AI got cheaper: why you are still paying too much
If you set up an AI automation last year and never touched it again, there is a good chance you are paying three times more than you need for the same result.
That is not an exaggeration. Over the past few months a queue of new models, lighter versions and open alternatives showed up delivering comparable quality at a fraction of the price. People are publishing comparisons showing equivalent results at a third of the cost. The market moved and almost nobody went back to check what was already running.
It is the usual oversight. Someone signs up for internet, the plans improve, the price drops, and they stay on a three year old contract. With AI it happens faster, because prices shift every few months.
Why nobody reviews it
Good automation is invisible. It runs, nobody complains, so nobody looks. The invoice lands in the "software" bucket and disappears among the others.
There is also a legitimate fear: touching what works. Nobody wants to swap the model, break quality and spend two weeks putting out fires. That fear is reasonable. The answer is not standing still, it is testing properly before switching.
Automation nobody reviews is not stable. It is forgotten.
What to look at first
Before changing anything, you need to know where the money goes. In most cases I see, the spend is badly unbalanced: one or two tasks eat almost everything and the rest is pocket change.
List your automations and answer for each:
- How much does it cost per month?
- How many times a day does it run?
- Is the task simple (classify, extract, summarize) or does it need polish (customer facing text, analysis, decisions)?
Put that in a ten line table. Usually the answer jumps out: you are using the most expensive model on the market to triage messages, which any light model handles.
The rule: expensive models only where they show
This is the criterion I use, and it holds up most of the time.
Tasks the customer never sees, with short, patterned output: classifying, tagging, extracting a field, deciding who to route to. Light model. The quality gap between expensive and cheap here is nearly nothing, and the cost drops a lot.
Tasks the customer reads, or that decide money: writing a proposal, answering a delicate complaint, analysing a contract. Good model, no savings. Getting it wrong here costs more than any subscription.
A well built system uses both. The cheap model handles triage and volume, and only calls the expensive one when the case deserves it. It is the same reasoning as having a junior and a senior on the team. You do not put the senior on stamping forms.
Test before switching, always the same way
Swapping models without testing is like changing supplier without sampling the goods. The honest procedure is simple and takes an afternoon:
- Pick 30 to 50 real cases that already went through the system, including the weird ones.
- Run them through both models, the current one and the candidate.
- Compare side by side. Count errors, not impressions.
- If the candidate ties or loses a little and costs much less, switch. If it loses badly, do not, and you spent an afternoon to be sure.
Keep those cases. They become your standard test, and next time a new model appears you repeat it in an hour instead of starting over.
The hidden cost: bloated prompts
There is another place where money leaks, and it is not the model price. It is the size of what you send every time.
Plenty of automations were built sending the company's whole manual with every message, just in case. That multiplies the cost tenfold without improving the answer. In most cases, sending only the relevant piece works the same or better, because the model gets lost less often.
If your bill looks high for the volume you have, that is usually the reason. Worth checking before blaming the provider.
What this is worth in practice
A common case: a company spending around 900 reais a month on triage and response automation. Triage was 80% of the calls and it was all on the most expensive model available.
We moved only the triage to a light model, kept the good model for the answers that go to customers, and trimmed the text sent on each call. The bill landed near 250 reais. Response quality stayed the same, because the part customers read did not change.
That is not an extraordinary case. It is what usually happens when someone finally sits down and looks.
Put it on the calendar
AI prices will keep moving, in both directions. Your defence is a habit, not luck.
Schedule a review every three months. Half an hour: open the invoice, look at the table, check whether a relevant new model showed up, run the standard test if it is worth it. People who do this pay less and spot improvements before their competitors.
If you have automation running and never reviewed the cost, let's take a look together.