Supplier quote comparison: build a reviewable purchasing worksheet

I would begin with a small purchasing category and the fields the buyer actually uses. Those might include item identifier, description, quantity, unit, unit price, currency, delivery charge, delivery date, validity period, exclusions, and substitution notes. The exact fields depend on the purchase.

Each field needs a state beyond “filled” or “empty.” A value can be explicitly stated, calculated from stated inputs, unclear, or absent. A calculation should show its inputs. An inferred value should never silently become an explicit supplier commitment.

For example, a quote may list a pack price while the request specifies individual units. The worksheet should retain the pack size and expose the conversion. If pack size is missing, stop the normalization for that row. A model's guess could reverse the apparent price difference.

The extraction result should reference the supplier file and the page or section supporting the proposed value. The buyer should be able to open the source without searching a shared drive by filename.

I would keep the original quote immutable in the workflow's storage model and record later revisions separately. A replacement quote may change delivery terms without changing the visible total. Comparing revision identifiers prevents an older worksheet from being presented as current.

The extraction should also preserve qualifications. “Subject to site survey” belongs beside the price, not in a discarded paragraph. “Equivalent item proposed” needs a visible flag. The buyer can then decide whether a technical reviewer must confirm compatibility before any commercial comparison is meaningful.

Once the relevant inputs are verified, ordinary code should calculate totals, differences, and unit conversions. I would not ask a language model to be the final arithmetic engine for a purchasing decision.

The model can help identify a shipping clause or associate a line description with a requested item. The application should validate numeric formats, currency codes, quantities, and calculation rules. If suppliers quote different currencies, use an explicitly approved conversion source and timestamp, or leave the values unconverted for review.

Anthropic distinguishes predefined workflows from dynamically directed agents in Building effective agents. For this bounded task, I would choose a predictable extraction, validation, calculation, and review sequence. The buyer gains little from an agent inventing new comparison steps on each run.

A good worksheet may end with questions rather than a winner. If delivery is excluded, the next action is to obtain delivery cost. If a proposed substitute has a different specification, the next action belongs with the technical owner. If validity expires before the planned order date, the buyer needs a refreshed quote.

I would generate clarification drafts from the unresolved fields and attach the relevant quote revision. The buyer reviews the recipient and question before sending. The system should avoid suggesting that an order has been awarded or that the supplier's terms have been accepted.

When a reply arrives, preserve it as additional evidence. Do not overwrite the original quote with a model's interpretation of the reply. A reviewer should confirm which fields changed and whether a revised formal quotation is needed under the organization's process.

Supplier PDFs and emails are external content. They may contain malformed files, unexpected links, or text that attempts to instruct the model. I would allow the extraction component to read supported content while denying it the ability to change approval rules or place orders.

The purchasing worksheet should not contain live secrets or unrestricted internal notes. The extraction service only needs the relevant request and supplier material. Internal budget ceilings and negotiation strategy should be included only when the workflow genuinely requires them and the data owner approves that use.

OWASP's prompt injection guidance supports treating external instructions as untrusted. In this design, the stronger practical boundary is that the extraction step has no purchasing authority to exercise.

I would ask the buyer to help create synthetic examples with missing tax treatment, bundled items, delivery exclusions, alternate units, revised quantities, and a handwritten correction. Include a scanned quote that cannot be reliably read. The expected result for that file is an explicit extraction exception.

The test should verify source links, field states, arithmetic, revision handling, and the questions presented to the buyer. It should also verify that rerunning the same quote does not create a second purchasing request or discard an existing reviewer correction.

A reviewer correction needs provenance. Store who changed the field, what it was before, and why it changed. On the next extraction run, show any conflict between the new machine suggestion and the reviewed value rather than silently replacing the buyer's work.

I would scope the initial release to a reviewable comparison packet and clarification drafts. Supplier selection and purchase authorization remain with the existing owners. If the organization later wants approved selections written into its purchasing system, that becomes a separate action with its own permissions and acceptance checks.

Measure the time required to prepare a comparison, the number of fields corrected, missing qualifications caught before ordering, and the rate of unusable extractions. Keep the sample size and purchasing category visible. One successful stationery comparison does not establish performance on complex equipment quotations.

The implementation handoff should include supported quote formats, field definitions, calculation rules, exception ownership, and a replay procedure. The buyer should leave with a worksheet they can inspect and correct, plus a clear explanation of which decisions remain theirs.

My demonstration packet would contain three synthetic quotes for the same requested quantity. One would price individual units, another sealed packs, and the third a substitute item. I would deliberately omit freight from one quote and include an expired validity date in another.

The expected worksheet would normalize only the quantities it can establish, preserve the substitute flag, and leave the unresolved landed cost visible. It should not rank the cheapest headline number as the recommended purchase. The buyer should be able to open the source for each qualification and prepare targeted clarification drafts.

After replacing one quote with a revised version, I would check that the worksheet updates its source reference and invalidates any decision tied to the old terms. This is a more useful acceptance exercise than feeding the system three identically formatted quotes. It demonstrates whether the implementation respects the ambiguity the buyer actually needs to resolve.