AI Prices Keep Falling. Here's How to Find the Use Cases That Now Pay Off
Model providers are undercutting each other on price. That changes which AI projects are worth building — if you know how to re-test them.
The Spreadsheet You Shelved Last Year
A year ago, someone in your company ran the numbers on an AI-powered idea — maybe auto-drafting client replies, summarizing service tickets, or extracting data from invoices. The idea worked in a demo. Then someone multiplied the cost per API call by the expected monthly volume, and the number looked bad. The project got parked. That spreadsheet is probably still sitting in a shared drive somewhere, untouched since.
Here's the problem with leaving it there: the assumption baked into that spreadsheet — the cost per unit of AI work — is not stable. Providers of large language models are competing hard on price, and that competition shows up as repeated, sometimes steep, price cuts on the same capability you tested before. A use case that failed the cost test in 2024 can clear it easily in 2025, without a single line of your workflow changing. The only thing that moved is the price per token.
Most small and mid-sized businesses don't revisit these decisions. They tested once, got a no, and moved on. That's a reasonable instinct when testing is expensive and slow — but testing an AI use case today is neither. The real risk now isn't wasting money on a bad pilot. It's leaving a viable, cost-effective use case shelved because nobody checked the math again.
Why the Math Changes Faster Than the Workflow
The price of running a task through an AI model depends on how many tokens — roughly, chunks of text — it takes to read the input and produce the output. That price is set by providers competing against each other, and that competition is aggressive precisely because switching between models is now technically easy for anyone building the pipeline. A cheaper provider today doesn't need your process to change; it just needs your software layer to point at a different model.
This matters for a very practical reason: the tasks that get cheaper first are usually the routine, high-volume ones — classification, summarization, extraction, first-draft generation. Those are exactly the tasks that showed the worst cost-per-task ratio in early pilots, because volume is what kills a marginal-cost project. As the marginal cost drops, the same volume that made a project unaffordable last year can make it profitable this year.
The trap is assuming the test you ran once is still valid. It isn't. Cost per task isn't a fixed property of the task — it's a moving number set by a market that's still shifting.
How to Re-Test a Shelved Use Case Without Re-Running the Whole Project
- Pull the original pilot's numbers: tasks per month, tokens per task (input and output), and the cost per task you calculated at the time.
- Re-price the same task at current rates from two or three providers, using the same volume assumptions. Don't redesign the workflow — just re-run the cost line.
- Set a break-even threshold before you look at the new price: what cost per task would make this worth deploying, given the time or errors it saves? Decide that number first, so the result doesn't get rationalized after the fact.
- If the new price clears the threshold, run a small live test — real inputs, real staff reviewing the output — before committing to full deployment. Price dropping doesn't guarantee quality holds.
- Log the decision either way, with the date and the price you tested at. That log is what lets you re-test again in six months without starting from zero.
What Actually Tells You It's Working
Once a shelved use case goes live, don't judge it by whether the AI output looks good. Judge it by the number that killed it the first time: cost per task, measured against the time or error rate it replaces. Track it monthly for the first quarter, because provider pricing can move again, and a use case that clears the bar today should keep clearing it — not drift back into the red because nobody's watching the API bill.
The other thing worth watching is scope creep. A cheap task tends to attract more tasks bolted onto it — more fields extracted, longer summaries, extra formatting. Each addition adds tokens, and tokens add cost. The economics that made the use case viable apply to the task you tested, not to whatever it grows into six months later.
Revisit Your Shelved AI Projects Before You Build New Ones
ArkonLabs re-prices and re-tests AI use cases that didn't clear the cost bar the first time, and builds the measurement into the deployment so the numbers stay visible after launch. If you have a shelved project worth a second look, get in touch through www.arkon-labs.com.