AI Cost Optimisation

Analyse the following AI integration for cost optimisation opportunities. Review each area and give specific, actionable recommendations.

**Token Waste**
- Are prompts longer than they need to be? Identify redundant instructions, repeated context, or verbose system prompts.
- Are responses being generated at a length that exceeds what the application actually uses?
- Is context being stuffed with irrelevant history or documents?

**Model Selection**
- Is the task matched to the right model tier? Flag cases where a cheaper model (e.g. Haiku) would produce equivalent results.
- Are simple classification, extraction, or routing tasks using a frontier model unnecessarily?
- Could any steps be replaced with deterministic logic instead of an LLM call?

**Caching Opportunities**
- Are identical or near-identical prompts being sent repeatedly? Identify candidates for semantic caching.
- Could prompt prefixes be cached to reduce input token costs?
- Are there lookup-style queries that could hit a cache before calling the model?

**Batching**
- Are individual items being processed one at a time when they could be batched into a single call?
- Could async batch APIs reduce per-request costs?
- Are there fan-out patterns that could be consolidated?

**Architecture**
- Is the system making unnecessary round-trips (e.g. calling the model to format data that code could format)?
- Could retrieval be tightened to reduce context size?
- Are there fallback chains that could be reordered by cost?

For each recommendation, estimate the potential cost reduction (as a percentage or multiplier) and state the trade-off, if any.

```
[describe your AI integration, paste code, or share usage metrics]
```