Where should AI sit?

One label hides several jobs.

An assistant that compares quotations has a different responsibility from an agent that sends an offer. Curhan and Gratch distinguish four positions for AI: representing a principal, coaching a negotiator, supporting internal coordination, and facilitating between parties. Their framework preserves meaningful human participation across these roles. [S1]

For Procurement Lead, this suggests a deliberate progression: establish the facts and the mandate, support a person’s judgement, then consider bounded external negotiation. This progression is our product interpretation, rather than a sequence prescribed by the paper.

A mandate is a product object.

AgenticPay gives buyers and sellers private states such as valuations and price constraints, separate from shared product context and dialogue. Its parser turns messages into structured negotiation actions. That distinction matters: a fluent sentence is not sufficient evidence that the system understood a commercial term correctly. [S3]

Our proposed mandate would record the specification, volume, acceptable tradeoffs, disclosure permissions, and escalation rules. A structured proposal would carry price, quantity, lead time, freight responsibility, and payment terms for review. A person should resolve missing or conflicting terms before authorization.

Commercial context changes the answer.

Waditwar describes a procurement architecture connecting historical data, market intelligence, negotiation, risk analysis, and approvals. It also identifies weak data and unstructured conditions as obstacles. We take the architecture as a useful checklist of dependencies, without treating the paper’s illustrative savings figures as validated product outcomes. [S5]

A larger order might reduce unit price while increasing inventory cost. A shorter lead time might improve service resilience. The existing cost lab makes several such tradeoffs explicit; it excludes other costs and does not optimize contract terms. An agent should use an agreed calculation, rather than inventing a rationale for a preferred supplier.

Start with a role people can trust.

A first pilot can test whether the team spends less effort assembling a decision, whether reviewers can trace the assumptions, and whether proposed terms remain inside the mandate. External negotiation adds another test: does the supplier understand and accept the agent’s role? The human approval step becomes a concrete review of evidence and terms.