Procurement leadership for AI-enabled work

Define the outcome before selecting a product

Start with the user problem, current workflow, intended outcome, affected people, decision owner, data involved, and consequence of failure. Record non-AI alternatives and the option to change the process without buying a new system. Avoid writing requirements around a favored model or vendor before the team has established what must be achieved.

Assemble a multidisciplinary group and classify each member as owner, reviewer, adviser, or collaborator. Procurement can own market engagement, evaluation process, and contract mechanics. Functional owners define acceptable work; technology validates integration; security and privacy assess their risks; legal interprets obligations; and an accountable service owner accepts operational use.

The UK government's Guidelines for AI procurement recommend starting from the challenge, assessing data and risks, involving multidisciplinary expertise, setting output-based requirements, and planning ongoing contract management. The guidance is aimed at public-sector procurement and is not a universal legal template.

Sources: [1]

Build the AI sourcing evidence pack

Create sections for the problem statement, users, alternatives, source data, intended output, performance conditions, prohibited uses, human oversight, accessibility, security, privacy, portability, audit evidence, incident duties, change notification, exit plan, and renewal triggers. For each supplier claim, identify the requested evidence and the test that will be run.

Separate product capability from configured service behavior. A feature may exist but be unavailable in a plan, disabled by configuration, constrained by permissions, or changed by an integration. Require evidence for the exact offering and environment under consideration, including model or service updates that could affect the acceptance result.

The Cabinet Office's PPN 02/24 on transparency of AI use in procurement highlights identifying and managing AI-related risks in commercial activity and using proportionate controls for confidential procurement information. Its mandatory scope is limited to named UK central-government bodies; other buyers should apply their own rules and advice.

Sources: [2]

Worked artifact: a synthetic knowledge assistant purchase

Imagine a synthetic professional-services team considering an assistant that searches internal policy documents and drafts staff answers. One supplier states that the product provides grounded answers and enterprise privacy. The evidence pack converts these phrases into tests: citation accuracy against approved documents, behavior when sources conflict, tenant access checks, retention configuration, export controls, update notification, and a safe response when authority is missing.

During a sandbox test, the assistant cites an obsolete policy after a newer document has been uploaded. The team does not average that failure against fluent answers. It marks current-source selection as a release-blocking requirement, asks for configuration and remediation evidence, and keeps the service out of production.

A second test passes only after the supplier demonstrates source precedence and the buyer reproduces it with a varied document set. The contract draft assigns change notification, incident cooperation, evidence access, and exit responsibilities, while the service owner retains the deployment decision. This synthetic example is not evidence of Two Prune procurement outcomes.

Verify throughout the contract

Repeat acceptance tests after material model, feature, integration, data, permission, or policy changes. Review incidents, exceptions, unresolved defects, user feedback, actual review burden, and exit readiness before expansion or renewal. Supplier attestations, demos, and contract language are inputs, not operating proof by themselves.

Two Prune can provide task-bounded evidence about how a participant frames a sourcing problem, navigates supplier evidence, verifies claims, and escalates in a controlled simulation. It does not select vendors, certify products, provide legal advice, or make employment decisions.