AI in a small business: Which problem should you tackle first?

You want to use AI in your business, but the first decision already feels complicated: does your team need Claude, another application or an agent? Before comparing products, ask a smaller question: Which specific step in your work should improve?

“Less admin” is not yet specific enough. “Identify missing information in a customer enquiry and prepare follow-up questions” is a useful starting point. You can observe the task, check the result and end the experiment if it does not help.

This guide helps you choose that first experiment. You do not need a company-wide AI strategy or access to every business document.

Claude is our example, not a prerequisite: we explain the practical steps using Claude from Anthropic. The underlying method—choosing suitable tasks, preparing inputs, reviewing outputs and measuring value—is provider-independent. You can also apply it with OpenAI’s ChatGPT or suitable Chinese model families such as Qwen or DeepSeek. What matters is whether the particular model reliably handles your task in an approved setup.

Product features are not necessarily interchangeable: interfaces, document handling, integrations and permissions may differ. Data processing and deployment arrangements need a fresh review for each solution too. “The best model” therefore means the tool you have evaluated for your use case, not a universal ranking.

1. Start with a bottleneck, not a tool

Look at an ordinary working week. Where does someone repeatedly collect the same information? Where do messy notes need turning into a clear draft? Which task regularly gets left unfinished?

Do not just write down “quotes take too long”. Identify which part takes time. Is information missing from the enquiry? Is the team searching for old documents? Is the pricing decision unclear? Or does someone spend too long describing a scope of work that has already been agreed?

Those causes need different solutions. A better enquiry template can reduce missing information. Clear ownership can speed up approval. AI becomes a candidate where it can do useful preparatory work within a process you understand.

My recommendation: choose a recurring task with an output you can assess, where a mistake is caught before it affects anyone outside the business. I would not begin with binding prices or customer messages sent without review.

This is consistent with Anthropic’s general recommendation to start with the simplest solution and add agentic complexity only when needed. That is the vendor’s architectural guidance, not evidence that your particular task requires AI. [1]

2. Consider three practical starting points

The following examples come from Klarwerk Service, a fictional technical-services business with twelve employees. They illustrate decisions, not a customer case study or measured productivity gains.

Prepare customer enquiries for handling

An enquiry describes a problem but leaves questions unanswered. The office needs to identify the missing information. AI could prepare a structured summary and follow-up questions from an approved sample enquiry.

A person checks the summary and the questions. Dates, prices and commitments stay outside the assignment. A form with required fields might be a simpler alternative.

Turn meeting notes into an action list

A meeting leaves the team with unstructured notes. AI could extract tasks, explicitly named owners and unresolved points.

The boundary matters: if the notes contain no deadline, the output must not invent one. A shared meeting template might already solve the problem. The experiment should help you distinguish these options.

Draft answers to recurring questions about services

The office repeatedly writes answers about what the business provides. AI could prepare a draft using a current, approved service description.

If no reliable description exists, prepare that document first. A model can help with wording, but should not decide which services your business actually offers. A few saved text snippets may be enough for a small set of standard questions.

All three examples share a useful property: you can review a limited draft before it becomes an action.

3. Prepare your first AI test in six steps

Step 1: Observe actual cases

Review several typical cases and record the steps. Measure searching, follow-up questions and checking, not just writing time. Five to ten cases may help with an initial overview; that is not a representative effectiveness study.

Step 2: Define the assignment in one sentence

For Klarwerk, it could be:

Create a factual summary of a customer enquiry and list missing information; do not add prices, dates or commitments.

“Automate enquiry handling” would be too broad. A clear boundary makes errors easier to recognise.

Step 3: Consider the simpler alternative

Ask whether a template, checklist or fixed rule could solve the same problem. A missing required field does not automatically need AI. Varied, free-text enquiries may make a model worth testing.

Step 4: Agree on inputs and data permission

Start with entirely fictional examples or data explicitly approved for the chosen system. Removing names does not necessarily remove confidential or identifiable information.

If permission is unclear, resolve that before entering the information. This test plan does not replace a privacy or contractual review.

Step 5: Describe acceptable and unacceptable results

For enquiry summaries, acceptable means accurately describing the request, flagging missing details and inventing nothing. Adding prices or service commitments would be unacceptable.

Write these criteria down before the experiment. Anthropic also recommends specific, measurable success criteria as the starting point for evaluating LLM applications. [2]

Step 6: Assign review and evaluation

A knowledgeable person reviews drafts and records what was useful, what needed correction and how long the whole task took. Also specify which errors should stop the test.

The goal is not a perfect AI demonstration. It is a decision: continue testing, narrow the assignment or work without AI.

4. How Claude can help with the first experiment

For this starting point, I would use Claude as a tool for a reviewable draft, without broad access to business systems. Use only appropriate sample data and your approved setup.

An initial instruction could look like this:

You help prepare incoming customer enquiries.
Use only the fictional example below.

Produce:
1. a factual summary,
2. a list of details explicitly provided,
3. missing information and suitable follow-up questions.

Do not invent information. Flag uncertainty.
Do not make price, scheduling or service commitments.

Example enquiry:
[Insert a completely fictional enquiry here.]

Check more than the wording. Is something important missing? Has a guess become a fact? Is each follow-up question necessary? That review remains necessary even when the instruction says “do not invent information”.

Later, recurring instructions and approved documents could be collected in a Claude Project. Anthropic describes Projects as workspaces with their own context and instructions. That does not replace permission to use the documents or expert review of the answers. [3]

How to transfer the experiment: start with the same fictional enquiries, instruction and review checklist in the alternative solution. Adapt the instruction if necessary, record the change and compare again. Assess difficult cases and rework, not just the first fluent draft. ChatGPT also offers Projects for recurring context, for example, but this does not establish identical features or data terms. [4]

5. Before and after: Compare the whole task

Before: a person reads the enquiry, identifies missing information and writes follow-up questions.

With AI in the test: that person enters suitable sample text, checks the draft, corrects it and uses only the useful parts.

A benefit is established only if the second approach works better overall while maintaining adequate quality. A faster draft is not sufficient. Preparation, checking and rework belong in the comparison; a later rollout also needs to account for setup, subscriptions and maintenance.

Avoid this mistake: if outputs sound good but repeatedly contain false assumptions, better wording is not enough. Narrow the task, improve the input or decide against that use of AI.

Worksheet: Your first AI experiment on one page

Copy these questions into a document and answer them before selecting a tool:

Task: which single step should improve?
Problem: what makes it slow, unreliable or difficult today?
Alternative: what could a template, checklist or fixed rule solve?
Inputs: what suitable sample data is available?
Output: what exactly should the result contain?
Boundary: what must the AI not decide or do?
Review: who can recognise a correct or incorrect result?
Decision: what must the experiment demonstrate for you to continue?

Conclusion: Choose a useful task before choosing the setup

Your first useful AI application need not be large. A limited task with suitable inputs, a reviewable output and clear ownership gives you a stronger basis for deciding than a list of purchased tools.

Once you have a candidate task, check its prerequisites: Is your business ready for AI? A practical self-check. You do not need to prepare the whole company before beginning a small, controlled experiment.

Ask your AI tool / LLM / agent

Use an abstract description without confidential business information:

“Break this workflow into steps. Where might a template be sufficient instead of AI?”
“Help me turn this into a limited test assignment. Which decisions must stay with a person?”
“Which failure cases and review criteria should I write down before testing?”

AI in a small business: Which problem should you tackle first?

In this article

1. Start with a bottleneck, not a tool