· 8 min read · Field Notes

The true cost of AI for a small business (it's not the AI)

The model is the cheapest line item on the invoice. The real money goes to teaching a system how your business actually works — and there's a way to buy that without funding a $90,000 surprise.

A regional gym company came to us wanting an "AI membership agent."

They run several locations with somewhere between 10,000 and 15,000 active members. The vision was broad but reasonable: respond to new leads immediately, answer questions about memberships and classes, recommend the right package, follow up with people who toured but never joined, and eventually flag members at risk of cancelling.

On paper, it looked contained. They already had a CRM, gym-management software, a phone system, email and SMS tools, and years of customer data. The assumption was that an AI layer could sit across those systems and start doing useful work. A vendor had indicated a first version was deliverable for $30,000 to $40,000, so the leadership team believed they were funding a fairly contained AI implementation.

Then the work got concrete.

The AI was not the expensive part. Model usage at their volume would have run a few hundred dollars a month. The expensive part was making the system reliable enough to put in front of real members and prospects.

Their customer data was spread across four or five systems. The same person could exist as a lead in the CRM, a former member in the gym-management platform, and a call record in the phone system — often with a different email address or phone-number format in each, so matching records wasn't automatic. Membership plans were worse: different locations had run different promotions at different times. Some discounts lived in the billing system, some in spreadsheets, and some existed only as institutional knowledge in the heads of general managers.

Then came the discovery that stung. The company's written sales scripts did not reflect how its best membership advisers actually sold. The strongest advisers were making nuanced calls based on a prospect's goals, prior membership history, family situation, preferred location, and reason for hesitating. None of it was documented anywhere.

So the "AI agent" quietly accumulated workstreams:

  • CRM and gym-management integration
  • Record matching and duplicate cleanup
  • Standardizing membership plans and promotions across locations
  • Documenting sales and escalation rules
  • Consent and communication-preference handling
  • Security and access controls
  • Reporting, monitoring, and exception handling for anything the AI couldn't resolve safely

Notice that none of these are AI. Every useful AI capability had exposed an underlying operational problem. "Recommend the right package" revealed there was no single source of truth for current offers. "Follow up with old leads" revealed inconsistent consent records. "Predict cancellations" revealed that attendance, billing, and customer-service data had never met each other.

The revised estimate moved to roughly $90,000, with another $4,000 to $7,000 a month in ongoing support and software. Deployed across the full membership journey, it could have crossed $120,000. By my estimate, less than 15% of that budget had anything to do with the AI model. The rest was integration, data preparation, workflow redesign, and the controls required to make the system dependable.

The project was paused before the money was spent. What happened next is the important part I want to discuss, and I'll come back to it — because the same confusion that put this company in that position is sitting in front of thousands of small business owners right now.

Why is ChatGPT $20 a month and your AI project $90,000?

This is the question that makes AI budgets feel like a scam. The answer is that they're two different products.

The $20 gets you intelligence as a commodity. The labs have already eaten the grunt of training the model — you're renting the output. But a general-purpose model knows nothing about your business. It doesn't know that one of your locations still honors a family rate from 2019, that a "lead" in your CRM might be a lapsed member in your billing system, or that your best salesperson never opens with a discount.

The $90,000 is the cost of teaching it. Not teaching it to be smart — teaching it your plans, your data, your judgment, your edge cases, and what to do when it's unsure. That work has a name in the industry. A RAND study of why AI projects fail quotes one practitioner:

"80 percent of AI is the dirty work of data engineering. You need good people doing the dirty work — otherwise their mistakes poison the algorithms."

Those who follow me on X will recognize the phrase. The dirty work is always where the real cost — and the real value — hides.

And the data work is stubbornly large no matter how good the models get. Older surveys claimed data preparation ate 60–80% of a data scientist's time; more recent and much larger surveys put it near 40%. But even at 40%, Anaconda's survey found it's still the single biggest activity — more than model training, model selection, and deployment combined. The AI is the small slice. It was the small slice in 2016, and it's the small slice now.

Nobody's budget survives contact

If it were just one gym company, this would be an anecdote. It isn't.

A Benchmarkit/Mavvrik survey of 372 companies found that only 15% could forecast their AI costs within 10% of what they actually spent. Nearly one in four missed by 50% or more. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept — and now predicts that over 40% of agentic AI projects will be canceled by the end of 2027. In both predictions, "escalating costs" is a named cause. RAND, citing industry estimates, puts overall AI project failure above 80% — roughly twice the failure rate of ordinary IT projects.

But the most useful finding is buried in the RAND study. Of the five root causes of AI project failure they identified from 65 practitioner interviews, the number one cause was not data, not infrastructure, not the technology at all. It was leaders failing to communicate what problem needs to be solved and what metric defines success. Read that again.

The most common reason AI projects fail is that nobody decided what they were buying.

That was true of the gym company. "AI membership agent" is not a problem statement. It's a category of software — and an open-ended one, which is why the budget behaved the way open-ended software budgets always behave.

Buy a loop, not a platform

Here's what the salvage looked like.

We reframed the project from build an AI membership agent to improve the first 24 hours after a qualified lead comes in. One workflow. One clock. Numbers attached.

Instead of connecting every system and cleaning every historical record, the first version touched only new leads from the website and selected ad campaigns. It pulled the minimum information required, categorized each lead by stated goals and location, and drafted the next response using a privately maintained version of the company's sales playbook. A membership adviser reviewed every recommendation before anything went out.

The most valuable part of the build wasn't the AI. It was sitting with two of their strongest membership advisers and capturing how they actually decide: when to recommend a trial, when to call instead of text, when not to lead with a discount, when a prospect should go straight to a manager. That judgment had been walking around the building for years, undocumented. Now it's an asset the company owns.

The narrow version cost $12,000 to $15,000 — not $90,000. It did not automate the membership journey. It gave the company a working system that could be measured: response time, percentage of leads contacted within 15 minutes, staff hours spent researching leads, appointments booked, and eventual conversions.

That's the whole model, and it generalizes. Call it buying a loop: one repetitive, expensive workflow, with the proprietary judgment behind it captured, built inside the tools your team already uses, expanded only after the first loop produces measurable value. The alternative — the platform, the agent, the company-wide transformation — is how a $35,000 idea becomes a $120,000 infrastructure program wearing an AI costume.

So what should a small business actually budget?

Three tiers, from my own client work:

Off-the-shelf assistants: $20–30 per person per month. Real leverage for individual tasks — drafting, research, analysis. Knows nothing about your systems, carries none of your judgment. Start here regardless; it's the cheapest education you'll ever buy.

A first loop: roughly $10,000–$20,000, fixed. One workflow, minimal integration, your best people's judgment captured, a metric that moves within a quarter. This is the level where AI stops being a subscription and starts being an operating advantage.

A platform: $90,000–$120,000+, plus thousands a month ongoing. Sometimes justified — eventually. But it should be earned loop by loop, financed by measurable wins, never bought on faith from a pitch deck.

The trap is skipping the middle tier. The $20 subscription feels too small to matter and the platform pitch feels like "doing AI properly," so owners sign up for open-ended development — and open-ended is exactly the property that breaks budgets.

What to do this week

Before you take another AI vendor meeting, run this test on the proposal — and on yourself:

  1. Which single workflow is this for, and is it repetitive, expensive, and measurable?
  2. Whose judgment makes that workflow good today, and where is it written down?
  3. Can the first version run inside the tools we already use, touching the minimum data required?
  4. What number moves within 90 days if this works — and what's it worth if it does?

If the pitch can't answer those four questions, you're not buying an outcome. You're funding exploration — at your expense, on an open meter.

(This is the entire premise behind how we work at Product Layer: a $3,000 audit to find your first loop, a flat-fee $12,000 sprint to ship it. But run the test no matter who you hire.)

The model is the cheapest part of your AI project. The mess is the expensive part. The judgment your best people carry is the part worth owning.

Buy one loop. Measure it. Then buy the next one.