Most businesses that believe they have an AI problem have a subscription problem.
The pattern is consistent. Someone trials a tool. Someone else buys a different one for an adjacent job. A third arrives bundled inside software you already pay for. Two years later the card is charged by nine vendors, three of them do overlapping work, two are used by one person each, and nobody can say what any of it replaced.
This is a straightforward thing to fix, and fixing it usually funds whatever you actually want to build.
Why it happens
Not carelessness. Three structural reasons.
Per-seat pricing hides the total. Forty dollars a month reads as nothing. Nine of them across a dozen people is $50,000 a year, and no single approval ever looked like a $50,000 decision.
Trials convert silently. The free tier becomes a paid tier at a usage threshold nobody is monitoring.
Bundling obscures overlap. Your accounting platform, CRM and helpdesk have each added AI features. You are frequently paying separately for something you already own.
The audit
An afternoon's work. Do it in this order.
1. Get the actual list. Not from memory — from the card statement and the accounting system, twelve months back. Include anything billed annually, which is where the forgotten ones hide. Most owners find between two and four subscriptions they did not know were active.
2. For each one, answer four questions.
| Question | Why it matters |
|---|---|
| Who uses this, by name? | "The team" means nobody |
| How often — daily, weekly, monthly, never? | Monthly-or-less rarely justifies a seat |
| What did it replace, and did that cost actually go away? | This is the one that exposes the real position |
| What breaks if we cancel tomorrow? | Usually less than expected |
3. Sort into four piles.
- Load-bearing. Used daily, something breaks without it. Keep, and check you are on the right plan.
- Overlapping. Two tools doing one job. Pick one.
- Single-user. One person, one workflow. Sometimes justified. Usually it is a preference that became a line item.
- Dormant. Nobody has logged in for ninety days. Cancel today.
4. Check the seat counts on what remains. Licenses bought for departed staff are the single most common finding, and the easiest money in the exercise.
The question that matters most
The third one above: what did it replace, and did that cost actually go away?
This is where most AI spending fails to justify itself. A tool that speeds up a task without removing the cost of doing it has produced a nicer experience, not a saving. That is sometimes worth paying for — but you should know which one you bought.
Roughly 89% of US small businesses report using AI; under 9% use it in the production of goods or services. The gap between those figures is largely this: tools adopted, nothing structurally changed.
The test: for each tool, name the line in your P&L that is lower because you have it. If you cannot, you have bought convenience. Fine, if that is the intent and the price is right.
What to do with what you find
Do not simply pocket the saving.
The usual outcome is a five-figure annual recovery. The highest-return use of it is to spend a fraction on doing one thing properly: pick the single most repetitive, well-defined, expensive process in the business and automate it end to end, with a written procedure and a verification step.
That is the difference between the 89% and the 9%. Nine tools lightly used change nothing. One process genuinely removed from a person's day changes the cost structure — and that is what a buyer credits.
Keeping it from recurring
Three habits, none onerous.
One owner per subscription. A named person, not a department. Renewal is their decision and they answer for it.
A ninety-day review on anything new. Trials should have an end date in the calendar with a decision attached: adopt properly, or cancel. Most sprawl comes from tools that were never decided about.
Buy against a written procedure, not a demo. If you know what the process is and what a correct outcome looks like, you can test a tool against a specification. Without that, you are evaluating a sales presentation, and every vendor demo works.
That last habit is the one that compounds. Writing the procedure first tells you whether the task is automatable at all, makes the tool choice much easier, and produces a document that is useful even if you never buy anything — a point we cover in Skills: How to Tell an AI Agent How Your Business Works.
The diligence angle
Worth knowing if a sale is anywhere in view.
Buyers will ask for your software contracts. What they are looking for is not the total — it is transferability and dependency. Subscriptions in a founder's personal name, annual contracts that auto-renew past closing, tools holding customer data under agreements nobody has reviewed, and licenses that do not assign on a change of control all surface as diligence findings.
A tidy, documented software estate with named owners and current contracts is a small thing that reads well. A list nobody can explain reads as a business running on habit.
Related
How to Evaluate an AI Vendor covers buying decisions. What Does It Cost to Build an AI Agent? covers build economics. What Buyers Ask About Your Technology covers the diligence position in full.
This guide is for informational purposes only and does not constitute investment, financial, legal or technology advice.