The Owner's Audit
A simple audit to find the tasks in your business an AI agent can take over, and the ones it shouldn't. Free worksheet inside.
Disclosure: This article contains no paid links. If we add any, they'll be marked "paid link." We're vendor-neutral, and this method works with any tool. Read our full disclosure policy.
No savings claims: We can't tell you how many hours you'll get back. That depends on your business. This audit helps you find out for yourself.
Most owners start with AI the wrong way round. They pick a tool because someone on YouTube loved it, then go looking for a problem it can solve.
A few weeks later the subscription is still running, and nobody's sure whether it saves any time.
The better order is tasks first, tools second. Find where your hours actually go, score which tasks suit an agent, test the top few, and measure. This guide walks you through that in about a week, using a spreadsheet and some honesty.
Why bother auditing first?
Three pieces of evidence point the same way.
Most small businesses haven't adopted AI yet, and that's fine. The U.S. Census Bureau's Business Trends and Outlook Survey found that between December 2025 and May 2026, overall AI use by U.S. businesses "hovered between 17% and 20%." Use grew among firms with at least 20 employees but "didn't change significantly among firms with fewer than 20 employees" (U.S. Census Bureau). You aren't late. You can still be deliberate.
The time savings are real but modest, and they vary a lot. In a nationally representative survey, researchers at the Federal Reserve Bank of St. Louis found that workers who used generative AI reported saving an average of 5.4% of their work hours, about 2.2 hours a week for someone working 40 hours (St. Louis Fed). Those figures are self-reported and come from a November 2024 survey. A later update put savings across all workers, including non-users, at 1.6% of work hours (St. Louis Fed). Useful, but not "fire half your staff."
Gains depend heavily on the task and the person. In a study of 5,179 customer-support agents, access to an AI assistant raised issues resolved per hour by 14% on average. Novice and lower-skilled workers improved by 34%, while experienced, highly skilled workers saw "minimal impact" (NBER; published in the Quarterly Journal of Economics, 2025). For an owner, the lesson is that the same tool can be a big help on one task and a waste of money on another. You need to know which is which.
And the downside of skipping this step is real. Gartner predicted in 2025 that over 40% of agentic AI projects would be canceled by the end of 2027 because of rising costs, unclear business value or weak risk controls (Gartner).
An audit is how you make the value clear before you spend.
Step 1: Log a normal week (days 1–5)
For five working days, keep a simple log. Don't change how you work. You're measuring, not optimizing.
| Time block | Task (verb + object) | Minutes | Tool or app used | Notes |
|---|---|---|---|---|
| 8:00 | Sort and answer email | 45 | Gmail | Mostly scheduling |
| 8:45 | Write quote for new client | 30 | Docs, pricing sheet | Similar to last 3 quotes |
| 9:15 | Chase overdue invoice | 15 | QuickBooks, email | Same reminder as last month |
Tips:3 items
- Write tasks as verb + object ("draft quote," "reconcile receipts," "answer pricing question"), not vague labels like "admin."
- Log in 15-minute blocks, or jot notes at lunch and at the end of the day. Perfect accuracy isn't the goal. A rough picture is.
- If you have staff, ask one or two to do the same. Owners often don't see the repetitive work that lands on others.
Step 2: Group and total (day 6, 30 minutes)
Merge similar entries into a task list and total the weekly minutes for each.
Most owners end up with 15–30 distinct recurring tasks.
Sort the list by minutes per week, largest first.
Step 3: Score every task
Score each task from 1 to 3 on five questions. Higher means a better fit for an agent.
| Factor | 1 | 2 | 3 |
|---|---|---|---|
| Frequency | Monthly or less | Weekly | Daily |
| Repeatability: are the steps the same each time? | Different every time | Mostly the same | Same every time |
| Digital inputs: is the information already in apps or documents? | Mostly in heads or on paper | Partly | Fully digital |
| Easy to check: can you verify the output quickly? | Hard to tell if it's right | Takes some review | Obvious at a glance |
| Low stakes: what happens if it's wrong? | Costly, legal or irreversible | Annoying, fixable | Trivial |
Add the five scores for a total between 5 and 15.
The "easy to check" and "low stakes" rows matter most.
A task that scores 3 on frequency but 1 on stakes, such as issuing refunds, isn't a starter task no matter how much time it takes.
OpenAI's guide to building agents recommends human oversight for actions that are "sensitive, irreversible, or have high stakes," and gives examples like "authorizing large refunds, or making payments" (OpenAI).
Step 4: Sort into four buckets
| Bucket | Typical score | What it means | Examples |
|---|---|---|---|
| Eliminate | Any | The task shouldn't exist. Fix the process, don't automate it. | Re-typing data between two apps that could be connected, a report nobody reads |
| Automate | 13–15, fixed steps | A plain automation, with or without one AI step | Logging new orders, sending booking confirmations |
| Agent-assist | 10–14 | AI drafts or prepares, a person approves | Email replies, quotes, invoice reminders, meeting prep |
| Keep human | 5–9, or stakes = 1 | Judgment, relationships, money, legal | Pricing decisions, hiring, complaints, negotiations |
Always check Eliminate first. Automating a pointless task just makes the pointless work run faster.
Anthropic's advice to agent builders applies at small scale too: find "the simplest solution possible," which sometimes means "not building agentic systems at all" (Anthropic).
Step 5: Estimate honestly and pick your top three
For each Automate or Agent-assist task, estimate:
Net weekly time saved = current minutes − (minutes the new process takes + minutes you spend reviewing it)
The review time is where most estimates go wrong.
A drafted email still has to be read. Then subtract monthly tool costs (subscriptions plus any usage-based AI fees) to see whether it's worth it.
Illustration only (made-up numbers for a fictional landscaping company): quote writing takes 150 minutes a week.
With an agent drafting from a pricing sheet and past quotes, the owner estimates 40 minutes of review and edits, for a projected 110 minutes saved.
That's an estimate to test, not a promise, and the next step is how you test it.
Pick the three tasks with the best combination of net time saved and low risk. Not ten. Three.
Step 6: Pilot for two weeks and measure
For each of your three tasks:
- Run the agent alongside your current process for the first few days, and compare outputs.
- Track three numbers: actual minutes per task (including review), edit rate (the share of outputs you had to change) and error count (anything that would have caused a problem if it had gone out).
- Keep it drafts-only and read-only until the error count stays near zero.
At the end of two weeks, make one of three calls: keep, fix or kill. If the edit rate stays high, the agent is creating work rather than saving it.
Red flags in your audit
- Customer-facing promises. You're responsible for what your AI tells customers. In Moffatt v. Air Canada (2024), a tribunal held the airline liable for wrong information its website chatbot gave a customer (CanLII).
- Tasks that need sensitive data (health, financial, HR) in tools whose data terms you haven't read.
- "It'll save 20 hours a week" from a vendor. Your audit numbers beat anyone's marketing.
Repeat every quarter
Re-run the audit, in a shorter version, every three months.
Your business changes, and AI tools change even faster.
Something you kept human in January might be ready for an agent by June, and a workflow that worked might have drifted.
Key takeaways
- Tasks first, tools second. Log a real week before you buy anything.
- Score tasks on frequency, repeatability, digital inputs, ease of checking and stakes.
- Check Eliminate before you automate.
- Count review time and tool costs in every estimate.
- Pilot three tasks for two weeks, measure edit rate and errors, then keep, fix or kill.
Next in Upgrade Your Business: The 10 Agent Workflows Every Small Business Should Run First · An Agent That Clears Your Inbox Every Morning · Upgrade Your Business: all guides · Get the free 25-workflow guide
Researched and drafted by the Agents That Earn AI agent team, with owner oversight. Research figures describe specific studies and populations, not results you should expect. This is general information, not legal, tax or financial advice.
Sources
- U.S. Census Bureau, "Large Firms With at Least 20 Employees Biggest AI Users" (May 26, 2026): https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
- Federal Reserve Bank of St. Louis, "The Impact of Generative AI on Work Productivity" (Feb 27, 2025): https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity
- Federal Reserve Bank of St. Louis, "The State of Generative AI Adoption in 2025" (Nov 13, 2025): https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025
- Brynjolfsson, Li & Raymond, "Generative AI at Work," NBER Working Paper 31161; Quarterly Journal of Economics 140(2), 2025: https://www.nber.org/papers/w31161
- Gartner (June 25, 2025): https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- OpenAI, "A Practical Guide to Building Agents": https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
- Anthropic, "Building Effective Agents" (Dec 2024): https://www.anthropic.com/engineering/building-effective-agents
- Moffatt v. Air Canada, 2024 BCCRT 149: https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.pdf