A supplier may be described as unreliable after one difficult delivery, while another receives praise despite several undocumented problems. An assistant can help make the review more consistent by organising evidence and explaining approved measures. It should not convert staff impressions into facts or treat a tiny sample as a complete picture of the supplier.
This article proposes a source-linked review built from delivery, defect and response records. The worked example is entirely fictional. Its purpose is to prepare a fairer review conversation, not to declare a supplier suitable, impose consequences or claim that a generated score predicts future performance.
Define the review period and eligible evidence
Choose the supplier reference, covered period and approved record sources. Include the types of work represented, because different products and project conditions may affect comparability. A review of small repeat orders does not automatically describe the supplier's performance on a large installation.
Identify eligible delivery records, accepted inspection records and response events. Record exclusions with reasons, such as open orders that have not reached their agreed date or test records. Do not quietly remove difficult cases to improve the result, and do not treat unresolved records as confirmed failures.
Keep source coverage visible. If a mailbox or inspection register is incomplete, state which measure is affected. The reader should know whether a number describes every eligible record or only the available subset.
Set measures before asking the assistant for a judgement
A delivery measure needs an approved commitment date and actual receipt rule. A defect measure needs a defined unit and accepted inspection outcome. A response measure needs a start event, qualifying response and elapsed-time or working-time convention. These choices belong to the process owner.
Separate records from opinions. A comment that communication felt slow is useful feedback, but it is different from a measured response time. Label the comment, its source and context. Do not use it to fill a missing timestamp or transform it into an unsupported numerical rating.
OpenAI's Structured Outputs guide supports schema-constrained responses and refusal handling. That can keep metric inputs, evidence, limitations and feedback categories explicit. It does not verify the underlying record or prove that the chosen metric fairly represents the supplier. Source: OpenAI structured outputs
Source-linked supplier review
Use this complete fictional review as a reusable format. The process owner has approved the fixture definitions below; these are not default procurement rules.
Supplier: TEST-SUPPLIER-A. Coverage: four eligible completed delivery records, forty inspected units and three measurable response events within the selected test period. One additional response event lacks a required timestamp and is excluded from the response calculation with that gap reported.
| Review area | Verified fixture evidence | Calculation or summary | Limitation and review question |
|---|---|---|---|
| Delivery timing | D01–D04 contain approved commitment and receipt dates; three meet the fixture timing rule | Three of four eligible deliveries, or 75%, meet the rule | Small fictional sample; inspect the late case and context before judgement |
| Recorded defects | I01–I04 identify two accepted defects among forty inspected units | 2 ÷ 40 × 100 = 5% recorded defect rate | Applies only to inspected units under this definition; uninspected goods are not assumed defect-free |
| Response timing | R01–R03 show elapsed responses of two, four and eight hours | Median elapsed response is four hours | Three measurable events; one additional event has a missing timestamp |
| Staff feedback | F01 says escalation was difficult; F02 praises a clear explanation | Two labelled subjective observations | Do not treat either as a measured fact or representative survey |
| Open evidence gaps | Missing response timestamp and unresolved cause of one late delivery | Named follow-up items | Owner decides what evidence can resolve the gaps |
Attach a source register with event reference, supplier match, purchase or item context, original outcome, verification status and permitted access. Each metric stores its definition, numerator or input values, denominator where applicable, exclusions, calculation version and reviewer. Preserve accepted corrections rather than changing the historical result without explanation.
Review conclusion: the fixture supports the displayed delivery, defect and response measures within its coverage. It does not establish overall supplier quality. The review owner should inspect the late delivery, the defect context and communication feedback, and decide what follow-up is appropriate under the organisation's process.
Acceptance cases: a complete normal record; a missing commitment date; a defect report not yet verified; duplicate receipt records; inconsistent units; and a subjective comment stated as fact. The workflow must hold or qualify the affected measure, avoid duplicate counts, refuse incompatible denominators and retain the feedback label.
The review is accepted when every quantitative claim can be reproduced from approved evidence, every coverage gap is visible and qualitative feedback remains distinct. No automatic blacklist, award decision or future-performance claim is generated from this output.
Link summaries to the records actually reviewed
OpenAI's file-search guide describes retrieval from uploaded documents and citations. It can help locate delivery reports or inspection notes, but does not establish that the uploaded collection contains every relevant record or that a cited interpretation is correct. Source: OpenAI file search
Maintain an approved event inventory alongside retrieved excerpts. Check that the supplier and item references match, and that the source version was verified. If retrieval misses a record present in the inventory, record the coverage failure rather than treating the missing event as absent from reality.
Restrict source links to authorised reviewers. A supplier review may contain commercial details, staff comments and customer-related material. The visible summary should include only the detail needed by its audience, with original access controlled through the established record system.
Keep calculations deterministic and updates traceable
Calculate counts, proportions and median values through explicit rules rather than asking the model to estimate them. For the fictional response inputs two, four and eight, the middle value is four. For the delivery proportion, three divided by four is 75%. These examples demonstrate arithmetic, not measured business outcomes.
If a spreadsheet holds the review, keep its metric definitions and source references aligned with the cells. Google Sheets documents grouped updates to spreadsheet details and a separate values resource. Those editing capabilities do not validate your metric definitions or prevent an inappropriate comparison between suppliers. Source: Google Sheets batch updates
An approved correction should identify the changed source, affected measure and new calculation. Preserve the previous version's reference so the reviewer can understand why the result moved. Do not let an assistant rewrite the summary while leaving the table on an earlier dataset.
Work through normal, missing and duplicate events
In a hypothetical normal case, each delivery has a verified commitment and receipt record. The application calculates the timing measure and the assistant explains the result with coverage and references. The owner can inspect the exceptional case before deciding on follow-up.
In a missing-evidence case, a delivery lacks a verified commitment date. The workflow cannot determine whether it met that timing rule. It reports the unresolved event and its effect on the measure rather than classifying it as either on time or late to complete the percentage.
In a duplicate case, two imports refer to one verified receipt event. The proposed event-identity rule counts it once. If they disagree on quantity or date, hold the affected calculation until an owner resolves the source conflict. Similar supplier names are not sufficient grounds for merging event records.
FAQ about evidence-based supplier reviews
Does a high percentage prove a supplier is dependable?
No. Review the definition, sample size, coverage and work represented. The measure describes its verified dataset under the chosen rule, not every future delivery or the supplier's overall suitability.
Can staff feedback be included when it is subjective?
Yes, if clearly labelled with context and separated from verified measures. It can guide review questions, but should not fill missing records or become an unsupported score.
Should the agent recommend excluding a supplier automatically?
That is outside this proposed review output. The responsible procurement process decides the significance of evidence and appropriate action. The assistant prepares the traceable record and open questions.
If your business needs help defining this process, explore Custom AI agents, the wider AI automation services, and our custom-agent workflow guide. The agents and automation comparison and custom AI agent glossary explain the terms. To discuss your records and approval rules, get in touch.