SERVICES · SAMPLE DELIVERABLE
A BURN AUDIT, REDACTED
This is a real Burn Audit with the client removed — the same report you get, minus the name. One month of usage exports, four findings priced at their real volume, and the fixes ranked by what they save. Nothing here needed a meeting.
THE WASTE MAP
Every finding, ranked by monthly cost. Start at the top — that's where the money is.
Context re-sent on every turn
HIGHThe “chat with your docs” feature keeps one long thread per session. Average input is 31,000 tokens per request — the same document context re-billed on every follow-up. This is the history multiplier at feature scale.
Fix Reset context per question and retrieve the relevant passages instead of stuffing the whole doc. Fresh threads are free money.
No prompt caching anywhere
HIGHZero cached reads in the entire export. The triage agent sends the same 2,400-token system prompt on all 62,000 monthly calls, billed at full input price every time.
Fix Enable provider prompt caching on the static prefix (system prompt + tool schema). Cached context bills at ~10% of list.
Flagship model on triage work
MEDIUM71% of spend runs on GPT-5.5 — including ticket routing and labeling, classification-grade work that a model 6–10× cheaper handles with no measurable quality loss.
Fix Route triage to a cheaper tier (GPT-5.4 / a Flash model); keep the flagship for the hard ~15% that actually needs it.
Output-heavy: full rewrites
MEDIUMThe internal copilot returns whole files instead of diffs. Output/input ratio is 1.8× — and output is billed 4–6× higher than input, so this is the expensive end of the bill.
Fix Ask for diffs, not rewrites; cap generation length. Most “rewrite the file” requests only needed three changed lines.
Two things that keep this honest. “Costs today” overlaps: these are slices of the same $6,200 bill, not additive line items — finding 3's $4,400 of GPT-5.5 spend already contains the re-sent context and the un-cached calls above it, so the column deliberately sums to more than the bill. And the fixes don't simply add up: caching and context-reset touch some of the same tokens, so the naive recoverable sum ($3,050/mo) overstates it — the combined, non-double-counted figure is ≈ $2,300/mo. Savings are estimates at public list prices, labeled as such; your negotiated rates and the order you ship the fixes in will move them.
WHAT YOU DO WITH IT
It's yours to keep
A written report (PDF + the raw numbers). No dashboard to log into, no tool to adopt, no retainer. Hand it to your team and start at line 1.
The fixes are ordered by money
You don't have to do all four. Doing just the top two here recovers ~$2,150/mo — most of the total — for a few days of engineering.
Ask anything, in writing
Follow-up questions answered by email. No walkthrough call to schedule; the report stands on its own.
WANT ONE FOR YOUR BILL?
$290, one week, fully async. Or try the free browser version first — it runs the same four checks on your CSV, in your browser.
REQUEST THE AUDIT →