Most small business owners have now tried an AI chatbot. Far fewer have used AI to change how their business actually runs. The gap between those two things is where a lot of time and money gets wasted, usually on a tool bought after a good demo that never quite fits the way the team works.
Most small business owners have now tried an AI chatbot. Far fewer have used AI to change how their business actually runs. The gap between those two things is where a lot of time and money gets wasted, usually on a tool bought after a good demo that never quite fits the way the team works.
Federal survey data backs that up. In a May 2026 analysis of its Business Trends and Outlook Survey, the US Census Bureau reported that the share of businesses using AI had hovered between 17% and 20% since December 2025. Use climbs with firm size: 37% of firms with at least 250 employees reported using AI, while fewer than 20% of firms with four or fewer employees did.
Smaller firms are not behind because AI is out of reach. More often, they lack a clear way to decide where to start. Here is a simple way to choose a first automation project, test it without disrupting the business and decide whether to build it in-house or bring in specialists.
Why the first project matters more than the tool
The first automation sets expectations inside the company. If it saves visible time, people start suggesting the next one. If it produces errors that someone has to clean up, the team concludes that AI does not work here, and that belief can last for years. The choice of project matters far more than the choice between two similar tools.
A good first project is boring on purpose. It is a task people already understand, that happens often, that follows clear rules and whose results are easy to check. Invoice processing, sorting incoming emails, updating a CRM after calls and preparing weekly reports are typical examples. Strategic decisions, pricing and anything customer-facing with legal consequences are poor places to start.
A simple scoring method for candidate processes
List five to ten tasks the team complains about, then score each one from one to five on four questions. How often does it happen? How strictly does it follow a rule? How easy is it to check the result? How much damage does a mistake cause? Give the last question a reverse score, so that low-risk tasks score higher.
Add a fifth question that is often forgotten: how sensitive is the data involved? A task that handles customer names and email addresses is manageable. A task that handles identity documents, health details or bank account numbers needs more controls, more testing and more care with vendors. That does not rule it out, but it probably should not be the first project.
Add up the scores and look at the top two or three. The winner tends to be an unglamorous task where a working automation would save several hours a week and a failure would be caught quickly and cost little. In many small businesses, that turns out to be the inbox: sorting incoming requests and drafting routine replies.
Check the data before you build anything
AI automation runs on the records the business already keeps, and those records are often messier than anyone admits. Customer names spelled three ways, products with no consistent codes and spreadsheets with columns that changed meaning halfway through the year all cause problems that look like AI failures but are really data problems.
Spend a week looking at the data your chosen process uses. Where does it come from, who updates it and how often is it wrong? Fixing the worst issues before automation starts is cheaper than debugging an automated process later. It also tells you which systems the automation will need to connect to, which is often the hardest part of the job.
This is also the moment to decide what the automation does not need to see. If a process only needs an order number and a delivery date, there is no reason to feed it full customer profiles. Sending AI tools less personal data lowers your risk, simplifies conversations with vendors and makes the whole system easier to explain to customers if they ask.
Borrowing from the NIST framework, at small business scale
The National Institute of Standards and Technology published its AI Risk Management Framework on January 26, 2023, for voluntary use, and added a profile for generative AI in July 2024. It is written with large organizations in mind, but its four core functions translate well to a small business: Govern, Map, Measure and Manage.
In practice, that means four plain steps. Decide who owns the automation and who can switch it off. Map what the process does, what data it touches and what could go wrong. Measure the results against a baseline, including error rates. Manage the problems you find, whether by fixing the process, adding a human check or stopping the automation. None of that requires a compliance department.
Run a pilot with a baseline and an exit
Before switching anything on, record how long the task takes today, how many errors occur and how long customers wait. Then run the automation in parallel for two to four weeks, with a person reviewing its output before anything reaches a customer or a ledger. Track how often the reviewer accepts, edits or rejects what the system produces.
Set the exit criteria in advance. If the acceptance rate stays low after reasonable fixes, stop and pick the next project on your list. Dropping a pilot that misses its targets is the point of setting targets, and it costs much less than keeping a tool that everyone quietly works around.
When an outside partner makes sense
Simple automations, such as routing form submissions or sending reminders, are within reach of most teams using tools they already have. Projects that connect several systems, read documents or draft messages for customers are harder, mainly because the integration work and the edge cases take time to get right.
For those, an outside partner can shorten the path. Aivy, an Australian AI automation company, works with growing businesses in sectors such as accounting, real estate and e-commerce. A useful test for any partner is to ask them to score your candidate processes with you before they quote, then compare their pick with yours. If the two lists barely overlap, find out why before you sign.
The businesses that get real value from AI rarely start with the most ambitious idea. They start with a narrow, well-measured project, prove it works and use that success to fund and justify the next one. A deliberate first step will take you further than a rushed one, even if it looks less impressive in the first month.
Federal survey data backs that up. In a May 2026 analysis of its Business Trends and Outlook Survey, the US Census Bureau reported that the share of businesses using AI had hovered between 17% and 20% since December 2025. Use climbs with firm size: 37% of firms with at least 250 employees reported using AI, while fewer than 20% of firms with four or fewer employees did.
Smaller firms are not behind because AI is out of reach. More often, they lack a clear way to decide where to start. Here is a simple way to choose a first automation project, test it without disrupting the business and decide whether to build it in-house or bring in specialists.
Why the first project matters more than the tool
The first automation sets expectations inside the company. If it saves visible time, people start suggesting the next one. If it produces errors that someone has to clean up, the team concludes that AI does not work here, and that belief can last for years. The choice of project matters far more than the choice between two similar tools.
A good first project is boring on purpose. It is a task people already understand, that happens often, that follows clear rules and whose results are easy to check. Invoice processing, sorting incoming emails, updating a CRM after calls and preparing weekly reports are typical examples. Strategic decisions, pricing and anything customer-facing with legal consequences are poor places to start.
A simple scoring method for candidate processes
List five to ten tasks the team complains about, then score each one from one to five on four questions. How often does it happen? How strictly does it follow a rule? How easy is it to check the result? How much damage does a mistake cause? Give the last question a reverse score, so that low-risk tasks score higher.
Add a fifth question that is often forgotten: how sensitive is the data involved? A task that handles customer names and email addresses is manageable. A task that handles identity documents, health details or bank account numbers needs more controls, more testing and more care with vendors. That does not rule it out, but it probably should not be the first project.
Add up the scores and look at the top two or three. The winner tends to be an unglamorous task where a working automation would save several hours a week and a failure would be caught quickly and cost little. In many small businesses, that turns out to be the inbox: sorting incoming requests and drafting routine replies.
Check the data before you build anything
AI automation runs on the records the business already keeps, and those records are often messier than anyone admits. Customer names spelled three ways, products with no consistent codes and spreadsheets with columns that changed meaning halfway through the year all cause problems that look like AI failures but are really data problems.
Spend a week looking at the data your chosen process uses. Where does it come from, who updates it and how often is it wrong? Fixing the worst issues before automation starts is cheaper than debugging an automated process later. It also tells you which systems the automation will need to connect to, which is often the hardest part of the job.
This is also the moment to decide what the automation does not need to see. If a process only needs an order number and a delivery date, there is no reason to feed it full customer profiles. Sending AI tools less personal data lowers your risk, simplifies conversations with vendors and makes the whole system easier to explain to customers if they ask.
Borrowing from the NIST framework, at small business scale
The National Institute of Standards and Technology published its AI Risk Management Framework on January 26, 2023, for voluntary use, and added a profile for generative AI in July 2024. It is written with large organizations in mind, but its four core functions translate well to a small business: Govern, Map, Measure and Manage.
In practice, that means four plain steps. Decide who owns the automation and who can switch it off. Map what the process does, what data it touches and what could go wrong. Measure the results against a baseline, including error rates. Manage the problems you find, whether by fixing the process, adding a human check or stopping the automation. None of that requires a compliance department.
Run a pilot with a baseline and an exit
Before switching anything on, record how long the task takes today, how many errors occur and how long customers wait. Then run the automation in parallel for two to four weeks, with a person reviewing its output before anything reaches a customer or a ledger. Track how often the reviewer accepts, edits or rejects what the system produces.
Set the exit criteria in advance. If the acceptance rate stays low after reasonable fixes, stop and pick the next project on your list. Dropping a pilot that misses its targets is the point of setting targets, and it costs much less than keeping a tool that everyone quietly works around.
When an outside partner makes sense
Simple automations, such as routing form submissions or sending reminders, are within reach of most teams using tools they already have. Projects that connect several systems, read documents or draft messages for customers are harder, mainly because the integration work and the edge cases take time to get right.
For those, an outside partner can shorten the path. Aivy, an Australian AI automation company, works with growing businesses in sectors such as accounting, real estate and e-commerce. A useful test for any partner is to ask them to score your candidate processes with you before they quote, then compare their pick with yours. If the two lists barely overlap, find out why before you sign.
The businesses that get real value from AI rarely start with the most ambitious idea. They start with a narrow, well-measured project, prove it works and use that success to fund and justify the next one. A deliberate first step will take you further than a rushed one, even if it looks less impressive in the first month.
