How to sell AI automations to small businesses

What AI automation should you sell first?

The best first offer handles a repetitive task that already costs the business time or lost sales. Good targets have a clear trigger, predictable inputs and an action that can be checked by a person before anything important happens.

Start by interviewing the owner about work that happens every day or every week. Ask, “What do you copy between apps?”, “Which enquiries wait too long?” and “Where do staff chase missing information?”

Do not start with a general AI chatbot because it sounds impressive. A small workflow that sorts enquiries, drafts a reply and updates a customer record is easier to test, explain and price.

Look for at least 20 similar tasks per month. If the task happens only twice a year, the build may cost more than it saves.

How to sell AI automations to small businesses

Owners usually care about missed enquiries, slow service, staff time and errors. Describe the current problem, the new workflow and the expected business result in ordinary language.

A useful offer sounds like this: “I will connect your website form, email and customer system so every new enquiry is logged, sorted and given a draft reply within five minutes. Your team approves the reply before it is sent.”

That is stronger than saying you build AI agents. It tells the buyer what changes, which systems are involved and where a person stays in control.

1. Pick one type of customer. Choose a group you understand, such as local trades, agencies or property managers. A narrow focus makes your examples and outreach more relevant.
2. Find one repeated problem. Speak to five owners or office managers. Write down the apps they use, task frequency, average handling time and common mistakes.
3. Map the current workflow. Record the trigger, every handoff, required data and final action. Mark sensitive data and any step that needs approval.
4. Build a simple demonstration. Use invented data, not a prospect’s private customer records. Show the trigger, AI draft, approval step, error route and final record.
5. Offer a paid pilot. Limit it to one workflow, one team and a short test period. State what success looks like before you build.
6. Review the evidence. Compare handling time, response time, error rate or booked appointments with the old process. Use the result to decide whether to expand, revise or stop.

Your demonstration should take five minutes or less. Show the old process first, then run one realistic example through the proposed system.

Never claim guaranteed revenue or perfect accuracy. AI can misread requests and write incorrect text, so sell a controlled process with monitoring and human checks.

How should you package and price the service?

Package the work into discovery, build, launch and support. This stops a small request from growing into unlimited integrations, revisions and staff training.

A beginner might charge around $300 to $750 for a tightly limited pilot, around $1,000 to $3,000 for a tested workflow across two or three apps, and around $100 to $500 per month for monitoring and minor changes. These are practical starting ranges, not fixed market rates.

Charge more when the process has many branches, messy data, several users, custom code or sensitive information. Charge less only when the scope is genuinely small, not because the buyer calls it easy.

List software separately. At the time of writing, entry plans for tools such as Make, Zapier and n8n are often around $10 to $30 per month, while OpenAI and Anthropic API use is usually charged by consumption.

Your proposal should state the apps, workflow limits, test cases, approval points, data access, delivery stages and support period. Also state what is excluded, such as custom software, data cleanup, extra departments and third party subscription fees.

Where can you find your first prospects?

Start with businesses where you can see the process and reach a decision maker. Existing work contacts, local business directories, trade groups and businesses you already buy from are better starting points than sending thousands of generic emails.

Research each prospect before contacting them. Look at their enquiry form, booking process, response emails and job adverts for admin roles, but do not submit fake enquiries or collect private data.

Send a short message built around an observed process. For example: “I noticed quote requests go through your website form. I build small workflows that log each request, sort it by job type and prepare a reply for staff approval. Would a five minute example using sample data be useful?”

Follow up once or twice, then stop. Personal, relevant outreach to 10 well chosen businesses is more useful than an automated blast to 1,000 addresses.

Run a practical discovery call

Ask the owner to walk through one real example from start to finish. Find out who touches it, which systems hold the data, what goes wrong and what a mistake costs.

Then calculate a rough monthly value. If a task takes 15 minutes and happens 100 times per month, that is 25 staff hours, but do not assume all 25 hours become cash savings.

Some saved time will be used for customer work, training or other admin. Present it honestly as capacity unless the business can remove a direct cost or produce measurable extra sales.

End the call with one recommendation. If the process is not suitable, say so, because a manual checklist or a standard software feature may solve it more cheaply.

How do you deliver a safe, reliable project?

Begin with a process map and test data. Build the smallest working path first, then add validation, approval and error handling before connecting live customer information.

Make, Zapier and n8n can connect forms, email, spreadsheets and customer systems. ChatGPT or Claude can classify text, extract fields and draft content, but important outputs should use approved source material and a human check.

Use Airtable or Google Sheets for a simple pilot log when appropriate. For production work, follow the client’s security requirements and avoid placing sensitive data in a tool unless its terms and settings fit that use.

Test more than the happy path

• Try empty fields, duplicate submissions, unusual spelling and very long messages.
• Check what happens when an app is offline or an API limit is reached.
• Confirm that the workflow does not send two replies or create duplicate records.
• Test hostile instructions inside customer text, especially when AI can access tools or data.
• Create an alert and a manual fallback for failed runs.

Keep a simple record of each test, expected result and actual result. Ask the client to approve acceptance tests before launch, then train at least two people on pausing the automation and handling failures.

For ongoing support, review failed runs, tool usage, AI output samples and changes made by connected apps. Set a monthly limit for included fixes and quote larger changes separately.

Give the client a short operating guide with the workflow purpose, connected accounts, owner, approval step, common errors and shutdown method. The goal is a useful system the business can understand, not a hidden setup that depends on you forever.

FAQ