When three proposals describe three different realities
A procurement team asks three suppliers to quote the same service. One sends a spreadsheet with unit prices, another sends a PDF with a bundled amount, and the third includes important conditions only in the email body. Delivery dates use different references, taxes appear in different fields and service levels are not directly comparable. The analyst spends hours rebuilding the proposals before anyone can discuss the decision.
An AI agent can read these materials, extract terms and prepare a comparison. The risk begins when a fluent summary hides missing information or turns a business preference into an apparently objective ranking. The goal should be to improve analysis, not to make the decision disappear inside the system.
Define the comparison model before reading proposals
Start with a common schema approved by procurement and the requesting area. It may include total cost, payment terms, implementation time, service scope, mandatory certifications, support conditions, contractual restrictions and risks. Each field needs a definition, expected unit, source and rule for missing values.
Also separate eliminatory requirements from weighted preferences. A missing mandatory license is not equivalent to a slightly longer delivery time. The agent should apply the policy that the company has defined, not invent weights from the wording of the proposals.
NIST AI RMF Map guidance recommends documenting the specific tasks supported by the AI system, its knowledge limits, how outputs will be used and how humans will oversee them. In procurement, this means stating whether the agent extracts, checks, compares or recommends, and where its authority ends.
Separate extraction, normalization, recommendation and decision
Treat the workflow as four different stages. Extraction identifies clauses and values in the original files. Normalization converts dates, currencies and units into the agreed format without changing their meaning. Recommendation applies explicit business criteria to the normalized record. Decision belongs to the authorized people and considers context that may not be present in the documents.
Each stage should produce a structured output and a visible status. A value can be confirmed, inferred, conflicting or absent. If two documents from the same supplier disagree, the agent should preserve both references and open a question instead of silently selecting the most convenient one.
OpenAI guidance for building agents recommends well-defined tools, clear instructions and human intervention for sensitive or high-impact actions. Supplier selection may affect cost, continuity, contractual exposure and relationships. The agent can prepare the evidence, but approval and commitment should remain under the company's authorization rules.
Make every important field traceable to evidence
A comparison table is trustworthy only when reviewers can reach the original evidence. Store the document identifier, page or section, excerpt location, extraction time and file version for each relevant value. When the proposal changes, create a new version and identify which fields were affected.
Do not fill gaps with estimates unless the policy explicitly permits them. Show missing information as missing, record who requested clarification and prevent a final recommendation when an eliminatory field remains unresolved. This is more useful than a polished table that creates false certainty.
A hypothetical example: packaging contract
Consider a hypothetical food manufacturer, not an NM Solutions customer case. The company requests proposals from three packaging suppliers. The files contain prices by different quantities, freight rules, printing setup fees, minimum orders and quality certificates.
The agent extracts every commercial condition into a shared schema and converts prices to the same quantity range. It does not treat an omitted freight value as zero. Instead, it marks the field as absent and links the status to the source document. A deterministic rule checks mandatory certificates and blocks any proposal that lacks valid evidence from entering the recommended group.
The agent then prepares two scenarios using criteria approved by procurement: lowest expected annual cost and lowest supply risk. A buyer reviews the cited evidence, asks suppliers to resolve two conflicts and records the reason for the final choice. The scenario does not prove that the selected supplier will perform as expected, but it makes assumptions, exclusions and responsibility visible.
Evaluate the comparison, not only the final selection
Test the agent with completed purchases and with deliberately difficult proposals. Measure field extraction accuracy, unresolved conflicts, unsupported values, corrections made by reviewers and time spent until an approved comparison is ready. Evaluate different document layouts and supplier writing styles.
Monitor whether reviewers accept suggestions automatically or investigate the evidence. A high acceptance rate is not always a quality signal. It may indicate that the interface makes disagreement difficult or that people are treating the recommendation as a decision. Periodic sampling should compare the agent's record with the original documents.
Recognize limitations and keep the decision accountable
Scanned files, ambiguous clauses and attachments referenced but not delivered can limit extraction. Currency changes, taxes and volume discounts can make normalized comparisons unstable. Historical supplier performance may also reflect different contexts and should not become an unexplained penalty. Confidential proposals require access, retention and sharing controls.
Not every purchase needs an agent. A simple form and deterministic spreadsheet may be enough for standardized, low-volume acquisitions. NM Solutions helps companies map the purchasing workflow, comparison criteria, integrations, evidence and approval points before automation. A diagnostic can identify where an agent reduces analytical work without obscuring the business decision.