How to start with AI in a small business: six decisions and what they cost
Written against the systems we have shipped. Where the record does not support a claim, this article says so instead of making one.
Start with a process, not a product. Pick one task your team repeats every week, put a number on what it costs you in hours or fees, then find out whether an existing tool covers it or the work crosses too many systems to buy. Before anything gets built, settle two more things: what data is allowed to leave your building, and what the system is not allowed to decide on its own. Ship that one process, measure it, and take the next one after. The tool is the last decision, not the first.
Why the tool is the last decision
The usual sequence starts with a product. A demo looked sharp, a competitor mentioned a name at a conference, and the question becomes which one to buy. Then the seats get purchased, three people log in during week one, two log in during week two, and by the end of the month the old spreadsheet is still open on the second monitor. The software was never the problem, so replacing it changed nothing.
The sequence that works runs the other way. You name a piece of work that is costing you something, you describe what a correct result looks like, and the tool falls out of those two facts. That is six decisions, and only the last one involves software.
1. Find the process that repeats
Walk your week and look for the task somebody does over and over. Four tests tell you whether it is the right one.
- It happens on a schedule. Weekly or more. A quarterly task will not earn back the work of changing it.
- It arrives in the same shape. The same form, the same report, the same kind of email, the same file from the same place.
- One person owns the result. If three people each do a version of it, you have three processes, and you need to pick one.
- Correct is definable. Someone can look at the output and say yes or no without a meeting.
An orthodontic practice was rebuilding the same patient slide decks by hand every week. Same inputs, same layout, same person, every week, and a finished deck is either right or it is not. That is what a first process looks like. We turned it into a local pipeline that assembles the presentations and keeps missing clinical inputs visible to the operator.
If nothing in your week passes all four tests, the answer is not a different AI product. It is that the work is not defined yet, and defining it comes first.
2. Put a number on what it costs today
A process you cannot price is a process you cannot justify changing. The number does not need to be exact. It needs to exist.
Count it four ways. Staff time, at what that hour actually costs you. Direct fees, the kind that show up on a statement every month. Errors, at what a single one costs to catch and undo. Delay, at what the wait costs in revenue or goodwill. Most first projects are carried by one of those four, and knowing which one tells you what the system has to be good at.
A dental practice we worked with could name the processing fee on every recurring card payment. That fee is what made moving those payments to ACH worth building toward, and it set the target before a line of code existed. The integration is running in production. We do not publish a savings figure for it, because the practice has not shared one we can stand behind.
3. Decide whether to buy or build
Buy when one product handles the task without you rearranging the rest of the workflow around it. Build when the work crosses systems, depends on rules that exist only in your business, or needs a data boundary and an exception path that no product offers.
Test the buy option first, every time. It costs less and it can be running this month. Our AI Finder asks four questions about the business and the bottleneck and returns existing tools that fit, including the honest catch on each one. If a product covers the process, buy it and go back to step one with the next process.
The build case usually announces itself the same way: the work is not hard inside any single system, it is hard in the gap between two of them, and no vendor owns that gap.
4. Decide what leaves the building
This decision costs nothing to make now and is expensive to make later. Write down which data is allowed to leave your network, and which is not.
In the dental integration, the cloud service carries only the payment information needed to retrieve work. Patient linkage, account context, posting detail, and every exception record stay on the clinic workstation. That boundary was a design constraint before it was a feature, and it is the reason the system could be built at all in a practice handling protected health information.
Small businesses get this wrong in one direction: they assume the boundary is somebody else's problem because the vendor mentioned encryption. Decide it yourself, in one sentence, and hold every option against it.
5. Decide what the system refuses to do
The fast path is only safe when the uncertain path stops. Before you build anything, name the states where the system hands the work back to a person.
- No guess under ambiguity. An uncertain match becomes an exception a human resolves, not a best effort the system commits.
- No duplicate action. A retry must not post the same payment, send the same email, or create the same record twice.
- No silent success. The system reads the result back and records either confirmation or a structured failure. Nothing is assumed done.
- No judgment calls. Anything you would want to defend to a customer, a regulator, or an employee stays with a person.
These four are what let the common path move quickly. A system that never stops is not confident. It is unsupervised.
6. Ship one, measure it, then take the next
Ship the single process you picked, in production, with real work flowing through it. Then measure the number you wrote down in step two, on the same terms you measured it before.
Resist the platform. The instinct after a first win is to design the system that handles everything, and that is where small companies lose a year. Take the next process on the list and run the same six decisions again. Five processes handled one at a time beat one platform that handles none of them yet.
Ten questions that spec your business for AI.
Ten yes or no questions across seven dimensions, about two minutes, ending in a readiness spec sheet. It runs steps one and six of this article against your actual operation. Free, and there is nothing to install.
What each decision costs
Five of the six decisions cost time rather than money, and you can make them without hiring anyone.
| Decision | What it takes | Cost |
|---|---|---|
| 1. Find the process | One pass through a normal week with the four tests in hand | Free. The ten questions run it in two minutes. |
| 2. Price it | Staff hours, fees, error cost, delay cost | Free. An hour with your own numbers. |
| 3. Buy or build | Check whether one product covers it end to end | Free to check. The AI Finder returns fitting tools and their limits. |
| 4. Data boundary | One written sentence on what may leave your network | Free now. Expensive to retrofit. |
| 5. Stop conditions | The list of states that return to a person | Free now. This is the list that keeps you out of trouble later. |
| 6. Build and ship | The first process, in production, measured | Buy Your Time Back is $995 fixed, payable only after we confirm fit. Larger builds are scoped once we have seen the process. |
Prometheus pricing as of September 2026. The $995 entry engagement returns five hours a week or you do not pay. Custom build pricing depends on how many systems the work crosses, so we quote it after looking rather than before.
When a process is not ready
Three signals mean the answer is process work, not AI, and adding software will make the mess faster instead of smaller.
- It changes shape every time. If no two instances look alike, there is no pattern to hand off.
- Nobody owns the outcome. With no owner, there is nobody to define correct and nobody to catch a wrong result.
- Correct is an opinion. If two experienced people disagree about the right output, write the rule down first. That rule is the actual deliverable.
None of this is a reason to wait. Pick the messiest high-volume task, spend a week writing down what it actually does, and you will usually find a clean process hiding inside it.
What to hold on to
- Process before product. The tool is the sixth decision, not the first, and buying it early is how the spreadsheet survives.
- Four tests pick the process. It repeats weekly, it arrives in the same shape, one person owns it, and correct is definable.
- Price it before you change it. Staff time, fees, errors, delay. One of those four will carry the project.
- Try buying first. If one product covers the work without rearranging your workflow, that is the answer and it is available now.
- Two rules go in writing before any build. What data may leave your network, and what the system is never allowed to decide.
- One process at a time. Ship it, measure the number you wrote down, then run the same six decisions on the next one.
Questions owners ask us
How do I start using AI in my small business?
Start with a process, not a product. Pick one task your team repeats every week, put a number on what it costs you in hours or fees, and decide whether an existing tool covers it or the work crosses too many systems to buy. Then settle what data is allowed to leave your building and what the system is not allowed to decide on its own. Ship that one process, measure it, and take the next one after. The tool is the last decision.
What should a small business automate with AI first?
Automate work that repeats every week, arrives in the same shape each time, has one owner, and produces a result someone can check. Copying the same fields between two systems, rebuilding the same report or presentation, and sorting routine requests are the usual first candidates. Leave hiring, clinical judgment, credit decisions, and anything that requires the system to guess about a person with your team.
How much does it cost for a small business to start with AI?
The first three decisions cost time rather than money. Naming the process, pricing what it costs today, and testing whether an existing tool covers it can all be done in-house, and our ten-question diagnostic and AI Finder are free. Our fixed entry engagement, Buy Your Time Back, is $995 and is payable only after we confirm the process is a fit. Custom build work is scoped after we have seen the process, because the price depends on how many systems it crosses.
Should I buy an AI tool or build a custom system?
Buy when one product handles the task without changing the rest of your workflow. Build when the work crosses systems, depends on rules specific to your business, or needs a data boundary and exception handling that no product offers. Test the buy option first. It costs less and it can be running this month.
How do I know if my business is ready for AI?
A process is ready when you can point to it happening on a schedule, name the person responsible for the result, show where the inputs come from, and describe what a correct output looks like. If the work changes shape every time, or nobody owns the outcome, or correct is a matter of opinion, define the process before adding AI to it.
What should a small business not automate?
Do not automate a decision you would want to defend to a customer, a regulator, or an employee. Hiring, firing, clinical calls, credit and account approval, and pricing exceptions belong with a person. The system can gather the inputs and lay them out. It should not be the one that chooses.
Sources
Every claim above comes from work we have shipped or from pricing we publish. These are the records behind it.
- Dental payment integration. Recurring payments moved to lower-cost ACH without replacing Dentrix, with patient matching and ledger posting kept inside the clinic. Deployed and running in production. No savings figure is claimed.
- Patient presentation automation. A local-first system that assembles patient presentations and keeps missing clinical inputs and operator judgment visible.
- AI enabling a semi-private airline. Practical AI enablement for operating teams, with identifying details withheld.
- The evidence standard we hold case studies to, which is also the standard this article is written against.
- Buy Your Time Back, the $995 fixed entry engagement referenced in the cost table.