"Your package is in transit" costs the same tokens as a full dispute investigation. That's the problem.

The situation

How the spend got here

The reply workflow was built in a sprint: one prompt template, one premium model, one retrieval call that fetches everything about the shipment — full tracking event history, the carrier contract excerpt, prior tickets from the same customer, and the dispute-precedent library. It produced good replies, so nobody touched it. Volume tripled; the template didn't change.

Now a "where is my order?" query — answerable from the latest tracking event and the ETA — consumes 30,000+ input tokens on the most expensive model available. So does an actual damage dispute, which genuinely needs that context. The invoice can't tell the difference, and neither can the team, because nothing logs cost per query type.

The obvious fix — "use a cheaper model" — is exactly what the ops lead is afraid of: dispute replies carry liability language that cheaper drafting has botched before. The company needs to know which replies need the expensive path, not a blanket downgrade.

The work

What Taz does with this workflow

Taz works from exported usage logs, the prompt templates, and a sample of real replies across every query type. No access to live customer accounts is needed.

• Token-spend map: Every reply type — status, ETA change, delay apology, delivery exception, damage dispute, billing dispute — gets a measured profile: volume share, average input/output tokens, model used, and cost per reply. The map shows exactly where the waste concentrates.
• Context audit: For each reply type, Taz traces what the retrieval call fetches versus what the reply actually uses. Status replies using 2% of fetched context get a minimal context template; disputes that genuinely draw on precedent keep the full pipeline.
• Routing policy: A written decision table: which query types route to a small model with a tight template, which need the premium model, and which trigger human review regardless of model. Includes the escalation rule for when a "simple" status query mentions damage or a refund.
• Prompt rewrites: The one-size-fits-all template becomes per-type templates with the boilerplate stripped, liability language locked as fixed text rather than regenerated each time, and context slots that only load what the type needs.
• Governance pack: Approval gates for high-cost or externally-risky actions, a monthly spend-review checklist, and a lightweight tracker spec so cost per reply type stays visible after Taz leaves.

The deliverable

What the company receives

• Cost leak map — one spreadsheet: reply types by volume, current cost, projected cost under the new routing, and the assumptions behind each projection.
• Model-routing decision table — the per-type policy with escalation triggers, written so an engineer can implement it and an ops lead can audit it.
• Rewritten prompt templates — per reply type, with fixed liability language, minimal context slots, and inline notes on what was removed and why.
• Context-diet spec — what the retrieval layer should fetch per query type, replacing the fetch-everything call.
• Governance and tracking pack — approval gates, the monthly review checklist, and a tracker template that makes next quarter's invoice explainable line by line.