Repetitive, rule-based work
- Examples
- Data entry, scheduling, reminders, sorting, first drafts, routine answers.
- Why
- Same steps every time, easy to check, low judgment.
- Still needs
- A person to set it up and spot-check it.
■ · People & Tech
Technology and AI: tools do the routine, people do the judgment
Software and AI can take over a lot of repetitive work in a small business. But a computer is only as good as the person using it. The goal is a business where technology handles the routine and people spend their time on what machines can't do.
What to hand off
Much of the repetitive work in a small business can now be handled by software, and some of it by AI tools that you set up and train on your own processes. The best candidates are tasks that follow the same steps every time and are easy to check.
- Data entry and sorting: pulling numbers from receipts and invoices, categorizing transactions.
- Scheduling and reminders: booking, follow-ups, payment reminders.
- First drafts: emails, proposals, reports and social posts that a person then reviews.
- Answering common questions: hours, pricing, how-to questions, with a person handling anything unusual.
- Summaries: meeting notes, long documents, client histories.
How to "train" AI for your business
For most small businesses, training AI doesn't mean building your own model. It means giving an AI tool the right instructions and examples so it does a task your way.
- Write down the process first. If you can't explain the task step by step, AI can't do it either. This is where your written procedures pay off.
- Give it examples of good work. Past emails, finished reports or correct entries show it what "right" looks like.
- Start with one small task. Test it on real work, compare with what a person would do, and fix the instructions.
- Keep a person reviewing the output, especially anything that goes to a client, involves money, or touches private information.
- Protect client data. Check what each tool does with the information you put in before you use it for anything confidential.
■ The human edge
You still need people. Better ones, doing better work.
A computer is only as good as the person using it. Software follows instructions, and AI can produce confident answers that are wrong. Someone has to set it up, check its work, catch the mistakes and take responsibility for the result. The goal of technology isn't a business with no people. It's a business where people stop doing work a machine can do, and spend their time on what machines can't.
Judgment, trust and relationships
- Examples
- Client relationships, hard conversations, final decisions, ethics, unusual problems, selling.
- Why
- Clients trust people, not software. Responsibility can't be handed to a tool.
- Leverage
- Let technology do the prep so people spend more time here.
The research
What studies show about AI at work
These are real studies with real numbers, from specific settings. They won't transfer exactly to your business, but they show where the gains tend to be and where the risks are.
AI helps most on routine work
In a study of 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for newer, less-skilled agents, with little effect on the most experienced.
Brynjolfsson, Li & Raymond, NBER Working Paper 31161Writing tasks get faster
In an experiment with 453 college-educated professionals, ChatGPT cut the time to complete writing tasks by about 40%, and evaluators rated the quality 18% higher.
Noy & Zhang, 2023, SciencePeople still have to check
In a study of 758 BCG consultants, those using AI finished suitable tasks about 25% faster with higher quality. But on a task outside what the AI handled well, they were about 19 percentage points less likely to get the right answer than consultants without it.
Dell'Acqua et al., 2023, Harvard Business School Working Paper 24-013The pattern across all three: AI helps most on routine, well-defined work and with less experienced staff, and it can make people worse on work it doesn't handle well. That's why a person always checks.
Before you buy any tool
- Run the Dollar Test. Hours saved only count if they go to paying work.
- Start with one task, not a whole system. Prove it on a small scale, the way the pivot timeline does in days 31–60.
- Pick who checks it. Every automated task has a named person who reviews the output.
- Set a tripwire. If it isn't saving the time you expected by day 60, drop it.
- Plan the training. A tool nobody knows how to use saves nothing.
When technology changes which roles you need, see the Role Review on the Pivot page.
Go deeperCase file, step-by-step how-to, common mistakes and questions to work through
Case file: Training an AI assistant on one task
A fictional example
Hannah runs an insurance agency with five staff. Her team spends hours each week writing renewal reminder emails, each one slightly different depending on the policy and the client. She wants to try an AI tool, but she's worried about mistakes and client privacy.
She started small, with one task: first drafts of renewal reminder emails. She wrote down the process step by step: what each email must include, the tone the agency uses, and what must never be said, such as any promise about coverage or price that hadn't been confirmed. She collected five of the agency's best past emails as examples, with all names and policy numbers removed.
She gave the AI tool clear instructions: "Write a friendly renewal reminder using only the details provided. Don't promise coverage or pricing. Keep it under 150 words. End by inviting the client to call with questions." Her team tested it on 20 real renewals, using only non-confidential details like renewal dates and policy type. A staff member reviewed every draft before it was sent.
About three-quarters of the drafts needed only small edits. The rest had problems: one hinted at a price that wasn't confirmed, and two sounded stiff. Hannah tightened the instructions and added an example of the problem to avoid. The second round was better.
After two months, writing renewal emails took a fraction of the time it used to, and a person still read every email before it went out. She checked her AI tool's settings and her industry's privacy rules before allowing any client details at all. Her conclusion: "The tool writes the first draft. My people still own the last one."
How to bring in an AI tool safely, step by step
- Pick one repetitive, low-risk task with a clear right answer, like first drafts, summaries or sorting.
- Write the process down as if training a new employee.
- Collect three to five examples of good work, with private details removed.
- Write clear instructions: what to do, what never to do, length and format.
- Check data rules before sharing any client information: the tool's settings, your contracts and any rules in your industry.
- Test on 10 to 20 real examples, with a person reviewing every output. Track how many needed changes.
- Improve the instructions based on the mistakes, then retest.
- Set a tripwire and a reviewer: who checks the work, and what result would make you stop.
A starter instruction for any AI task
"You're helping [business] with [task]. Follow these steps: [steps]. Here are examples of good results: [examples]. Never [things to avoid]. If you're unsure about anything, say so instead of guessing. Keep it under [length]."
Common mistakes and how to fix them
| Mistake | Fix |
|---|---|
| Starting with a big, important task | Start small and low-risk. Build trust in the tool step by step. |
| Letting AI output go to clients unchecked | A person reviews anything client-facing, financial or sensitive. |
| Pasting private client data into a tool without checking | Check the tool's data settings and your obligations first. |
| Vague instructions, then blaming the tool | Give clear steps, examples and limits, like you would for a new hire. |
Questions to work through
- Which task does your team repeat most often with only small variations?
- Could you write that task as clear steps today? If not, that's the first job.
- What information must never leave your business? Does everyone know that?
- Who on your team would be best at checking and improving AI output?
- What would your team do with the hours saved?