84 PDFs Before Monday’s Procurement Review

The Monday procurement packet: 84 PDFs

Imagine a procurement team preparing for Monday’s supplier review. The packet contains 84 PDFs: contracts, pricing schedules, security questionnaires, and amendments. The job is to identify renewal dates, compare liability terms, and produce a short decision memo with traceable evidence.

Uploading everything into every conversation makes the work expensive to repeat and difficult to audit. A cheaper workflow separates document preparation, structured extraction, and final judgment. Claude Desktop can remain the workspace for discussion, while a relay endpoint handles selected API tasks through a separate client or an explicitly configured tool integration.

Claude Desktop and the relay endpoint have separate roles

Start with a compatibility check. Do not assume Claude Desktop offers a setting that replaces its native chat backend with an arbitrary API base URL. A desktop subscription and API usage through a relay are separate services with separate billing.

For a practical setup, use Claude Desktop to develop the review rubric and discuss findings. Run bulk extraction in a small script using an OpenAI-compatible SDK pointed at https://api.relayvano.com, with your relay API key and an exact model identifier from the provider’s catalog. Confirm the supported request format before processing the packet.

If you want Desktop to initiate those jobs, an MCP server can expose your own extraction tool, subject to your Desktop version’s integration support. That tool calls the relay; it does not replace Desktop’s underlying model. Keep the API key in the tool’s environment rather than in document text or prompts.

Page references make cheaper extraction useful

Build a manifest before calling a model. Give every file a stable document ID, record its version, and extract text with page boundaries intact. Apply OCR to scanned pages, then inspect samples containing tables, footnotes, and signatures. Missing columns can turn a low-cost extraction into an expensive correction cycle.

Ask Claude Desktop to help define a fixed extraction schema. For this procurement review, each result should contain:

• Document ID, supplier name, and effective date.
• Renewal date, notice period, and termination conditions.
• Liability cap, exceptions, and relevant amendment references.
• A supporting quotation and page number for each finding.
• An explicit missing or ambiguous status when evidence is insufficient.

Require the extraction job to preserve the source quotation. A polished summary without evidence forces the reviewer to reopen the entire contract, which erases much of the time saved.

Selective escalation controls model spending

Use a lower-cost model such as Claude Haiku 4.5 for straightforward field extraction, after checking current pricing. Reserve Claude Sonnet 5.5 for conflicting amendments, complicated exceptions, and synthesis across suppliers. Check both models against the same representative documents before assigning the bulk work.

Route cases using observable failures: a required field is missing, the quotation does not support the answer, or two documents disagree. A model’s confidence label alone is a weak escalation rule. For high-impact clauses, have a reviewer verify the source even when the extraction looks complete.

RelayVano is a useful option for this workflow because its pay-as-you-go relay lets you run selected API jobs without committing the whole document process to one model. Sign up for RelayVano and test a small, representative packet before processing the full collection.

Repeated context is the first cost to remove

Save each validated extraction under a key containing the document hash, schema version, prompt version, and model identifier. When one supplier sends an updated schedule, rerun that file and any dependent comparison. Reprocessing all 84 PDFs adds spending without adding evidence.

Keep extraction instructions stable, request compact structured results, and set an output limit large enough for the required quotations. Handle truncation explicitly. Provider prompt caching may offer additional savings where supported, but your own saved results provide a more predictable way to avoid duplicate work.

The Monday memo needs an acceptance check

Bring the validated results and selected source passages back into Claude Desktop for the decision memo. Compare the workflow against a baseline on several representative contracts: field accuracy, quotation support, reviewer correction time, and total API cost, including retries.

Expand only when those checks meet the team’s standard. The saving comes from sending fewer repeated pages and escalating fewer routine tasks while preserving evidence for every consequential finding.