# Sample $299 AI Bill Roast

This is the lightweight artifact for the 24-hour paid roast. Numbers are illustrative and do not represent a real customer.

## Snapshot

Buyer: Example coding-agent team  
Monthly AI spend reviewed: $8,200  
Primary stack: Claude Code, Codex, Cursor, OpenAI API, Anthropic API  
Verdict: worth a deeper audit if monthly spend stays above $8K.

## One Visible Leak

The team is paying premium reasoning prices for work that does not need premium latency or deep reasoning:

- repository indexing and repeated file summaries
- test-output interpretation after deterministic failures
- issue triage and duplicate-ticket grouping
- nightly eval reruns
- repeated tool output carried across long conversations

Estimated visible waste: $1,850-$2,400 per month.

## Fast Savings Model

| Waste source | Current pattern | First fix | Estimated monthly savings |
| --- | --- | --- | ---: |
| Repeated repo context | Same architecture notes reloaded into long sessions | Cache stable repo summaries and pin short contracts | $620 |
| Eval reruns | Full paid eval sweep on every prompt tweak | Run changed-path evals during work, full sweep nightly | $480 |
| Tool-output replay | Large command outputs copied into context repeatedly | Summarize command output before re-entry | $420 |
| Model tier overuse | High-reasoning default on classification and cleanup tasks | Route low-risk tasks to cheaper model/tier | $350 |
| Failed-loop retries | Agent retries without budget caps | Add retry ceilings and stop reasons | $270 |

Modeled monthly savings: $2,140.

## What I Would Change First

1. Put a hard budget cap on coding-agent tasks by workflow: exploration, implementation, test repair, review.
2. Store short repo summaries and architecture contracts outside the conversation instead of rediscovering them.
3. Split evals into cheap changed-path checks and a scheduled full suite.
4. Route formatting, issue triage, log summarization, and deterministic test diagnosis away from premium reasoning.
5. Track cost per accepted pull request, not just tokens or messages.

## Do Not Send

Do not send source code, API keys, credentials, raw prompts, raw outputs, customer data, account IDs, or unrelated personal data.

## Inputs Needed For A Real Roast

- monthly spend by provider or coding tool
- rough usage by workflow or team
- which tasks use high-reasoning modes
- retry/failure examples with sensitive data removed
- eval or test-loop frequency
- current budget caps and alerting, if any

## Paid Upgrade Trigger

Upgrade to the 72-hour audit only if one of these is true:

- monthly AI spend is above $8K
- AI cost is above 15% of gross margin for a product line
- coding-agent spend is rising faster than engineering output
- the team needs a board- or finance-readable savings plan