Enterprise AI checklist — pre-pilot and pre-production gates

Sponsor and scope: executive sponsor named; workflow owner with authority to change process; single use case scoped with baseline metric you own (not vendor ROI).

Data: representative samples available; classification rules documented; prohibited data classes listed; legal/security sign-off on samples used in proof.

Process: workflow documented including exceptions; escalation paths exist today; success and stop criteria written in pilot charter.

Integration: source systems identified; read/write requirements defined; SSO approach agreed; entity boundaries mapped for multi-company groups.

Governance: policy or interim rules for approved tools; logging requirements defined; human override requirement classified by risk tier; model/vendor change notification process drafted.

Talent: named internal product or process owner; support model for hypercare identified; ops capacity for monitoring acknowledged.

Commercial: build-vs-buy decision recorded—/resources/guides/build-vs-buy-ai; contract exit clauses reviewed if buying; inference cost owner named.

Integration complete: production connectors tested; SSO enforced; orphan accounts eliminated; write-back rules validated on messy data.

Security: encryption and access control verified; DLP alignment confirmed; penetration test completed if required; vendor subprocessors documented.

Observability: logging active per retention policy; alerting on error spikes; dashboards operators will use; prompt/output sampling method approved by security.

Human oversight: override UI tested under load; escalation paths exercised; high-risk decisions require human sign-off by design.

Performance: latency and throughput at peak volume (month-end, campaign, hiring season); async design where needed.

Runbooks: incident response; rollback and version pin; vendor escalation contacts; inference cost monitoring.

Enablement: role-based training complete for launch cohort; hypercare schedule published; champions named per function.

Sign-off: business sponsor, model owner, security, and support—not engineering alone.

Not every workflow needs the same bar. Tier examples:

Tier 1 — internal draft assist, low sensitivity: lighter logging; manager review optional. Tier 2 — operational workflows with PII: full logging, override, retention rules. Tier 3 — customer-facing or credit-adjacent: human decision on outcomes; enhanced audit; legal review before go-live.

Document tier in the checklist. Auditors sample Tier 3 first. Applying Tier 3 bureaucracy to Tier 1 slows adoption unnecessarily; applying Tier 1 laxity to Tier 3 creates findings.

See /resources/guides/ai-governance-framework and /ai-consulting/ai-governance-risk for programme design.

Pilot on synthetic data only—production edge cases break trust at cutover. Policy published but managers never briefed on enforcement. Licences purchased before integration owners assigned. Logging "enabled" but nobody reviews samples. English-only training for bilingual operators. Go-live without rollback plan when model updates degrade quality. Shadow mode skipped for high-risk workflows. Hypercare ends before metrics stabilise.

Keep a short incident log when gates fail post-cutover—feeds the next checklist revision and onboarding for new managers.

We embed checklist gates in pilot charters before build starts. Readiness assessments produce evidence for pre-pilot items; production cutovers use the pre-production section as sign-off criteria.

We report fail gates honestly with remediation sequencing rather than ceremonial go-lives. Handover packages include completed checklists, runbooks, and dashboard links.

Engagements reference /products/approvals and /solutions paths when workflows need governed sign-off. Global enterprises use the same gates whether operating in Bangladesh, UAE, US, or UK—local policy overlays sit on top, not instead of, operational evidence.

Pre-pilot pass requires: sponsor, owner, metric, samples, policy interim, integration map, governance tier, ops owner, commercial record.

Pre-production pass requires: prod integration, security sign-off, logging live, override tested, performance at peak, runbook accepted, training done, multi-party sign-off.

Re-run performance and override tests after every model or vendor update before expanding cohorts. Schedule six-month checklist revisit on every production system.

Store signed PDFs or ticket links in GRC or intranet—production claims need evidence, not Slack reactions alone.

Internal audit should sample one completed checklist per quarter in 2026. Missing artefacts indicate governance theatre—programmes expand on narrative instead of evidence.

Assign a single checklist owner per programme—often PMO, risk, or enterprise architecture—not a rotating vendor PM. Store checklists in systems auditors already use: GRC, Jira epics, or SharePoint libraries with version history. Email threads fail audits.

Use the pre-pilot section at charter sign-off; freeze scope when pre-pilot fails without written risk acceptance. Use pre-production gates as the cutover meeting agenda—each fail gets remediation ticket, owner, and date before expansion.

For multi-country groups, add a local overlay appendix: residency, language, sector regulator—without removing global gates. Bangladesh RMG overlays differ from UAE free-zone overlays; both still need integration and logging evidence.

Arcloops exports completed checklists at engagement end so internal teams inherit operational gates, not consultant memory alone.

Link each gate to a ticket template in your PM tool so pass/fail is searchable. Auditors request samples by gate ID—not by asking who remembers the go-live meeting.

Failed gates should not always stop the programme—they should stop uncontrolled expansion. Document fail severity: hard stop (no production data until fixed), soft stop (limited cohort only), or accepted risk (executive sign-off with expiry date).

Hard stops include missing SSO, no logging on Tier 2+ data, or absent override on customer-impacting workflows. Soft stops include incomplete hypercare schedule or training below eighty percent of launch cohort—expand only to champions until remediated.

Programme managers should review gate status weekly during pilot, biweekly during hypercare, monthly in production. Arcloops flags fail gates in steering notes with remediation owners—no surprise cutovers at month-end to satisfy vendor commercials.

Export gate status to steering dashboards sponsors already read—duplicate PMO spreadsheets nobody opens do not count as governance.