Guide · · 6 min read
The hidden cost of your AI stack
Six subscriptions at $20 each is $1,440 a year. Here is how AI spend creeps up on teams, and the questions that stop it.
By Manish
Nobody decides to spend three hundred dollars a month on AI. It arrives one twenty-dollar subscription at a time, each individually reasonable, none of them ever cancelled.
The problem is not the price of any single tool. It is that AI pricing is structured to be difficult to total, and most teams have genuinely never added it up. When they do, the number is usually two to three times what they guessed.
Where the money actually goes
Five patterns account for almost all of the surprise, and they compound.
- Seat sprawl. You bought five seats for a trial. Two people still use it. You are billed for five, every month, indefinitely.
- Overlapping capability. Three tools in your stack all generate images. You are paying three times for one job because each arrived to solve a slightly different problem.
- Credit systems. Usage priced in credits rather than currency makes real cost genuinely hard to compare, which is not an accident.
- The annual discount. Paying yearly saves twenty percent and removes the monthly moment where you would have noticed and cancelled.
- API spend nobody owns. A developer wires something up during a prototype, it ships, and the bill grows quietly on a card nobody reviews.
The last one is the most dangerous because it scales with usage rather than headcount. A subscription that is wasted costs the same every month. An API integration that is wasted costs more every month.
The costs that never reach the invoice
Direct spend is the part you can see. Three others are usually larger and never get counted.
Switching cost is the first. Every tool you adopt accumulates prompts, templates, saved workflows and habits. Moving off it later costs real time, which is why teams stay on tools they have outgrown. That lock-in is a cost you paid at adoption and pay again at exit.
Context-switching is the second. A person using nine AI tools spends a meaningful part of the day deciding which one to open. Consolidating to five slower tools frequently beats nine fast ones on total output, because the decision itself is expensive.
Verification is the third and largest. Every AI output that matters needs checking, and checking takes longer than people budget. A tool that is ninety percent accurate on work you must be certain about has not saved you ninety percent of the time. It has changed your job from producing to reviewing, which may be worth it and is not free.
Any productivity claim that ignores verification time is measuring how fast you can generate work for someone to check.
Run the audit, quarterly
Export the card statement, filter for anything that looks like software, and put it in a sheet with four columns: monthly cost, who used it in the last thirty days, what it replaced, and what would break if it disappeared tomorrow.
Most teams find several rows where they cannot answer the fourth column. Those are the cancellations. Cancel them and see whether anyone notices, which is a faster and more honest test than a discussion about whether the tool is valuable.
Also check the second column against the seat count you are paying for. Seat sprawl is usually the single largest line, and it is the easiest to fix because fixing it requires nothing except attention.
Three questions before adding anything
These take a minute and prevent most of the creep.
- What does this replace? If the answer is nothing, you are adding to the stack rather than improving it, and that needs a stronger justification than enthusiasm.
- Who owns it? A tool with no named owner has nobody to notice it stopped being used.
- When do we review it? Put a date on it. A tool with no review date is a tool you will still be paying for in two years.
The second question does most of the work. Ownerless subscriptions are the ones that survive, because cancelling something requires somebody to decide, and nobody decides about things that are not theirs.
Consolidation is usually the biggest saving
Before optimising individual prices, look for duplication. Most stacks have two or three tools doing substantially the same job, acquired at different times for different reasons by different people.
The general-purpose assistants now handle a wide range of tasks adequately, and adequately is the operative standard for anything that is not your core work. Keeping a specialist for the thing you do constantly and using a generalist for everything else is usually both cheaper and simpler than a specialist per task.
The counter-argument is real: generalists are worse at each specific job. That matters for the work that defines your output and matters much less for the work that merely supports it. Being honest about which is which is the entire exercise.
What good spending looks like
None of this is an argument for spending nothing. A tool that genuinely saves an hour a week pays for itself many times over, and refusing to pay for that is a false economy with a spreadsheet attached.
The goal is a stack where you can name, for every line, what it does and who uses it. Teams that can do this usually spend less than teams that cannot, and they get more from it, because the spend followed a decision rather than an accumulation.
Check the free tier honestly before every upgrade. The answer is more often yes than the pricing page would like you to believe, and finding out costs a month.
API spend is a different animal
Subscriptions have a ceiling. API usage does not, and that difference deserves its own controls rather than being lumped in with software spend.
A subscription that is wasted costs a fixed amount every month. An API integration that is inefficient costs more as you grow, which means the problem is smallest exactly when you would have caught it and largest once it is embedded.
- Set hard spending limits at the provider, not just alerts. An alert tells you after the money is gone.
- Separate keys per project or environment, so you can attribute cost rather than staring at one total.
- Watch cost per request over time, not just the monthly bill. A rising bill on flat traffic means something changed.
- Cache aggressively. Repeated identical prompts are common and paying twice for the same answer is pure waste.
The single largest saving available to most teams is routing. Sending every request to a frontier model when most of them are classification or extraction is the default and it is expensive. A cheaper model handles the bulk, the expensive one handles what genuinely needs it, and the difference is frequently a majority of the bill.
The build-versus-buy line moved
Teams historically bought AI features because building them was hard. Building has got much easier, which changes the calculation for a specific band of tools.
If a tool is essentially a prompt with a nice interface, and you are paying per seat for a team, building an internal version is now plausible in days rather than months. Several of the products in a typical stack are exactly this.
The counter-argument is real and usually wins for anything non-trivial. Buying gets you maintenance, updates, someone else handling model deprecations, and no internal owner. Build the thing that is genuinely specific to you and buy the rest, which is the same rule as always, applied at a line that has shifted.
Making the number visible
The structural fix is not discipline, because discipline decays. It is making the total visible to the person who feels it.
One shared document listing every AI subscription, its cost, its owner and its review date. Updated when anything is added. It takes ten minutes to create and it changes behaviour immediately, because adding a line to a visible list is a different act from expensing something quietly.
Teams that keep this list spend less than teams that do not, and the difference is not that they are more frugal. It is that the total exists somewhere a person can see it, and an invisible total only ever goes up.
Cut your AI bill without losing anything
Many paid tools have a free plan that is good enough. Check before you renew.