The cheapest AI answer may create the most expensive work

The team switches to a cheaper model. The inference bill drops. Someone sends a cheerful note to the sponsor.

Three weeks later, analysts are reopening more cases than before. The summaries miss conditions. Some join the wrong records. Each answer costs less, and each accepted result now costs more. It’s an illustrative story, but the pattern behind it is worth settling before anyone negotiates unit prices.

Part of the trouble is that AI costs are scattered. Model usage shows up in a technology bill. Document preparation sits in a project budget. Human review hides inside a business team’s workload. Optimise one line and you can simply push cost into another.

Pick a unit of work with a finish line

Decide what “done” means before you calculate anything. For a supplier review, it might be an approved comparison. For customer support, a resolved case that meets your quality bar. For an internal analyst, a briefing the decision owner accepts.

The FinOps Foundation’s guidance on unit economics distinguishes efficiency measures, such as cost per token, from business measures, such as cost per resolved case. You need both. The first helps engineers tune consumption. The second tells a sponsor whether the workflow deserves money.

Keep the maths honest. Abandoned outputs don’t count as results, but their cost still counts. Treat reopened cases consistently. If the definition of “done” changes halfway through an experiment, say so before you show anyone a trend.

Count the work that happens outside the model

A complete cost view covers source preparation, retrieval, inference, evaluation, operational support and business review. Separate the variable costs from the shared and fixed ones. A small pilot can look expensive because it’s building something reusable. A big rollout can look cheap because that same foundation was quietly left out.

Measure review time from real samples, not an assumed percentage. Some cases need a quick check of the source. Others need a specialist to redo the analysis from scratch. Record the spread, so a difficult minority doesn’t vanish inside an average.

Watch the retries, too. People rephrase questions, switch models or give up and ask an analyst. If you only log the final successful answer, you never see what it cost to get there. Where your workflow and privacy rules allow, connect the attempts to the task they belonged to.

Compare options at the same quality bar

Run the cheap and the expensive configuration on the same representative workload, including the ordinary cases, the exceptions and the ones that should be refused. Agree the minimum acceptable quality first. Then look at price. A model that produces fewer acceptable results shouldn’t win just because each request is cheaper.

Routing can be smart. Straightforward extraction rarely needs your most capable model; a tricky comparison across conflicting documents might. But the routing itself needs testing. Send hard cases down the easy path and the saving disappears.

The FinOps Framework gives finance, engineering and business owners a shared language for technology value. Applying it to AI workflows like this is our own practice, not something the framework prescribes.

Look for what actually moves the number

In an executive review, find the biggest controllable cost drivers and the quality limits around each one. Better source data might cut retries. Clearer scope might cut unnecessary retrieval. Showing the evidence more clearly might shorten review. A lower model price is one lever among several, and often not the biggest.

Where a better data foundation helps several workflows, separate the shared investment from each workflow’s extra cost, and reconcile the totals with finance before anyone claims a saving.