Every prompt costs the same whether it formats a table or writes a strategy section. Nobody knows which is which.

The situation

Why the AI bill doubles and the output stays the same

The firm has three practice areas — strategy, operations, and technology — and each has adopted AI tools independently. Strategy uses a premium model for proposal writing. Operations uses the same model for engagement tracking and time-entry summaries. Finance uses it for month-end close reconciliation and variance narratives. Nobody decided on one model. Nobody evaluated whether the tasks require one. The premium model that writes a nuanced client-facing strategy section also writes "Please find attached the November time summary" — at the same per-token cost.

Context windows are the bigger problem. When a consultant starts a new proposal, the system loads the last six proposals from the same practice area as "reference context." The idea was to maintain tone and structure consistency. In practice, the model rarely references these prior proposals — it generates new content based on the brief. But the tokens are billed. A 4,000-token proposal generation call carries 80,000 tokens of context that the model processes, charges for, and ignores. Multiply by 15 proposals per month across three practices, and the firm is paying for 3.6 million tokens of context that produces no output.

Month-end close is a different kind of waste. The reconciliation prompts pull the full general ledger extract — every line item for the month — into the context window. The model's job is to flag variances above a threshold and write a one-paragraph narrative for each. But the GL extract includes hundreds of line items that are within tolerance. The model processes all of them to find the five that need narratives. A pre-filter that extracts only the variance rows would cut the context by 90% and produce the same output.

The work

What Taz actually maps and delivers

Taz audits every AI workflow the firm runs, maps the cost of each, identifies routing and compression opportunities, and delivers a governance pack the firm can implement without changing its tools.

• Workflow-level cost map: Every AI-powered task is cataloged: what triggers it, which model it uses, how much context it loads, how many tokens it consumes per run, and what it actually produces. The map shows cost per task, per practice, and per month — and highlights the tasks where cost is disproportionate to complexity. A table-formatting task that costs $0.12 per run sits next to a strategy-drafting task that costs $0.45, and the firm can see that one of them should not be using the same model.
• Routing analysis: Each task is classified by the model capability it actually requires: formatting and extraction (cheap model), summarization and synthesis (mid-tier model), judgment and original composition (premium model). The analysis shows which tasks can move to a cheaper model with no quality loss, which tasks need the premium model, and which are borderline — with a recommended test to confirm.
• Context compression audit: For each workflow that loads reference context, Taz measures how much of the context the model actually uses in its output. Prior proposals that are loaded but never referenced get flagged. GL extracts that include within-tolerance rows get flagged. Each finding comes with a specific compression recommendation: remove the context, pre-filter it, or summarize it before injection.
• Prompt efficiency review: Individual prompts are reviewed for redundancy, over-specification, and instruction bloat. A proposal prompt that spends 400 tokens on formatting instructions ("use bullet points, keep paragraphs under 5 sentences, use professional tone") when the model already defaults to that format gets trimmed. Each trim is measured in tokens saved per run and dollars saved per month.
• Governance framework: A set of rules the firm can apply going forward: which task types get which model tier, when context loading is justified, how to evaluate whether a new AI use case belongs on the cheap or expensive tier. Written as a one-page decision tree, not a policy manual.

The deliverable

What the firm receives

• Cost map spreadsheet — Every AI task in one sheet: task name, triggering workflow, model used, context tokens loaded, output tokens generated, cost per run, runs per month, monthly cost. Sortable by cost, by practice area, and by cost-to-complexity ratio. The firm can see in one view where the money goes.
• Routing recommendation table — Each task mapped to its recommended model tier with a rationale. Tasks marked "move to cheap model" include a quality-test description so the firm can verify the output does not degrade. Tasks marked "keep on premium" include the specific capability that requires it.
• Context diet plan — Each workflow with bloated context gets a specific prescription: which context to remove, which to pre-filter, which to summarize. Each prescription includes the token savings, the dollar savings, and the expected output impact (none, minimal, or needs-testing).
• Prompt compression diffs — Before-and-after versions of the top 10 highest-cost prompts. Each diff shows tokens removed, the reasoning, and a side-by-side output comparison confirming equivalent quality.
• Governance one-pager — A decision tree for evaluating new AI use cases: Does it need original composition? Route to premium. Does it need summarization? Route to mid-tier. Does it need formatting or extraction? Route to cheap. Does it load reference context? Justify it or remove it.