AI Strategy for Small Business: Start With the Job

An AI plan for a small business is a short list you can say out loud. It names the job that changes, the facts it needs, and the person who checks the work. It is not a list of tools, and it is not a big speech about the future. If your page cannot name those three things, you have a pile of apps and a hope that the pile will turn into a plan.

Cited in this note

75%

The assessment scores the same groups of facts we want on that page. The implementation framework is the order of work once the job has a name. The readiness framework is the score sheet under both of them. We start with the job. The tools can wait in the other room until the page exists.

We have watched owners skip the page because it felt too small to matter. In a company this size, a good plan is one job that still works in a busy week. A thick binder that covers every department you do not have is just for show.

Use is ahead of the plan

Microsoft’s 2024 Work Trend Index says 75% of knowledge workers use AI at work.¹ Of those users, 46% started less than six months before the survey. Also, 60% of leaders worry that their company has no clear plan for how to use AI. About 78% of AI users bring their own tools to work. At small and midsize businesses, that goes up to 80%.

Leaders also worry about finding talent, and 55% said so. Yet only 25% of companies were planning to offer training on generative AI. We read that as a truck cab and a front office that have not met yet. People already use the tools. But the plan, the training, and the hiring worries are all behind them.

NIST’s risk framework groups the work into four parts called Govern, Map, Measure, and Manage.² A plan that cannot point to those four parts has not really started. Govern means who is allowed to decide. Map means where the work and the facts live. Measure means how you will know when the answer is wrong. Manage means what you do in the week it is wrong. If your page is only a list of product names, it skips all four.

Salesforce asked 3,350 leaders at firms with 200 workers or fewer.³ The survey ran in August and September 2024. About 75% said they were at least trying out AI, and that rose to 83% among growing small and midsize businesses. Among growing firms, 78% planned to spend more on AI. Only 55% of the firms that were shrinking did. Of the leaders who already used AI, 91% said it helps their revenue. Those are leaders talking about their own companies, so we will not repeat the claim as if your revenue had already agreed.

These numbers come from different groups of people. Microsoft asked knowledge workers and leaders, while Salesforce asked small business leaders. Neither one can take the place of a job you can name on a single page. A worker who brings a tool from home and a leader who says revenue went up can both be telling the truth. They are just answering different questions. Your page still has to pick one job.

A worked example

Picture a heating and cooling company that runs a handful of trucks. This is only an example, and nobody measured it for a report. The owner prices jobs from the driver’s seat. He has started pasting notes about a broken unit into a chatbot on his own phone. He likes the draft it gives him, but he has not told the dispatcher.

The dispatcher builds the day’s board from calls, texts, and a whiteboard that is more right than the software they pay for. Two techs trust the board. They do not trust a long message the owner forwards at night.

So there is no plan yet, just people quietly using their own tools. That is the same habit we saw in the Microsoft numbers, where most AI users bring their own tools and even more do at small and midsize businesses. The owner is not a bad guy. He found something that helps him write a quote when he is tired. The trouble starts when a customer reads the draft back to the dispatcher the next morning. It names a part the shop does not stock and a price the owner would have changed at a desk.

The office manager finds out in the worst way, which is from the customer. She has no page that says which jobs the tool may touch. She has no rule that a quote only leaves after she or the dispatcher has seen it. Nobody trained her, because there was nothing to train on except that the boss has an app. The low training numbers from Microsoft will sound familiar to her. She lives in the gap between using the tool and having a plan.

Here is the page we would write with them, and it still has no tool name on it. The job is to draft estimates for swap-out jobs. It is not for complaints or anything that sounds like a safety problem. The facts come from the price book on the office computer and the notes the tech already puts on the ticket. They do not come from the owner’s memory of a house that looked similar. The dispatcher checks the draft, because she already knows which crews can do the work this week. When she is out, the office manager takes her place, and both names are written down. If a draft names a part they do not stock, it does not go out. That is a real plan, and it is short enough to tape inside a cabinet door.

We also left some things off the page on purpose. We did not write that the company will look for ways to use AI everywhere. We did not put a chatbot on the website. The techs need the board to stay right. The owner’s late night drafts can stay only if they become drafts the dispatcher expects, built from facts the office already trusts. Until then, the honest rule is to stop forwarding the chat to customers.

How we’d run it

We build the page before we add a login, because the order matters more than the paperwork.

First, we ask the owner what he already does in private. The chatbot in the truck is a fact, so we write it down and we do not scold him for it. We call it bring-your-own use, which is common. We keep it apart from a company plan, which is not. A habit that stays hidden is risky. But once we name it, we can change it.

Second, we pick one job and draw a fence around it. Swap-out estimates are inside the fence. Complaints, safety calls, and anything the price book cannot back up are outside it. The fence is the plan. A job with no fence turns into every job by Friday, because the tool will answer any question you type. Being able to answer is not the same as having permission.

Third, we name the facts and the checker in the same meeting. For this company the price book and the ticket notes are in, and the owner’s memory is out. The dispatcher checks, and the office manager covers. We put time on the dispatcher’s calendar so that checking is a real slot and not just a good idea. This is where the Map and Measure parts of the NIST framework live, even if we never say those words in the room. Govern is the owner agreeing that the fence is real. Manage is what they do when a draft breaks the fence. They hold it back and say the reason out loud once, so the techs hear the rule.

Fourth, we train only the people named on that page. That means the dispatcher and the office manager, working on this one job with real tickets from last week. We do not hold an all-hands talk about the future of the trades. Only 25% of companies in the Microsoft survey planned any training on generative AI. We would rather be in that group for one role than outside it with a pep talk. The techs get a simpler lesson. Quotes to customers come from the office and not from a forwarded chat.

Fifth, we leave the revenue claim alone. Salesforce heard leaders say that AI helps their revenue, including a large share of those who already used it. We will not print that on this company’s page as if it were their result. Their result, if they earn one, will be a dispatcher who can clear draft estimates without a surprise part. If they want a score before they start, the assessment and the readiness framework are the sheets to use. If they want the order of work after the page exists, that is the implementation framework.

Common mistakes

The first mistake is writing the plan as a shopping list, which is what the sellers hope for. A list of products is not a job, a place for facts, or a checker. It is a stack of bills. In the Microsoft survey, 60% of leaders already worried that their company lacked a plan for AI. A shopping list will not calm that worry. It will only add more logins.

The second mistake is pretending that bring-your-own use is not happening. About 78% of AI users bring their own tools, and that is 80% at small and midsize businesses. Your owner in the truck is normal. If the plan ignores him, he will keep forwarding drafts, and the plan will describe a company that does not exist. You can fold his habit into the plan, or you can ban it in one sentence he agrees with. But please do one of the two.

The third mistake is training the staff in a vague way, which keeps the training imaginary. The people who need it are the dispatcher and the office manager. A general session lets everyone nod and changes no ticket. The Microsoft numbers show that training plans were rare, and copying that gap is easy. A short lesson on one job is the fix.

The fourth mistake is borrowing another group’s revenue number. Of the Salesforce leaders who already used AI, 91% said it helps their revenue. They run firms with 200 workers or fewer, and the survey ran through August and September 2024. They are free to say that about themselves, but you should not treat it as your forecast. The 75% who were at least trying AI, and the 83% among growing small and midsize firms, still describe leaders and not your whiteboard.

The last mistake is skipping Govern, Map, Measure, and Manage because the names sound like a big institution. The names are just drawers. They ask who decides, where the facts are, how you notice a bad answer, and what you do next. A heating company can fill those drawers on one page. If it will not, the plan has not started, no matter how new the owner’s app feels.

What we’d do

We would tape one page inside the cabinet, and we would not start any tool until the page has names on it. That is the view we bring into the room. Use is already ahead of the plan. In the Microsoft survey, 75% of knowledge workers use AI at work, and 46% of those users had started less than six months before they were asked. The plan is the thing that is hard to find. Chasing more use, when you already have use in the truck, is how you stay among the 60% of leaders who worry they have no clear plan. We would make the plan a fence around one job.

We would not tell this owner that AI helps revenue just because 91% of the leaders who already used it said so to Salesforce. We think leaders answer surveys about their own story. We do not think that story carries over to a dispatcher who has not seen the draft. About 78% of growing firms planned to spend more, against 55% of the firms that were shrinking. Spending without a checker is how a shrinking firm gets a new bill. We would spend on the price book, the dispatcher’s time, and a rule the techs can repeat.

We would also be plain about talent. In the Microsoft survey, 55% of leaders worried about it, and we share a smaller version of that worry. You probably do not need a new kind of worker. You need the dispatcher to be allowed to say no to the owner’s draft. That is a people choice dressed up as a software choice. If you cannot give her that right, do not buy anything. Go back to the readiness framework until she has it. Then the implementation framework can run, and the assessment will have something real to score. A plan that the dispatcher can spoil is a real plan. A plan that only the owner understands is a hobby with a title.