← Back to Blog

How to Find the Tasks in Your Business Worth Automating With AI (A Two-Week Audit)

One team on this site reports cutting 96 hours of weekly busywork with AI agents. Before you build anything, here is how to find your own highest-value task in two weeks, score it, and pilot it for under $50 a month.

Strategy8 min read•

Most owners automate the wrong task first

A team documented on this site reports that AI agents took 96 hours a week of repetitive work off their plate, which they put at 2.4 full-time employees. Support tickets went from 45 hours a week to 12, with about 70 percent of questions deflected by a chatbot. Content production went from 25 hours to 8. Code review saved 14 hours a week. One caution before you copy it: the three line items in the write-up add up to less than the 96-hour headline, so treat the total as rough. The direction is clear, though, and so is the lesson. They found the work eating the hours first. The tool came second.

Most small business owners do it the other way around. They watch a demo of a voice agent or an email assistant, build it, and discover the task they automated took 20 minutes a week. This guide is a two-week audit that fixes that. It needs a spreadsheet, a timer and one Claude or ChatGPT account.

The two-week audit Days 1 to 10 Log every task Minutes and frequency Whole team, no editing Day 11 Score each task Hours a month Rules, cost of error Day 12 Pick exactly one Write the steps down Set a pass mark Days 13 to 14 Pilot with a human checking every output Most owners guess wrong about which task is the real time sink. The log decides.

Why a time log beats your gut

Ask an owner what eats the week and you hear "email" and "admin". Ask them to log it and the answer usually moves. The task is smaller and more specific: re-typing the same booking details into three systems, chasing the same four documents from every new client, answering the same pricing question from the website.

The consultant on this site who runs eleven Claude workflows across Gmail, Calendar, Notion and Slack describes the pattern in the admin work of a solo business: lots of small recurring jobs, each too minor to notice and together a large share of the week. The agent that handles one of them well saves little. The audit's job is to find the few that add up.

A worked example: a five-person HVAC office

Picture a five-person HVAC company with one office manager. For ten working days, everyone logs each task they do in a shared sheet: what it was, how many minutes, whether they had done it before. By day ten the sheet has about 400 rows. The office manager guesses that scheduling is the problem. The log says otherwise.

Scheduling is 6 hours a week. But two things nobody mentioned show up: chasing customers for their equipment model numbers and photos before a quote visit, 5 hours a week; and re-entering completed job details from tech texts into the invoicing tool, 7 hours a week. The second one is rule-based, happens every day, and has a clear right answer: the tech's text becomes an invoice line. That is the one to automate first. The numbers are made up for illustration, but this shape is what real logs look like.

Step by step: run the audit

Step 1: Set up the log

Make a shared sheet with five columns: date, who, task, minutes, and a yes or no for "did this exact thing before". Tell everyone to log tasks over five minutes as they finish them, not at the end of the day from memory. Do it for ten working days. Do not clean up the entries.

Step 2: Cluster the rows

On day 11, paste the whole sheet into Claude or ChatGPT and ask it to group the rows into recurring tasks, then total hours per month for each group. Tell it to list its grouping so you can correct it, and spot check three groups against the raw rows. The point is a ranked list of recurring tasks by monthly hours.

Step 3: Score the top ten

For each of the top ten, give a score from 1 to 5 on four questions. How many hours a month does it take? Could you write the steps on one page? Is the input digital already (email, forms, texts), or is it paper and phone? What happens if the output is wrong: a typo, or a lost customer? Multiply the first three scores and divide by the fourth. High volume, clear rules, digital input and low cost of error float to the top.

Step 4: Pick exactly one and write the steps

Pick the highest scorer, and only that one. Write the current process as numbered steps on one page, including the odd cases: what you do when the customer does not reply, when the photo is blurry, when the amount does not match. If you cannot write the steps down, the task is not ready for automation and you should pick the next one.

Step 5: Choose the lightest tool that works

Match the task to the tool. A pure writing or summarizing job with a person in the loop may need only a saved prompt in Claude or ChatGPT. A trigger-based job (a new email, a new form, a new row) fits n8n or Make. A job that needs a phone call answered needs a voice platform. Start with the cheapest option that could work; you can move up later.

Step 6: Set a pass mark before you build

Decide in advance what success means: for instance, 90 percent of outputs need no edit, or time on the task drops by half. Without a pass mark, a pilot drifts forever.

Step 7: Pilot for two weeks with a human checking everything

Run the agent beside the person doing it. A person reviews every output before it reaches a customer or a ledger. Log how often they change it and why. At the end of two weeks, compare against the pass mark. If it hit, reduce the review to a sample. If not, fix the prompt or drop the task and pick the next one.

What it costs

Cheaper than most owners expect, as of November 2026. Anthropic lists Claude Pro at $17 a month on annual billing or $20 billed monthly, and Team at $20 a seat on annual billing or $25 monthly. n8n Cloud lists Starter at 20 euros a month on annual billing with 2,500 workflow executions. Self-hosting n8n costs only the server. The listings on this site put typical startup cost for a single-task agent at $50 to $200 a month, and the 96-hour case sits in that same band.

The bigger cost is your time. Budget about four hours for the log setup and clustering, two to three hours to write the process and the prompt, and an hour a week during the pilot. Skipping the audit does not save that time. It just moves it into building the wrong thing.

When the audit says do not automate

Sometimes the honest answer is no. Tasks that happen a few times a month are rarely worth it. Tasks where a wrong answer is expensive, such as quoting, legal wording, or money leaving the account, need a person in the final step, which cuts the saving. And some of the time is relationship time: the call where a customer decides to trust you is not a task to remove. If your top-scoring task saves under five hours a month, fix the process with a template instead and move on. Also be wary of automating a broken process. If the HVAC office re-enters job details because the tech texts are inconsistent, the cheaper first fix is a one-line template for the techs, and the agent only has to handle the cleaned-up version.

Frequently Asked Questions

How do I know which tasks to automate with AI?▾

Log every task for ten working days with minutes and frequency, then score the recurring ones on volume, how clearly the steps can be written, whether the input is digital, and the cost of a mistake. Automate the highest-scoring task first and pilot it with a person checking the output.

How long does a task audit take?▾

Two weeks of passive logging, plus about half a day to cluster and score the results with an AI assistant. The logging is the slow part, and it is where the surprises come from.

What tasks are best for AI agents in a small business?▾

Repetitive, rule-based work with digital input: sorting and replying to common emails, moving details between systems, drafting standard documents, answering the same customer questions, and following up on leads. The 96-hour case on this site was mostly support, content and code review.

Do I need to be technical to automate a task?▾

Not for the first one. A saved prompt in Claude or ChatGPT needs no code, and visual tools like n8n or Make handle trigger-based jobs. Phone and system integrations get harder, and for those you may want a freelancer for the first build.

How much does it cost to automate one task?▾

Builders documented here report $50 to $200 a month for a single-task agent, and some run on far less. Check current pricing for the tools you pick, and add the hours you will spend on setup and review.

What if the pilot fails?▾

That is a result, not a disaster. Look at the reviewer's corrections to see whether the prompt, the input or the task itself was the problem. If the task was too variable, drop it and pick the next one on your list.

Keep Reading

Explore Related Niches

Related Strategies

Marketing & SalesProfessional ServicesInternal Tool

A sales team of 10 SDRs and AEs was replaced by 20 AI agents managed by 1.2 humans, maintaining the same business performance.

Advanced$200+/moVery High ($10K+/mo)
View Strategy
Dev ToolsProfessional ServicesSaaS Product

A non developer learned to code using AI tools and built a portfolio of SaaS products generating $28K per month.

BeginnerUnder $50/moVery High ($10K+/mo)
View Strategy
Dev ToolsProfessional ServicesSaaS Product

A SaaS playbook combining Claude Code for development and Google Search for distribution reached $40K MRR in 7 months.

IntermediateUnder $50/moVery High ($10K+/mo)
View Strategy

Keep Exploring the Directory

Browse more case studies, compare tools, or dig into the full tool stack behind these workflows.

Want more strategies like this?

Get weekly AI agent case studies in your inbox.