Agent Procurement Desk

Agent procurement desk

Vendor selection goes wrong when the comparison is built from vendor marketing rather than from your requirements, so every option looks best on its own page. A procurement desk fixes the criteria first and scores everything against the same grid.

Team

• Requirements owner: turns the need into weighted, testable criteria before any vendor is looked at.
• Researchers: one per shortlisted vendor, filling the same grid from documentation and trials.
• Risk reviewer (vendor-data-processing): checks security, data handling, lock-in, and viability.
• Synthesiser: produces the comparison and a recommendation.

Shape: criteria first, parallel per-vendor research, one comparison pack.

Method

1. Write weighted criteria before looking at vendors. Must-have, should-have, and nice-to-have with weights, derived from your use case. Criteria written after demos always match whoever demoed best.
2. Score every vendor on the same grid. Same questions, same evidence standard, and an explicit unknown where the answer is not published, since unknowns are themselves a finding.
3. Test the two or three that matter. Documentation describes the happy path; a trial against your real workload finds the limits that decide the outcome.
4. Price the total cost. Licence plus implementation, migration, training, overage, and the cost of leaving. Headline price is rarely the deciding number.
5. Review data and security terms as a gate, not a column. Where data goes, who processes it, deletion and export paths, and subprocessors (see cross-border-transfers, data-classification).
6. Assess exit before entry. Export formats, contract notice, and what breaks when you leave. Lock-in discovered later is the most expensive finding in procurement.
7. Recommend with the runner-up and the reason. A single option presented as obvious hides the trade the decision maker needs.

Run it

In Claude Code, write the criteria file first, spawn one researcher subagent per vendor to fill an identical grid file, run the risk reviewer over all of them, then synthesise into a comparison. Purchase and signature remain human. Port to CrewAI as parallel research tasks into one synthesis task.

Signals it works

• Criteria and weights predate the vendor list.
• Every grid has explicit unknowns rather than optimistic assumptions.
• Total cost and exit cost appear alongside the headline price.

Boundaries

Agents research and compare; humans negotiate, sign, and pay. Never let an agent enter into agreements, accept terms of service, or share confidential material with a vendor during evaluation. Contract terms need legal review (see agent-legal-desk), and vendor claims about security should be verified rather than believed.