A cabinet shop owner outside Hickory showed me his credit card statement in August. Four AI subscriptions, two per-seat licenses, and one API bill nobody in the building could explain. Nineteen hundred dollars a month. He could not tell me what a single dollar of it bought.
That is the most expensive kind of AI spending there is. Not the big number. The unexamined one.
Here is the part nobody selling you AI wants to say out loud: most small businesses are already spending real money on AI. Intuit's 2026 AI Impact Report found 77% of U.S. small and midsize businesses now use AI regularly, up from 48% eighteen months earlier. Seventy-eight percent say it improved productivity. So the question is not whether to spend. It is whether you know what your spending buys.
This is the no-pitch version. Real bands, real second-year costs, and a one-page worksheet you can run before you buy anything.
The four bands, honestly described
Every AI project a small business buys falls into one of four bands. Knowing which band you are in tells you what questions to ask.
Band 1: Subscription and configuration. Hundreds per month.
You buy an off-the-shelf tool and someone sets it up for your business. A missed-call text-back system. An AI receptionist. A review request sender. The tool exists, the configuration is yours.
What it buys: one workflow, handled. What it does not buy: anything custom. If your process is unusual, the tool will fight you. Total first-year cost usually lands between two and six thousand dollars, mostly subscription.
Band 2: Guided setup. Low thousands, one time, plus subscription.
Someone who has done this before designs the workflow with you, connects your actual systems, and hands you the keys with training. This is where most 10 to 50 person companies should start, because the failure mode of Band 1 is buying the wrong tool, and the failure mode of Band 3 is buying too much project.
Band 3: Custom workflow builds. Tens of thousands.
Your process does not fit any subscription, so software gets built around it. Quote automation tied to your ERP. A triage system that reads your inbox and routes like your best dispatcher. The build is the smaller half of the cost. We will get to the bigger half in a minute.
Band 4: Embedded systems. Tens of thousands and up, ongoing.
AI is built into how the business runs, monitored, improved, and owned as infrastructure. This is a program, not a project. Most small businesses do not need this on day one. The ones who end up here usually arrived through Bands 2 and 3 first.
Our own Blueprint engagements start at [CONFIRM]. I am not going to pretend this article has no perspective. But the bands above are market reality whether you hire us or anyone else.
The second-year cost nobody quotes
Here is where the real money hides. Every vendor quotes year one. Almost nobody quotes year two.
Maintenance hours. Software changes. Your phone provider updates their API. The AI model you built on gets replaced with a new version that answers slightly differently. Someone has to notice, diagnose, and fix these things. Budget 15 to 20 percent of the build cost per year in attention, whether that is your time, your IT person's time, or a retainer with whoever built it.
Per-execution creep. Every AI action has a unit cost, and unit costs multiply. Take texting. Twilio's published base rate is $0.0083 per SMS segment in the U.S., before carrier fees. That sounds like nothing. At 2,000 texts a month, it is under twenty dollars. But texts are the cheapest AI action there is. AI phone calls run dollars per call, not fractions of a cent. Document processing is priced per page. When your automation works, volume grows, and the unit costs grow with it. The businesses that get surprised are the ones that priced the subscription and never priced the usage.
Silent failures. This is the expensive one. An automation breaks on a Tuesday. Nobody notices until the following Monday, because the whole point of automation is that nobody watches it. A week of missed follow-ups, unanswered quote requests, or unsent reminders. The cost is not the fix, which takes an hour. The cost is the week.
If you want the technical version of this section: every automation has a unit cost equation. Monthly executions times cost per execution, times twelve, plus a maintenance buffer. The cost per execution itself breaks down into model tokens, API calls, and carrier or platform fees. The single biggest win in AI cost control is routing: cheap models handle classification, extraction, and summarization, while the expensive frontier models only touch the work that actually needs judgment. Run everything through the most expensive model by default and you can easily pay ten times what the work requires. This is also, not coincidentally, what the Govern function of the NIST AI Risk Management Framework is for. Governance sounds like enterprise paperwork. In practice, for a small business, it means one person can answer the question "what did our AI spend last month and what did it do." If nobody in your building can answer that, you do not have an AI strategy. You have a pile of subscriptions.
That is the deep end. You do not need it to use the worksheet. But now you know what the worksheet is protecting you from.
The one-page worksheet
Before you buy anything, fill this out. It takes twenty minutes and it will save you from every bad AI purchase I have ever seen.
1. Name the workflow. One sentence. "Every missed call gets a text within one minute." Not "improve customer engagement." If you cannot name it in one sentence, you are not ready to automate it.
2. Count the monthly executions. How many missed calls? How many quotes? How many invoices? Get the real number from your phone log, your inbox, or your CRM. Guessing here corrupts everything downstream.
3. List every per-unit cost. The subscription, divided by month. The per-text, per-call, or per-page fees at your expected volume. Add 20 percent for the fees you have not thought of yet.
4. Add setup, spread over 24 months. Take the one-time cost and divide by 24. AI projects should be evaluated on a two-year horizon because year one is always the honeymoon.
5. Add maintenance. Your hours times your loaded hourly rate, or the retainer. Be honest about whose calendar this lands on.
6. Total it. Then divide by hours saved. Multiply the hours the workflow currently eats by the loaded cost of whoever does it. If the automation costs less than the hours, buy it. If it costs more, do not. If it is close, the tiebreaker is speed: does doing it in minutes instead of days win you business you would otherwise lose?
That last question is where most worksheets turn positive. A follow-up system that costs $400 a month and saves ten hours looks marginal on labor alone. But if responding in five minutes instead of five hours wins you two extra jobs a month, the math is not close at all.
This connects directly to what we wrote in AI Spend Per Employee: the problem was never the total, it was the invisibility. And if a vendor cannot walk through this worksheet with you before quoting, read Your AI Vendor Is Lying to You before you sign anything.
What to do Monday morning
Pick your most painful workflow. The one that makes you sigh when someone mentions it. Run the worksheet on it with real numbers, not guesses. If the math works, you now know which band you are shopping in and exactly what to ask vendors. If the math does not work, you just saved yourself thousands of dollars and you know precisely why.
Bring the worksheet to a free 30-minute diagnostic and we will pressure-test it with you. No pitch. If the numbers say do not buy, we will tell you that too.
— Greyson