AI Adoption in Small Businesses

Two-thirds of US small businesses now claim they use AI, nearly three times the share in 2023. ^[1] However, far fewer businesses seem to actually be using it in their day-to-day operations. When small businesses were asked whether they had used AI in any business function in the past two weeks, fewer than 25% said yes. ^[2]

In addition, most small businesses have been using AI in a limited (and oftentimes ineffective) capacity. Last year the most common uses of AI related to writing content (83%) and personal productivity (61%), while only 21% used AI to automate a process. About half of businesses were still experimenting with AI, and only 7% had fully integrated it into their company operations. ^[3]

Research suggests the businesses seeing the largest financial gains from AI are far more likely to have redesigned core workflows around it. ^[4] With about half of small businesses still experimenting, many may not have seen those gains yet.

This can be a daunting and unclear evolution for small businesses though because what does "redesigning a workflow around AI" really mean? And how does a small business with limited resources actually work towards this milestone?

In a small business training program, many participants cited the "barriers to deeper integration" related to "lack of technical expertise", difficulty "choosing tools", and "data privacy concerns". Nearly 75% of participants in this program said they would benefit from more training and help putting AI in place. ^[5]

I've observed this same sentiment while working with small businesses over the years. As such, I've decided to write up this article to demonstrate an iterative path a small business can adopt using practical examples.

Four Roles for AI

While every organization's use of AI may differ slightly, for the sake of discussion, we'll classify AI usage into the following four roles:

1. Assistant: Using AI individually to help out with your tasks
2. Builder: Using AI to build tools that systematically improve operations
3. Teammate: Using AI to interpret information within a process for people to act on
4. Operator: Using AI to make decisions without a person needing to intervene

It's important to note that these roles should be assessed at a task-level. For example, some critical tasks may never be suitable for the Operator role due to risk factors or non-determinism, but they might still benefit from one of the other roles.

This breakdown will help us navigate a few different ideas:

• To use AI within the workplace you do not need to blindly trust its output
• You can use AI in a way that mitigates risk and data privacy concerns
• Different tasks within organizations may be suitable for different roles of AI
• There is a natural progression path for maturing AI adoption within an organization

To start, let's discuss the most basic way to use AI: using it as an assistant for individual tasks.

AI as an Assistant

Most small businesses that use AI use it in this capacity. To put it simply: A person asks AI for a draft or an answer and decides what to do with the result.

This can seem very helpful because AI can summarize information and output these drafts or answers very quickly. However, there are some challenges when leveraging AI in this capacity.

• AI may output inaccurate information for a variety of reasons
  • Trained on outdated information
  • Over-simplifying a nuanced topic
  • Hallucinating ^[6]
• The individual using AI becomes a "reviewer" and must verify claims
  • May not have the subject matter expertise or attention to detail to properly assess the AI's output for accuracy
  • May begin to experience "automation complacency" over time, reducing quality ^[7]
  • As corrections are made, they need to be shared organization-wide or other team members may experience the same incorrect result in the future

These challenges played out publicly in 2025, when Deloitte agreed to refund part of its fee for an Australian government report that cited research that didn't exist. Deloitte had disclosed using generative AI to help produce the report. ^[8]

Shared Instructions

This is where the concept of "Shared Instructions" or a centralized knowledge base that AI assistants can reference comes into the picture.

Having a centralized repository of information can help by:

• Enabling specific knowledge-holders to maintain info pertaining to their expertise
• Serving as a source of truth for specific organization information
• Allowing for iterative improvement of the AI assistants because inaccuracies can be corrected within a shared space

Most importantly, these shared resources enable both new and experienced employees to benefit from this shared knowledge.

A related study measured the effect of using AI to spread top performers' knowledge. Among more than 5,000 customer support agents, an AI assistant trained on top performers' conversations raised productivity by 15% on average and by about 30% for less experienced and lower-skilled agents. ^[9]

Whatever tool a business uses, shared instructions need five things:

• A shared home: Everyone works from the same instructions, kept in a team ChatGPT or Claude project or in a common folder.
• Clear ownership: The people who know each subject best maintain the instructions, and everyone else uses them.
• Version history: A record of changes means a bad edit can be traced and corrected.
• Answer rules: The instructions tell the AI to answer only from the business's own documents and to flag anything it is unsure about.
• Business accounts: Employees use company accounts, so business information stays out of personal ones.

Shared instructions also give employees an approved way to use AI. In a global study, 57% of employees said they hide their use of AI at work and present AI-generated work as their own. ^[10] When a business provides the tools and reference material, employees have less reason to hide their AI use.

Refined over time, the instructions become useful beyond the assistant role. A store's written rules for special orders or supplier preferences, for example, can serve as the requirements to build out an automated process. AI can use these refined instructions to build out these automations, which leads us to discussing the builder role.

AI as a Builder

As a builder, AI helps create tools for deterministic workflows that might otherwise take too much time or technical skill for a small organization. Once built, these tools often don't need to call an AI model to run, so a business can benefit from AI without depending on it day to day.

Building tools this way avoids a few concerns I hear most often about putting AI into operations:

• Inconsistency: The same request to an AI model can produce a different answer each time. Catching those differences requires a review step or a very detailed set of rules.
• Cost: AI is often billed per token, and token use can be hard to predict once a process runs at volume. In my experience, unexpected usage costs are one of the most common surprises for businesses running AI in their processes.
• Data Privacy: You can have AI help build automations using test data so that sensitive data is never exposed to the AI models. This allows you to run these automations within your private networks for privacy or compliance purposes.

Because an automation behaves the same way every time, a business can test exactly what will run before the automation goes live. AI can also help write the documentation, so the business depends less on technical resources for upkeep or questions.

Building an automation also brings a process's exceptions to light. Cases that staff used to handle from memory, or that went unhandled, have to be spelled out before the automation can follow them. AI can help find these cases by asking about situations the written rules don't cover yet.

As new scenarios are uncovered, staff can use AI to refine the automated process's rules to better handle these situations. Using AI as a builder in this way creates an iterative path for AI to improve operations without incurring risks of inconsistent answers, unpredictable costs, or data privacy concerns.

However, there are limits to this approach. Although AI-built automations can prove reliable, they only work for deterministic workflows where someone has defined the rules in advance. Many other business tasks may be ineligible for this approach because they require interpretation or are based around unstructured data. For these tasks, you may need to evolve your usage of AI and incorporate it more like a teammate.

AI as a Teammate

As a teammate, AI handles a step that requires interpretation each time a process runs. The AI's output helps the people working on a task, and they can use it, adjust it, or set it aside. The output should generally be internal, such as context added to an internal work item or a drafted reply (needing review) to a customer inquiry.

Whether that output needs approval depends on where it goes. Context added to an internal ticket can simply be corrected if it turns out to be wrong. Likewise, a reply drafted for a customer can be reviewed, approved, or adjusted by a person before it's sent. These serve as quality-control levers organizations can use to mitigate inaccuracies and ensure alignment with AI tools.

Tasks That Fit This Role

In my experience, tasks that work well in this role share the following traits:

• They repeat often, so small time savings add up.
• They involve unstructured input, such as messages, notes, or descriptions, that fixed rules can't interpret.
• Someone already works on the task and can put the AI's output to use.
• Mistakes in the AI's output are easy to spot, correct, or disregard.
• The data isn't sensitive, or the sensitive parts can be removed before the AI sees it.

Doing It Safely

Because the AI step runs within a business process, inconsistent results, unpredictable costs, and data privacy become real concerns. The following practices can be used to limit these concerns:

• Strip sensitive data: Before the AI gets the data to interpret, sensitive information is stripped to ensure data privacy or compliance requirements
• Interpretation and drafts: The AI step interprets, matches, summarizes, or drafts, and people decide what to do with the result.
• Limited access: The AI step can read what it needs and add notes or drafts, but it can't send messages, close tickets, or edit records.
• Approval for customer-facing output: A person approves anything that reaches a customer or changes a record before it takes effect. Internal context, such as ticket notes, can be corrected after the fact.
• One run per item: Running the AI step once per ticket or order keeps token costs in line with the volume of work. Clear limits in the instructions, such as which sources to check and how long the output should be, also keep the AI step's work predictable and its token use in check.
• Tracked corrections: The corrections people make to the AI's output show where the AI step's instructions need refinement.

In the teammate role, people still carry out every action. That keeps risk low, but for high-volume tasks, waiting on people can become a bottleneck. Once a process is stable and the AI step's output no longer needs frequent refinement, the business can let the AI step take the action itself. In that sense, an operator is a more mature teammate. It is the same AI step, given more responsibility once it has proven to be reliable.

AI as an Operator

As an operator, AI makes decisions and acts on them without a person approving each decision. AI in this role is often called an agent. While people may still be involved to review the cases the agent can't resolve within its limits, at least some routine decisions are allowed to be acted on by AI operators.

This is the hardest role to get right. Results can vary, and token costs can be hard to predict. With no one approving each decision, a mistake can reach a customer or another system before anyone notices.

Few companies, even large ones, have given AI this much responsibility. At US companies with 1,000 or more employees, two-thirds of departments use AI agents, but only 17% of departments let an agent complete a task without a person approving each step. ^[12]

Day to day, an agent works from written rules. It handles routine cases on its own and sends the rest to a person. Those escalated cases show where the rules need work.

Getting There Safely

A careful way to reach this role is to start with a task that already works well with AI as a teammate, then let the AI take over actions gradually, starting with the ones that are easiest to reverse.

Luckily, you can use evidence from past results to decide whether or not a task is ready to incorporate an AI operator:

• If people often correct the AI teammate's output, the task isn't ready and you should continue to refine the AI teammate's instructions until you reach a stable recommended path.
• If people rarely correct the AI teammate's recommendations, then that task might be ready for an AI operator to act without a person approving prior.

Once a task runs without approvals, these safeguards keep the agent's decisions within bounds:

• Small, reversible decisions first: Flagging, routing, or holding an order is safer than issuing a refund or canceling the order.
• Hard limits: Define which actions are allowed and set caps, such as a maximum order value, so anything beyond them goes to a person.
• Written rules: The agent follows documented rules that build on the shared instructions from the assistant role.
• Escalation: When a case doesn't fit the rules, the agent hands it to a person.
• Logging: The agent records every action, and someone regularly reviews a sample of routine decisions along with every escalated case.
• An off switch: Staff can pause the agent and return to the manual process.

Customer-Facing Agents

A support chatbot would be considered an operator in this context too, because each answer it gives is a decision about what to tell a customer.

Customer-facing agents carry two risks. The first is legal liability. Courts and tribunals have held businesses responsible for what their chatbots tell customers, even when no one reviews each answer. In 2024, a Canadian tribunal ruled against an airline on this point. ^[13]

Jake Moffatt asked Air Canada's chatbot about bereavement fares after their grandmother died. The chatbot told him he could request the bereavement discount after their flight, but the airline's actual policy said otherwise. When Air Canada refused the discount, the airline argued that the chatbot was responsible for its own actions. The tribunal disagreed:

"It should be obvious to Air Canada that it is responsible for all the information on its website."

Moffatt v. Air Canada, 2024 BCCRT 149

In May of this year, a German higher regional court reached a similar conclusion about a company's website chatbot. ^[14]

The second risk is service quality. Klarna said in 2024 that its AI assistant was doing the work of 700 full-time customer service staff. ^[15] A year later, Klarna's CEO said cost savings had played too large a role in the company's move to AI customer service, and that service quality had suffered as a result. He described investing in human support as the way forward. ^[16]

Both risks shrink when the chatbot answers only from approved content, hands uncertain questions to a person, and makes a person easy to reach.

Managing AI Operators

In my experience, businesses do best with AI as an operator when they treat an agent the way they would treat a new employee. The agent starts with a small set of tasks and limited access, and it takes on more work only after its results show it can handle that work. It's also a good practice to have a "manager" for AI operators that can periodically review the output for feedback and improvement opportunities.

Closing Thoughts

AI has been widely available for a few years now, and the data increasingly links AI's financial gains to businesses that build it into core operations, though few see that payoff yet. ^[4] As with most new technology, small businesses have a harder time adopting it due to time, knowledge, and budget constraints.

But this time it's a bit different. AI can ease some of those limitations by filling in missing technical knowledge and iteratively solving problems at a reasonable cost. One technical person working with AI can now build what once took a team of engineers. This may make AI one of the most empowering technologies small businesses have had the opportunity to engage with since the early internet.

The bottleneck has moved as a result. For most small businesses, the bottleneck is now the time it takes to define their workflows and willingness to evolve their operations. A business that writes down their operational workflows can now automate or improve them quickly with AI's help. In the reordering example from the builder section, ten hours of building with AI could save a store about three and a half hours every week.

In addition, documentation used to be de-prioritized and viewed as a tedious administrative step, whereas now it might just be the first tangible step you can take to improve your business:

1. If you aren't using AI yet, start with it as an Assistant and document your workflows along the way.
2. Use that documentation to automate time-consuming tasks with AI as a Builder.
3. Add AI as a Teammate to interpret and recommend actions that align with your workflows and procedures.
4. As those recommendations prove reliable, let AI act on them as an Operator without a person approving each one, and give it more well-defined actions over time.

It's rare for a tool this powerful to also offer a gradual path to adoption that adds value at each step. Whether you aren't using AI yet or already use it regularly, I hope this article shows how you can gradually incorporate it into your business operations as the technology keeps advancing.