Chapter 11 · Enterprise AI Readiness Checklist

Enterprise AI readiness is the combined ability to deliver a defined business outcome, control non-deterministic behavior, operate the workload, protect data, investigate failures, and recover verified business state.

11.1 Assessment method

Complete the review with business, product, architecture, AI engineering, operations, security, data, legal or compliance, and support representatives. Score the current production candidate rather than the intended target state.

0Unknown or not addressed

1Documented intention

2Implemented and tested

3Production evidence and recovery

Critical rule: a high average cannot override missing identity enforcement, unresolved data exposure, unverified high-impact action, legal prohibition, or absent recovery.

Gate 01

Business and use-case readiness

The target workflow, users, autonomy level, and measurable outcome are defined.

A named business owner accepts accountability for results and failure consequences.

Baseline cycle time, cost, quality, or risk is measured.

Success metrics distinguish adoption from business value.

Manual fallback and escalation remain usable.

Pilot scope matches the intended population, language, region, and channel.

Required evidence

Business case, workflow map, baseline metrics, success thresholds, autonomy statement.

Launch blocker

The use case is described only as productivity improvement without an owner or measurable task.

Gate 02

Ownership, governance, and lifecycle

Business, product, technical, data, security, and operations owners are named.

Agents, capabilities, models, prompts, policies, and knowledge sources are registered and versioned.

Risk acceptance has an accountable approver, conditions, and expiry.

Change, emergency suspension, access review, and retirement procedures exist.

Legal, regulatory, privacy, records, and sector obligations are assessed.

Expansion of autonomy or impact requires a new review.

Required evidence

RACI, asset registry, risk register, approval record, lifecycle and retirement policy.

Launch blocker

A pilot has become business-critical but has no permanent owner or support model.

Gate 03

Architecture and integration readiness

Current logical, deployment, trust-boundary, and data-flow diagrams exist.

Identity, policy, workflow state, credentials, validation, and audit are deterministic.

Workflow state is durable and separate from conversation memory.

Business capabilities replace raw SQL, shell, generic HTTP, and broad admin APIs.

Capability contracts define schema, authorization, idempotency, verification, and compensation.

Integration patterns match latency, transaction, approval, and recovery requirements.

Required evidence

Architecture diagrams, capability catalog, integration contracts, ADRs.

Launch blocker

One agent loop owns permissions, workflow state, tools, and final business execution.

Gate 04

Data, grounding, and knowledge readiness

Sources have owners, authority status, classification, and approved purpose.

Retrieval applies user, tenant, field, and resource access before model context.

Freshness thresholds and conflict precedence are defined.

Indexes preserve source references, versions, ACLs, and deletion state.

Context budgets limit irrelevant and excessive data.

User, tenant, workflow, and agent memory are isolated.

Required evidence

Data inventory, source contracts, retrieval tests, lineage, retention and deletion evidence.

Launch blocker

The model can retrieve data outside the initiating user's access.

Gate 05

Security, privacy, and threat readiness

Human, application, agent, orchestrator, and connector identities are distinguishable.

Protected actions use short-lived scoped credentials and deterministic authorization.

Prompt injection cannot change identity, policy, tool scope, or destinations.

Tool parameters are schema and domain validated.

DLP covers input, context, output, files, tool parameters, and logs.

Incident response can suspend agents, models, capabilities, sources, and connectors independently.

Required evidence

Threat model, identity design, policy tests, red-team results, DLP and incident playbook.

Launch blocker

Security controls exist only in prompts or the model has broad raw system access.

Gate 06

Evaluation, testing, and quality readiness

A versioned evaluation suite covers representative normal tasks.

Edge, adversarial, multilingual, and high-impact scenarios are included where relevant.

Grounded claims can be traced to approved sources.

Unauthorized data and tool attempts are tested and blocked.

Release and rollback thresholds are approved.

Production sampling, drift, corrections, and incident-derived regression tests are implemented.

Required evidence

Evaluation dataset, metric definitions, report, reviewer guide, canary and monitoring plan.

Launch blocker

Quality approval depends on a few hand-selected demonstrations.

Gate 07

Reliability, continuity, and recovery readiness

Task-specific SLOs cover latency, availability, completion, freshness, and verified execution.

Timeouts, retries, queues, circuit breakers, and bulkheads are coordinated.

Write actions use stable idempotency keys and target post-condition checks.

Unknown transaction state is reconciled before retry.

Partial completion has compensation or reconciliation.

Rollback, restore, and disaster-recovery exercises have succeeded.

Required evidence

SLOs, failure analysis, resilience tests, DR exercise, compensation and rollback evidence.

Launch blocker

A timeout can trigger an unverified retry of a high-impact action.

Gate 08

Operations and GenAIOps readiness

Distributed traces connect request, sources, model, policy, tools, and outcome.

Dashboards cover service, model, retrieval, agent, security, cost, and business metrics.

Sensitive evidence is redacted or restricted.

Alerts have owners, severity, thresholds, and runbooks.

Deployments support canary, progressive rollout, rollback, and emergency suspension.

Support and on-call teams are trained and can access required evidence.

Required evidence

Dashboards, alerts, trace example, runbooks, deployment pipeline and on-call readiness.

Launch blocker

The team can observe only the final response and cannot reconstruct an incident.

Gate 09

Performance, capacity, and scale readiness

Latency budgets are defined by task class and stage.

Average, peak, and burst demand are documented and load-tested.

Model, search, database, API, and connector quotas are monitored.

Admission control and priority rules protect critical workloads.

Tenant fairness and context limits are enforced.

Fallback routing preserves quality, region, security, and data policy.

Required evidence

Capacity model, load-test report, quota dashboard, scaling and saturation behavior.

Launch blocker

Production estimates extrapolate from a single-user demonstration.

Gate 10

Cost, efficiency, and value readiness

Cost is allocated by use case, tenant, model, workflow, and owner.

Per-task budgets limit model calls, context, steps, runtime, and tools.

Human approval, correction, support, evaluation, and recovery costs are included.

Cost per successful business outcome is measured.

Budget alerts and restriction thresholds are defined.

The business case remains positive under realistic adoption and error rates.

Required evidence

Unit economics, budget policy, cost dashboard, scenarios and finance sign-off.

Launch blocker

The business case counts token cost but ignores integration and operating cost.

Gate 11

People, process, and change readiness

Target users receive role-specific training.

The interface distinguishes source facts, generated interpretation, draft, approval, and executed state.

Users know how to verify output and report suspicious or incorrect behavior.

Human approvers receive action, target, evidence, risk, and rollback details.

Support, security, legal, compliance, and operations teams are prepared.

A sanctioned platform and clear data-use rules reduce shadow AI.

Required evidence

Training, user guidance, support model, communications, approval UX and feedback workflow.

Launch blocker

Users are expected to control risk without training, evidence, or escalation authority.

Gate 12

Production launch decision

All critical gates meet required scores and no critical blocker remains.

Residual risks have owners, conditions, evidence, and expiry.

Launch scope matches the evaluated population, data, regions, languages, and autonomy.

Canary size, duration, stop conditions, rollback owner, and metrics are defined.

Support and incident response are active before the first production user.

A post-launch review date and expansion criteria are scheduled.

Required evidence

Signed readiness record, risk acceptance, canary plan, rollback plan and review date.

Launch blocker

The launch relies on monitoring closely without measurable stop conditions or rollback authority.

11.2 Readiness scorecard

Decision	Meaning	Required action
Ready	Critical gates meet target scores and no blockers remain.	Launch through the approved canary plan.
Conditionally ready	Noncritical gaps have owners, deadlines, and compensating controls.	Limit scope and track conditions to closure.
Pilot only	Controls support a restricted, reversible use but not target scale or impact.	Keep population and autonomy bounded.
Not ready	One or more critical blockers remain.	Do not launch; remediate and repeat the review.

11.3 Minimum production evidence pack

The evidence pack must be versioned and tied to the exact agent, model, policy, connector, source, and evaluation versions covered by approval.

Illustrative readiness record

readiness_review:
  workload: enterprise-contract-assistant
  release: 2.4.0
  decision: conditional_ready
  approved_scope:
    users: legal-operations-eu
    autonomy: recommend_and_create_draft
    data_classification: confidential
  critical_gates:
    security: 3
    evaluation: 3
    reliability: 2
  conditions:
    - owner: Platform Operations
      action: complete secondary-region recovery exercise
      due: 2026-09-15
  evidence:
    architecture: ADR-set-2.4
    threat_model: TM-2026-08
    evaluation_report: EV-2.4.0
    rollback_test: RB-42
  next_review: 2026-10-01

11.4 Chapter summary

An enterprise AI workload is ready only when the organization can define its purpose, constrain its authority, evaluate its behavior, observe its operation, and recover verified business state after failure.

Core conclusion: production readiness is the ability to operate uncertainty with evidence, ownership, and reversible control.

Reference foundations

Microsoft Azure Well-Architected Framework — AI workload documentation, assessment, operations, and architecture patterns.
NIST AI Risk Management Framework: Generative AI Profile.
AetherStaff Enterprise AI Integration Chapters 1–10.