Time to acceptance
Compare median minutes per accepted checklist against matched manual tasks. Include review and rework. Proposed target: 20% reduction.
Governance gate
Use fictional inputs to explore the proposed review route. The highest applicable tier determines the decision.
Proposed policy · v1.0
Applies to employees, contractors and vendors supporting internal or white-label delivery. Proposed accountable owner: executive sponsor. The AI adoption lead maintains the register and coordinates quarterly reviews.
For a workforce of 350+ professionals, assign a central adoption lead, security and privacy reviewers, a champion in every participating function, and one accountable owner per client workflow. Confirm actual headcount, contractors, shift patterns, availability and seats by function during discovery. Begin with two teams of five to eight volunteers each, an illustrative cohort size subject to manager approval.
Train champions first. Use role-specific examples and supervised tasks for operators, separate sign-off practice for reviewers and decision briefings for leadership. Coordinate with the existing AI practice on vendor evaluations and ensure white-label partner enablement uses permitted material.
Classify inputs before use. Strip identifiers and use synthetic samples for testing. Confirm agency and end-client instructions for the assigned account, including permitted tools, subprocessors, location, attribution, IP rights and deletion periods. A partner’s approval only covers the rights it holds.
Use a separate workspace and access group per client. Keep credentials in a secret manager, restrict exports and sharing, and prohibit training on client material unless specifically authorized. Log minimum necessary metadata with an approved retention and deletion schedule.
IT maintains an approved-tool register covering owner, purpose, model version, terms, data location, retention, training controls, access controls and incident contacts. Reassess changes to vendors, integrations, model versions or client scope.
Start integrations read-only. Scope each service identity to one client and task. Treat retrieved webpages, emails and documents as untrusted data. Validate tool requests against an allowlist and require a named human to approve consequential writes. Define spend caps, timeouts, a stop control and revocation owner before trials.
The use-case owner records purpose, tier, data class, approved tool, accountable reviewer, baseline, model version, observed errors and exit criteria. Review accuracy, accessibility, bias, confidentiality and brand fit against examples representing typical and difficult work.
Train each pilot cohort on the approved task, escalation path and manual fallback. Require a supervised example before access. Reviewers sign off on deliverables in the agency’s existing delivery system. Champions hold weekly office hours and record recurring failure patterns.
Any operator can stop a workflow. Disable the integration, revoke exposed access, preserve minimal evidence in the authorized incident system and notify security and the account owner immediately. The incident lead assesses affected clients, required notices and contractual deadlines. Revert to the manual process.
For T1–T3 exceptions, record scope, rationale, compensating controls, approver and expiry. T4 requires redesign. Resume after the owner verifies containment and corrective evaluations and obtains tier-appropriate approval. On retirement revoke integrations, delete or return retained data under contract and close the register entry.
Pilot scorecard
Suggested decision thresholds for a two-week pilot. All baselines and observations await E2M participation.
Compare median minutes per accepted checklist against matched manual tasks. Include review and rework. Proposed target: 20% reduction.
Compare independently reviewed acceptance rates and critical omissions. Require zero confidentiality incidents, zero critical omissions and acceptance at least equal to baseline.
Track weekly active trained users divided by eligible trained users, plus tool and review cost per accepted task. Interview non-users and resolve blockers before expanding.
Evidence and interpretation
E2M describes its team and partner reach in its role postings. LinkedIn provides a broader employee band. The proposed controls adapt the Govern, Map, Measure and Manage functions of the NIST AI Risk Management Framework. E2M approval and internal baselines remain pending.