Case study

How we cut LLM costs by about 80% on a consumer finance platform

On a consumer finance and investing platform, we rewrote the AI pipelines and AI costs fell by about 80%. Redundant model calls are gone, each call gets only the context it needs, and spend caps and rate limits are in place. Every AI-generated report logs what it cost. The features are the same as before.

Book a scoping call

Key numbers

~80%less spent on AI

13slides per report, still with KPI extraction and output checks

The problem

The platform, built at Mokka Studios on AWS, uses AI to write financial reports on single stocks and to make images for news articles. The pipelines we inherited called the model more often than needed and sent it far more context than it used. Nothing put a hard ceiling on spend.

What we did

Removed redundant model calls
Cut each call down to the context it actually needs
Added spend caps and rate limits
Capped image generation per day and per category, and generated images only for articles above a score threshold
Cached answers in the report pipeline and logged the AI cost of every report

Architecture and stack

Backend on AWS (API, admin dashboard, market data ingestion, AI pipelines), a background queue for long jobs, Anthropic models with web search, and versioned prompts.

Results

About 80% less spent on AI
The reports kept every feature: a 13-slide deck with KPI extraction, source-domain checks, output verification and Italian translation

What it means for your company

Most LLM bills are high because of how the system is built: duplicate calls, bloated prompts, no caching, no caps. Model prices are rarely the main cause. If you audit where the tokens go, you usually find the savings before you need to change model.

Frequently asked questions

How can I reduce my LLM API costs?

Measure what every call costs first. Then remove redundant calls, send each call only the context it needs, cache repeated answers, and set spend caps and rate limits. On this consumer finance platform those changes cut AI costs by about 80%, with the same features.

Do I need to switch to a cheaper model to save money?

Not necessarily. On this platform the ~80% saving came from rewriting the pipelines: fewer calls, smaller context, caching and caps. The model was never downgraded. Fix the architecture first, then see if a model change is still worth it.

How do you stop AI costs from spiking?

Put hard limits in the code: daily and per-category spend caps, rate limits, and rules for when a feature runs at all. On this platform, for example, images are generated only for articles above a score threshold. Since every report logs its cost, you see any drift straight away.

Does cutting costs hurt output quality?

It shouldn't, as long as the quality checks stay. The financial reports still run source-domain checks and verify the model output. The cuts hit only waste: duplicate calls and context the model did not need.

Next step

Is your AI bill growing faster than your usage? We can find what to cut without losing quality.

Tell me the workflow. We scope it on a free call, then a working version is live within 10 days.

Book a scoping call