AI can compare supplier quotations by mapping different descriptions to a reviewed item list, while preserving what each supplier actually offered. The useful output is not simply a ranked total. It is a comparison that shows confirmed equivalents, quantity conversions, exclusions and unresolved questions. People must decide whether uncertain products or services meet the requirement before those prices enter a like-for-like comparison.
1. Define the required items before reading prices
Start with a buyer-approved requirement list, not a list copied from one quotation. Otherwise, the comparison may quietly treat that supplier’s specification as the standard.
Give each required item a stable identifier and record its purpose, required quantity, comparison unit and essential attributes. For goods, these might include dimensions, material, model compatibility and pack contents. For services, record the deliverable, duration, location and included work. Separate mandatory requirements from preferences so that a different brand is not automatically rejected when brands are interchangeable.
Ask the relevant technical owner to confirm which substitutions are acceptable. Procurement can organise the comparison, but it should not infer engineering equivalence from similar wording.
The proposed matching rules should also distinguish an equivalent offer from an alternative. An alternative may deserve consideration, but it belongs in a separate view until the buyer accepts the change. Freeze the requirement version used for the comparison. If requirements change, identify the affected rows and review them again rather than silently updating the totals.
2. Extract quotation lines without losing their context
Capture source text and document references alongside the extracted fields. A clean spreadsheet is useful only if a reviewer can trace each value back to the quotation.
For each line, retain the supplier, quotation reference, revision, page, original description, quantity, unit, unit price and line total. Capture document-level terms separately, including currency, stated VAT treatment, delivery charges, validity, lead time and exclusions. A footnote can change the meaning of several lines.
Source: Microsoft’s Document Intelligence overview describes extraction of text, tables and document structure. This supports the extraction stage, not a claim that the service natively determines whether supplier offers are equivalent.
Use document processing as the starting point for preparing records from PDFs, scans or spreadsheets. Check unreadable pages and table alignment before matching. Preserve blank fields as unknowns. A missing delivery charge is not evidence of free delivery, and a missing quantity is not permission to infer one from the buyer’s request.
3. Ask AI for candidate matches with reasons
Use AI to propose mappings, not to settle equivalence. Each proposed match should identify the required item, supporting attributes, conflicting attributes and missing evidence.
Define a consistent record format. Useful fields include required_item_id, source_line_id, proposed_status, matching_attributes, conflicts, unknowns and review_question. Proposed statuses could be confirmed-equivalent, review-needed, alternative, unmatched and duplicate-suspected. Only a human reviewer should set confirmed-equivalent in this proposed process.
Source: OpenAI’s structured outputs guide explains schema-constrained responses and handling for refusals or incomplete output. A schema can make records consistent; it does not establish that their contents are correct. Validate the evidence and route failed responses to review instead of treating them as empty quotations.
In the matching brief, explicitly prohibit guessing specifications, tax treatment or pack contents. Ask for no match when evidence is insufficient. Do not use an AI confidence score as the sole acceptance rule: a plausible label can still conceal a material difference in size, grade or included work.
4. Resolve quantities, bundles and repeated lines
Convert quantities only when the conversion has supporting evidence. Keep the quoted quantity and unit separate from the calculated comparison quantity and unit.
For a hypothetical carton containing 20 units, three cartons become 60 units. Record the conversion factor, its source and the calculation. If the carton size is absent, stop the conversion and ask the supplier. Do not assume that all suppliers use the same packaging.
Handle bundles as explicit relationships. One supplier may quote a kit while another lists its components. Record which required items the kit covers and what remains unknown. Do not distribute a bundle price across components unless the supplier supplies a breakdown or the buyer approves a clearly labelled analytical allocation. Such an allocation is not a supplier price.
Repeated descriptions also need investigation. They may represent separate delivery phases, different sites, optional extras or an extraction duplicate. Compare source locations and line identifiers before removing anything. Keep the original row and the reviewer’s disposition so that a later check can explain why it was included or excluded.
5. Align commercial assumptions before comparing totals
Compare the same scope on the same commercial basis, while keeping unknown costs outside the confirmed total. A lower quotation total may reflect fewer items or a different delivery arrangement.
Show quoted totals separately from any calculated comparison totals. Label currency and whether the quoted amounts include or exclude VAT. Finance should determine the appropriate VAT comparison basis and treatment of unclear tax information. Do not infer recoverability or apply an unstated tax assumption.
Record delivery destination, freight, unloading, installation and commissioning where relevant. A lead time should retain its trigger: after order, after deposit or after approval of drawings. A statement of availability is not necessarily a delivery commitment to the buyer’s site.
Where scope is incomplete, show a confirmed subtotal plus unresolved costs, not an apparently complete adjusted total. If the buyer proposes an allowance for planning, label it as a hypothetical allowance and keep it separate from the supplier’s offer. Payment terms, liability clauses and other contractual consequences require the appropriate finance or legal reviewer’s judgement.
6. Route exceptions to the right person
Assign each unresolved question to someone who can answer it, with a clear next action. The exception queue should distinguish extraction errors from commercial and technical uncertainty.
An extraction reviewer checks whether a quantity was read correctly. A technical owner assesses compatibility or an alternative specification. Procurement requests supplier clarification. Finance reviews the price basis, while legal reviews consequential contract wording where needed. Avoid sending every exception to a single general reviewer.
A proposed application could retrieve approved catalogue records or prepare clarification drafts through tools. Source: OpenAI’s function calling guide describes model requests for functions that application code executes. That mechanism does not itself authorise supplier contact, purchases or payments.
Use read-only retrieval and human-approved communications as proposed boundaries. Treat quotation text as evidence, never as operating instructions. Have the security owner decide document access, retention and provider suitability before processing confidential files. The AI agents versus automation comparison can help frame whether flexible matching needs an agent or a fixed extraction-and-review workflow.
Reusable like-for-like comparison procedure
Use this proposed procedure for each procurement comparison. Copy the fields and gates into a spreadsheet or workflow; adapt them with the responsible reviewers.
- Set the requirement: record item ID, description, mandatory attributes, required quantity, comparison unit and requirement version.
- Register evidence: record supplier, quote reference, revision, page or sheet, source line ID and original wording.
- Capture the offer: record quoted quantity, unit, pack contents, currency, unit price, line total and stated VAT basis. Mark missing values unknown.
- Propose the mapping: record candidate item ID, supporting attributes, conflicts, missing evidence and whether the offer is an alternative.
- Review exceptions: assign an owner, supplier question, next action and resolution evidence. Preserve suspected duplicates until checked.
- Confirm conversion: record conversion factor, evidence, comparison quantity and any bundle relationships. Leave unsupported conversions unresolved.
- Align scope: record inclusions, exclusions, freight, destination, lead-time trigger, validity and payment terms.
- Release for decision: retain reviewer, decision date, confirmed subtotal, unresolved costs and separate alternatives. Do not rank incomplete offers as fully comparable.
Completion gate: every required item has either a reviewed mapping or an explicit gap; every calculated value has evidence; unresolved material differences remain visible to the purchasing decision-maker.
Worked example: a match, an unknown pack and a repeated line
A hypothetical buyer requires 60 stainless-steel M8 × 30 mm bolts of a specified grade, delivered to a Johannesburg site. All descriptions, quantities and prices below are fictional.
Supplier A lists “SS hex bolt M8x30, specified grade”, 60 each at R4 each. Supplier B lists “stainless hex screws, 8 mm × 30 mm, specified grade”, three packs at R75 per pack, with a note confirming 20 units per pack. AI proposes that both lines match the requirement. The technical reviewer confirms that the differing bolt and screw labels refer to an acceptable equivalent in this example.
The confirmed comparison quantity is 60 for each. A’s goods subtotal is R240. B’s is R225, calculated as three packs multiplied by R75. Those fictional subtotals do not yet establish the preferred supplier: delivery and the stated VAT basis still need alignment.
Supplier C lists “M8 stainless fasteners”, three packs at R70 per pack, with no length, grade or pack contents. The expected handling is review-needed, with no inferred unit price. Procurement asks for the missing attributes and pack size.
If B’s description also appears in a delivery schedule, retain both source references. A reviewer checks whether it repeats the priced line or represents an additional order. Until resolved, do not count it twice or delete it automatically.
7. Evaluate the process before relying on its comparisons
Evaluate against a human-reviewed reference set, including difficult quotations rather than only clean tables. This proposed process should earn trust through observed results, not through polished output.
Include differing terminology, poor scans, unknown pack sizes, bundles, alternatives and revised quotations. Record extraction correctness separately from mapping correctness. A correctly extracted description can still be mapped to the wrong requirement.
Measure false equivalence, missed matches, unsupported conversions, lost exclusions and unresolved exceptions wrongly shown as complete. Also record reviewer effort and how often supplier clarification remains necessary. Decide acceptance criteria with procurement and technical owners before evaluation; there is no universal safe matching percentage.
Re-evaluate when the item category, supplier layouts, extraction method or matching rules change. Potentially easier review is a benefit to test, not a guaranteed result. For design options, the custom AI agents workflow resource and custom AI agent glossary entry provide related context. A fixed workflow may still be the more suitable choice when the inputs and rules are stable.
FAQs about matching quotation lines
Can AI match a generic description to a branded part?
It can propose a candidate, but a generic description may omit compatibility, grade or warranty requirements. Ask the technical owner what evidence establishes equivalence. If the branded part is mandatory, keep the generic offer as an alternative or an unresolved match. Do not accept it solely because the words describe the same broad product category. Preserve the supplier’s wording and request a datasheet or exact part reference where necessary.
What should happen when a supplier leaves out a required item?
Show the required item as a gap for that supplier, not as a zero-cost line. Ask whether it is excluded, bundled elsewhere or accidentally omitted. If the supplier later provides a price, retain the clarification with its date and update the comparison version. Until then, present the confirmed subtotal with the missing scope visible. A planning allowance, if authorised, should remain separate and explicitly hypothetical.
Can the system select the cheapest supplier automatically?
This proposed workflow should prepare evidence for a person, not award the purchase. Even confirmed prices can sit alongside different delivery commitments, payment terms or contractual risks. The buyer must decide which trade-offs are acceptable, with specialist review where consequences warrant it. Keep the recommendation traceable to reviewed rows and unresolved conditions, rather than letting an apparently precise grand total conceal an incomplete offer.
If your business needs a repeatable way to prepare these comparisons, explore AI automation and get in touch about a scoped workflow. Start with a representative quotation set, the required comparison fields and named reviewers, then evaluate the output before using it to support purchasing decisions.