Getting Started
One Task, One Folder, One System
By John J. BakerJuly 30, 20263 min read
Half the people working in small businesses now use AI at work. That number stopped being interesting a while ago. The interesting number is what they use it for.
Sixty four percent use it for personal productivity: drafting, summarizing, thinking out loud. Twenty six percent point it at a task that comes around every week. Six percent have it running a workflow with minimal human involvement.
Nearly everyone adopted. The real gap is between the businesses whose use stayed personal and the ones whose use reached the operation, and the distance between them is three specific moves.
64%
Personal productivity
Where most use sitsDrafting an email, summarizing a document, thinking a problem through out loud. One person, one session, and it is gone when the tab closes.
26%
A task that recurs
The same job, the same shape, every week. This is the first tier where the time saved is predictable enough to plan around instead of just noticing.
6%
A workflow that runs
Reaches the operationThe work happens with minimal human involvement and a person checks it rather than does it. The only tier where the business, not the individual, got faster.
The question stopped being whether your business uses AI. It is whether the use ever left somebody's browser tab.
Every tier is genuinely useful, which is why the first one feels like enough. It is also the only tier that disappears entirely when the person using it is out for a week.
Those figures come from the Main Street AI Monitor, a quarterly study Ipsos runs for the U.S. Chamber of Commerce Foundation, which surveyed 1,070 workers at small businesses in May 2026. The same study found about one in ten had received any formal training from their employer, which explains most of the shape of that ladder.
The three moves below take roughly a month, and none of them require buying anything new.
Move one: pick the task you would hand a new hire
Start with the task you would give a competent new hire on their first Monday, not the one that has bothered you longest. The hardest problem in your business is hard because it needs judgment, relationships, or information nobody has written down. That is the worst possible first project and the one most owners reach for.
A good first task recurs, follows rules somebody could write down, and produces something you would recognize as wrong in under a minute.
Score one real task, not the category
Name the specific thing that happened last Tuesday and answer all four. Vague candidates pass every filter ever written.
Does it happen at least every week?
Repetition is what turns a saved hour into a saved week. A task that comes around twice a year will cost more to set up than it ever gives back.
Would you spot a bad result in under a minute?
You are the quality control for the first several months. If checking the output takes as long as producing it, you have moved the work rather than removed it.
Could you write the rules on one page?
If the task runs on rules, they can be handed over. If it runs on judgment nobody has written down, the first job is writing it down, and that is a different project.
Does it stay inside your walls until you approve it?
A first workflow that can reach a customer without a person in between puts your reputation on a system nobody has calibrated yet.
Four yeses and you have your starting point. One no and you change the task, not the tool.
Run a real candidate through those four questions before you run a tool. Most of the disappointment in this category comes from good software pointed at a task that was never going to work, which is the same conclusion I reached from the other direction in what to automate first.
Move two: give it the folder you would give that hire
Everyone has access to the same models, so the model is not where your advantage lives. What the system knows about your business is the only part you own, and almost nobody assembles it deliberately.
You would not hand a new estimator a desk and no files. You would give them your last several winning proposals, the prices you do not go below, and the two or three things you always end up correcting on their early work.
Your best finished work
Roughly ten examples of the output done right. Not a representative sample, and not the average. The ones you would want copied.
The rules that do not move
Pricing floors, terms you never negotiate, who has to sign off above a certain number, the customers who get handled differently and why.
The corrections you always make
The three or four things you fix on someone else's work every single time, with the reason attached. The reason is the part that transfers.
The boundaries
What it must never do on its own, and who it hands to when it hits the edge. This is the piece almost everybody skips, and it is what lets you leave the thing running.
The whole shared drive.
A drive is a record of your inconsistency, and pointed at all of it the system will produce a confident average of your worst work alongside your best. This folder is curated, not archived. Four items and an afternoon is the entire assignment.
That is the whole exercise, and it is an afternoon rather than a project. The instinct to point the thing at your entire shared drive is the one worth resisting, for the reason above. I have written at more length about why context is the part you actually own.
Move three: let it reach one system you already pay for
The step from a tool that answers to a tool that works is access. A model that can read your calendar, your inbox, or your job costing data stops guessing about your business and starts reporting on it.
Reach comes in levels, and the mistake is starting at the top.
Read
It can see. It cannot touch.
The system gets access to the data and no ability to change anything. Every question you currently answer by opening three tabs gets answered in one sentence.
- Sounds like
- What did we bid on this address, and how did it close?
- Exposure
- None worth managing. Start here on day one.
Draft
It prepares. A person sends.
Where to settleThe output is real work, finished and waiting, and a human presses the button. You get nearly the whole time saving and keep the judgment where it belongs.
- Sounds like
- Six follow-ups written from this week's pipeline, sitting in your drafts.
- Exposure
- Contained. This is where most workflows should live for a long time.
Act
It finishes and tells you after.
The task completes without a person in the loop. Appropriate for narrow, well-tested, low-variance work, and earned one task at a time rather than switched on.
- Sounds like
- Logging the job, filing the receipt, updating the record.
- Exposure
- Real. Requires a log you actually read and a way to undo.
Move a task up one level only after it has been right many times in a row at the level below.
Promote the task, never the system. Earning level three on invoice coding tells you nothing about whether it should have level three on anything a customer will read.
Start read only. A system that can see your job costs but cannot touch anything carries no risk and still answers questions you currently open three tabs to answer. A contractor who lets the system read the pipeline and draft the follow ups, then sends them personally, has taken most of the value with almost none of the exposure.
The honest limits
A month of small moves does not produce an autonomous business. It produces one workflow that runs and a team that has watched it happen, which is the only foundation the second one gets built on.
Trust is the real budget here. An output that goes out under your name and is wrong costs more than the hours it saved, and people stop opening the thing quietly rather than telling you. I have written about how a technically correct automation can still lose the room. Read only access and a person on the send button are not training wheels. For plenty of workflows they are the permanent design.
And none of this scales sideways on its own. The second workflow is easier because you know what you are doing, not because the first one taught the system anything about it.
Where ClearOak comes in
Most of our work with owners starts exactly here, and it is less about tools than people expect. Choosing the task honestly, getting the reasoning out of the person who holds it, and building the thing so the team still runs it the week after we leave.
If you have tried this and it worked in flashes and then stalled, that is a normal place to be and it is usually a scoping problem rather than a technology one. Schedule a call and bring the task you were about to start with.
Prefer email? Reach me at john@clearoakconsulting.com and tell me which task in your week comes around most often. That is usually the one worth the month.