By John Wheeler
I have spent more than 30 years running, advising, and working inside businesses. Much of that work has involved accounting, costs, operations, and business software. I am used to asking practical questions: What does this cost? Who owns the work? What happens when something goes wrong? Does the result help the business?
Now I am applying those same questions to a different kind of company.
Through Wheeler Intelligent Systems, or Wheeler IS, I am exploring how far one person can go with AI helping to research, build, test, and organize the work. The central question is simple: How far can AI take a company when the human owner does not continually supply the next task?
I do not have a finished answer. I have an experiment in progress. But the work is already teaching me lessons that I think matter for small business owners trying to put AI to use.
AI expands what I can attempt
The first lesson is about capability.
I can use AI to explore a business problem, draft a plan, work through software code, and develop something I can review. That lets me attempt work that would have required a much longer learning path on my own.
Business Central development is one example. I understand the business problems around product information and ERP systems. Learning the development language and building every part myself would be a substantial project. AI gives me another way to move from a business need toward a tool.
That does not mean every result is correct or ready for customers. It means I can get far enough to inspect, question, and test a possible solution.
I find that more useful than trying to put a precise number on how many minutes AI saves me. Some of this work is work I would not have attempted at all. The value includes the chance to build something new, alongside any time saved on familiar tasks.
Doing a task and owning the work are different
AI can be useful when I give it a clear assignment. Keeping a business moving requires more.
Someone has to remember what was decided, identify the next step, notice when work has stalled, and bring forward the decisions that need the owner. That is the part of the experiment I am especially interested in.
If I have to return every day and rebuild the plan, I remain the point through which all the work must pass. I might produce more, but the business still depends on my constant attention.
I want AI to help maintain continuity. What is in progress? What can move forward? What is blocked? What needs my judgment? What should be finished before we start something else?
Those are familiar management questions. Assigning them to AI makes it necessary to be much clearer about responsibilities, limits, and what a useful update should contain.
A clear role helps more than a broad request
“Help me run the company” leaves too much open to interpretation.
A useful role needs a purpose, a small set of priorities, and clear authority. It also needs a way to report exceptions. For Wheeler IS, I have defined an operating role for my AI assistant that includes research, internal coordination, building, testing, and bringing forward decision-ready recommendations.
Certain actions still require my approval, including spending money, contacting customers, and publishing or deploying work.
That division gives AI room to do useful internal work while preserving my control over actions that create outside commitments.
For another business, the role might be much narrower. An AI assistant could review overdue invoices and prepare suggested follow-up. It could identify missing product facts. It could organize job estimates for review. The role should fit the business problem and the evidence that the tool can handle it.
Focus still matters when ideas are cheap
AI makes it easy to generate ideas and begin new projects. I have plenty of interests: product data, accounts receivable, business directories, education, software tools, and AI use in business.
Each idea can turn into a plan quickly. Each plan can produce drafts, screens, research, and code. That creates a management problem: starting work can feel like making progress even when the useful result is still far away.
I am trying to keep the active priorities small and give more weight to finishing. A limited number of current priorities makes it easier to see what is moving and what needs attention.
There is a practical test behind this. Can I point to something useful that has reached the next stage? Can I explain what remains before someone can use it? If those answers are unclear, generating another plan may add to the pile without helping the business.
A review can pass while the user still has a problem
One recent experience brought this home.
A mockup had gone through repeated review and a large set of checks. The reports covered layout, text, identifiers, and other details. The review work had value. But when I tried to use the result, I could not get the interactive experience I expected.
I wanted a working demo that responded to my mouse.
That exposed a gap between the result being reviewed and the result I needed. A visual mockup can meet its visual requirements and still leave the next business need unmet.
This is why I want the intended outcome stated in plain language before the work starts. If the goal is a working demo, I should be able to open it, click a control, and see the expected response. Screenshots and review reports support that result; the actual experience has to be checked too.
I also want a clear distinction between code that has been inspected, behavior that has been demonstrated, and a workflow that remains proposed. Each tells me something different about readiness.
Independent review helps when it checks the right thing
I have used separate review in the software work because the builder’s confidence is not enough for me.
A reviewer can look for missing cases, unclear claims, and problems that the builder overlooked. But the review needs to connect to the purpose of the work.
For an invoice assistant, that might mean checking whether it recognizes a disputed invoice and stops for human review. For a product-data tool, it might mean checking whether missing facts remain visibly missing rather than being filled with guesses.
I want evidence I can return to later: what was tested, what happened, what was fixed, and what remains open. That makes review useful for the next decision, rather than simply producing a reassuring verdict.
Human attention belongs in the cost of the work
Subscription costs are easy to see. The time I spend explaining, reviewing, correcting, and restarting work is harder to track.
Yet that time matters. If a tool produces a draft quickly but requires repeated rounds of repair, the total effort can be much greater than the first result suggests.
One of my goals is to understand how much useful work AI can complete for each hour of my attention. I do not yet have a reliable measure of that across the company. It is something I want to track, alongside errors, cycle time, and the usefulness of the outcome.
A good operating update should help me decide. It should explain what changed, what is blocked, what happens next, and whether my input is needed. Long reports that leave me searching for the decision use up the attention I am trying to protect.
Business experience gives me a way to judge the output
My accounting and operations background remains useful in this experiment.
I can ask whether a cost model makes sense, whether a workflow has a clear owner, and whether a proposed tool addresses a real problem. AI helps me explore and build, while experience helps me question what comes back.
This is an encouraging starting point for people who know their industry well. You do not need to begin with a grand plan for an AI company. You can begin with a recurring problem you understand and a clear description of a better result.
Use a fictional example or a suitable test environment. Give AI a bounded assignment. Review the output. Try the tool yourself. Record what needed correction. Then decide whether to expand the task.
What I am trying to learn next
Wheeler IS is still an experiment. I am working toward useful, repeatable products and services, and commercial results will matter. Producing more drafts and code does not, by itself, establish customer value or a profitable business.
My next questions are about continuity, finished outcomes, and demand. Can AI keep authorized work moving with fewer prompts from me? Can we produce tools people can actually use? Can those tools solve problems that customers are willing to pay to address?
I remain enthusiastic because AI has expanded what I can attempt. I also see the need for clear roles, focused priorities, direct testing, and honest reporting.
That is what I want to share as the experiment continues: the useful results, the gaps, and the changes those gaps lead me to make.
If you run a small business, what recurring task would you most like AI to help carry forward? I would welcome a conversation about the problem, the result you need, and what would make you trust the work.