The mismatch that drains the budget
Every request you send to a large model costs money. Most of those requests don’t need a large model. That mismatch doesn’t show up on any line of the invoice, but it builds up month after month until someone asks why AI spending is growing faster than usage.
54% of requests to AI models use capabilities beyond what the task requires, according to Accenture’s survey of 750 executives in 17 countries reported by Turing Magazine. The same analysis estimates that only 10% of tasks truly require a frontier model. Classifying, extracting, summarising and formatting can be handled with smaller, cheaper models.
Companies spend on average 25% of their AI budget on tokens, and one in three has already exceeded its annual budget, according to the same source. Executives expect a 78% increase in consumption over the next two years. Lowering the price per token fixes nothing: when the cost per token falls by 25% or more, 95% of the executives surveyed say they will increase total spending instead of saving the difference.
Which model to use for each task
Choose the minimum model that solves the task with the quality the business requires. This table is for assigning by default and moving up a step only when there is a reason.
| Task | Minimum viable model | Signal that you’ve fallen short |
|---|---|---|
| Classification and labelling | Light | Inconsistent labels in edge cases |
| Field extraction | Light | Empty or misaligned fields |
| Summarising long documents | Medium | Loses figures or conditions |
| Writing with brand criteria | Medium | Constant human rewriting |
| Reasoning over several sources | Frontier | Conclusions that don’t hold up |
| Customer service with exceptions | Frontier with escalation | Escalates badly or promises what it can’t deliver |
Move up a step when the third column repeats in a measurable way, not on a feeling.
The metric that changes the conversation
The price per token is the number on the pricing page. The cost per task is the one that appears on the invoice after a month of real use, according to Nexforce. The formula for the next budget meeting:
cost per resolved task = (tokens + retries + human review) / correctly resolved tasks
A model that is cheap per token but needs three calls to get it right ends up more expensive than one that gets it right the first time. The figure a finance director asks for is how much it costs to resolve 100% of tickets, not the million tokens.
Sequence for Monday
- List your use cases and write the real task of each one in a sentence.
- Assign each task the minimum viable model from the table.
- Measure the cost per resolved task for two weeks, with retries included.
- Move down a step where quality holds and document where it doesn’t.
- Leave a routing layer that decides by task, not by habit.
The numbers back up the exercise. A system that estimates output quality before invoking the model and chooses the most suitable one for each query reduces costs by between 40% and 90% and improves quality by between 4% and 7%, according to this work on LLM usage costs. Another query assignment framework achieves cost reductions of between 2.40% and 49.18% with the same accuracy as the best model, according to OptLLM.
Where this approach fails
Routing by task isn’t free. It adds middleware, caching, reservation across providers and observability. The model cost is only one part of the inference stack, and the rest goes on that plumbing and on redoing integrations when a model is discontinued or changes price, as Nexforce explains.
The API accounts for roughly 15% to 25% of the real cost of a generative AI system. Infrastructure takes 25% to 35%, engineering time 30% to 40%, and silent costs (retries, loops, token overconsumption) 10% to 15%, according to Datalaria. A Stravoris brief puts the hidden costs that most budgets don’t consider at 40% to 60%.
There are two more traps. The first is choosing by benchmark: of 445 LLM benchmarks reviewed, 47.8% use disputed definitions, only 16% use statistical tests and 27% rely on convenience sampling, according to the same brief. The second is that a more accurate model doesn’t always give a better return: the higher accuracy of expensive models may justify the investment with more revenue, but not necessarily a higher RoI, according to this decision model.
And although executives consider routing by model the most important measure, only 30% have implemented it, according to Turing Magazine. Less than 20% of token spending can be quantitatively linked to financial results.
If your invoice grows without explanation
Choosing a model is an architecture decision that is reviewed every quarter. At Data Innovation we have more than 20 years with CRM and data from Barcelona, and we build and operate Human+AI intelligent systems with that co-evolution logic: measure first, choose the model afterwards. If your model spending is growing and you can’t attribute it to a specific use case, we can help you set up the measurement and routing. You can also see how we approach a corporate AI strategy.