None represents an init.news customer result, a vendor quote, or a savings promise.

Monthly estimate | Illustrative calculation
--- | ---
Current effort | 30 tasks × 10 minutes = 300 minutes.
Review after automation | 30 tasks × 3 minutes = 90 minutes.
Exceptions and maintenance | 30 minutes of corrections + 30 minutes of upkeep = 60 minutes.
Capacity recovered | 300 − 90 − 60 = 150 minutes, or 2.5 hours.
Assumed value after tool cost | 2.5 hours × €30 per hour − €15 software = €60.
One-time setup effort | 3 hours × the same assumed €30 = €90.

Under those assumptions, €90 divided by €60 gives 1.5 months before the estimated benefit covers setup effort. This is a planning comparison, not a cash-payback forecast. Owner time has not turned into cash unless a cost actually falls or you use the capacity for work that produces a return. Check lower task volume and longer review times too.

## Run one small pilot before adding another tool

- Write the current steps and save a few representative examples, excluding unnecessary personal or confidential information.
- Choose one existing tool or a manual template that can handle the job; check the required features before purchasing anything.
- Run the proposed process alongside the current one for a limited batch. Keep external messages and consequential changes under review.
- Log total working time, missed items, corrections, and whether customers receive a useful result.
- Name the person who handles a failure and keep the original process available.
- Keep, simplify, or stop the workflow based on the actual result. Expand only when the benefit survives review and maintenance.

Your next step is small: choose one recurring job and time the next ten examples. That gives you a starting record, not a statistically complete study. If the process becomes clearer and the total effort falls without worse outcomes, you have a reason to continue. If it does not, you have learned before building a larger system.

## Sources

- [NBER: Generative AI at Work, Working Paper 31161](https://www.nber.org/papers/w31161) — [1] Inspected the original research landing page and abstract for the 2023 working-paper version: 5,179 support agents and heterogeneous reported effects. No independent reproduction or small-business savings forecast is claimed.; checked 2026-10-09
- [OECD: Generative AI and the SME Workforce (5 November 2025)](https://www.oecd.org/en/publications/generative-ai-and-the-sme-workforce_2d08b99d-en.html) — [2] Inspected report abstract and executive summary: late-2024 survey of more than 5,000 SMEs across seven countries; self-reported benefits and adoption barriers.; checked 2026-10-09