Industrial Agentic AI Readiness Diagnostic

Readiness methodologyresearch_methodology

Interactive diagnosticcontrolled_demo

Evidence Agentcontrolled_demo

Operational actionnot_designed

Direct definition

Industrial Agentic AI Readiness is evidence for deciding whether agency is justified.

It evaluates whether a variable industrial workflow genuinely requires goal-driven planning and tool selection, then checks the foundations required to keep that behavior bounded, source-traceable, secure, and accountable.

Current AI2COE boundary

AI2COE offers an equal-weight readiness methodology and a deterministic synthetic read-only demonstration. It does not claim a production operational agent, live MCP/A2A/OPC UA integration, autonomous remediation, or customer-verified outcomes.

Agentic necessity test

Use the least complex operating model that can solve the problem.

Approach	Use when	Control boundary
Analytics	A stable question can be answered from prepared data	No dynamic tool planning
Deterministic workflow / RPA	Steps and exceptions are known in advance	Rules own the sequence
Copilot	A human remains the primary planner and operator	Advice or drafting only
Read-only evidence agent	The goal is bounded but evidence requires multi-step context and tool selection	R0-R2 with human review
Approval-gated workflow agent	A future workflow requires transaction drafts and named approvals	R3-R4; not implemented by AI2COE

Ten readiness dimensions

Assess the complete operating system around the proposed agent.

Use-case fit

Does the workflow require variable multi-step planning rather than analytics, rules, RPA, or a copilot?

Data readiness

Are required fields, lineage, freshness, ownership, retention, and access understood?

Industrial context readiness

Can materials, stock states, sites, suppliers, assets, BOMs, work orders, and criticality be related?

Tool and interoperability readiness

Are approved tools, schemas, errors, timeouts, permissions, and transaction boundaries defined?

Identity and authorization readiness

Does the proposed agent have a unique identity, scoped credentials, least privilege, and auditable authorization?

Decision-rights readiness

Are retrieve, analyse, recommend, draft, submit, and execute rights explicitly separated?

Human-oversight readiness

Are owners, approval thresholds, escalation, override, rejection, deferral, and review service levels defined?

Security and safety readiness

Are injection, data disclosure, tool misuse, cross-tenant exposure, replay, poisoning, and runaway-loop risks controlled?

Evaluation and observability readiness

Can the team measure tool choice, sequence, evidence coverage, abstention, policy compliance, latency, and repeatability?

Value and deployment readiness

Are the baseline, target outcome, owners, pilot population, review capacity, support model, and exit plan defined?

Interactive controlled demo

Run the Industrial Agentic AI Readiness Diagnostic.

Choose the current planning state for each dimension. Inputs are not persisted and must not contain private operational data.

Bounded use-case goalDescribe the decision, not a command to a production system.

Use-case fitDoes the workflow require variable multi-step planning rather than analytics, rules, RPA, or a copilot?

Data readinessAre required fields, lineage, freshness, ownership, retention, and access understood?

Industrial context readinessCan materials, stock states, sites, suppliers, assets, BOMs, work orders, and criticality be related?

Tool and interoperability readinessAre approved tools, schemas, errors, timeouts, permissions, and transaction boundaries defined?

Identity and authorization readinessDoes the proposed agent have a unique identity, scoped credentials, least privilege, and auditable authorization?

Decision-rights readinessAre retrieve, analyse, recommend, draft, submit, and execute rights explicitly separated?

Human-oversight readinessAre owners, approval thresholds, escalation, override, rejection, deferral, and review service levels defined?

Security and safety readinessAre injection, data disclosure, tool misuse, cross-tenant exposure, replay, poisoning, and runaway-loop risks controlled?

Evaluation and observability readinessCan the team measure tool choice, sequence, evidence coverage, abstention, policy compliance, latency, and repeatability?

Value and deployment readinessAre the baseline, target outcome, owners, pilot population, review capacity, support model, and exit plan defined?

Equal weights are a versioned methodology starting point and are not presented as scientifically validated.

Decision rights

Authority increases only after evidence and named approval.

Class	Capability	Current boundary
R0	Retrieve approved information	Allowed in the controlled demo
R1	Analyse and generate evidence	Allowed in the controlled demo
R2	Recommend a human-reviewed action	Allowed in the controlled demo
R3	Draft a transaction or work package	Future; approval design required
R4	Submit after named approval	Future; not implemented
R5	Execute a bounded reversible action	Not near-term
R6	Closed-loop physical control	Out of scope

Human Approval Matrix

Every consequential interpretation has a named decision owner.

Decision	Owner	Required evidence	Safe fallback
Accept a review candidate	Named domain owner	Source references, limitations, confidence, and exception context	Request more data or defer
Assign remediation analysis	Process owner	Accepted finding and accountable assignee	Reject or return to review
Use a financial exposure band	Finance owner	Approved valuation and currency assumptions	Keep as unquantified hypothesis
Advance to a controlled pilot	Executive sponsor, security, data owner	Isolation, identity, policy, evaluation, incident, and exit controls	Remediate foundations first
Execute an operational change	Outside current AI2COE scope	Not available in the controlled demo	Use customer-controlled systems and approvals

Industrial context graph

Connect operating entities before asking an agent to reason across them.

MaterialManufacturer / MPNInventory positionSite / storage stateSupplier / POAsset / BOMWork order / criticalityEvidence / owner / policy

The methodology starts from exported data. OPC UA is assessed as a future semantic-readiness option; this page does not claim a live OPC UA connection or digital-twin feed.

Eight governed skills

The solution coordinates existing engines; it is not a ninth engine.

Skill contract	Industrial IQ engine	Authority	Responsibility
partscleanse.resolve_material_identity.v1	PartsCleanse AI	R1	Material identity and duplicate-family evidence
inventorymind.analyse_stock_posture.v1	InventoryMind AI	R1	Stock posture, reservation, quality-state, and availability evidence
procuremind.review_purchase_path.v1	ProcureMind AI	R1	Purchase history, supplier aliases, emergency flags, and price context
financemind.qualify_exposure.v1	FinanceMind AI	R2	Assumption-labelled financial exposure for owner review
assetmind.map_asset_context.v1	AssetMind AI	R1	Asset, equipment, functional-location, and BOM context
reliabilitymind.review_work_context.v1	ReliabilityMind AI	R1	Work-order, maintenance, and criticality context
readymind.validate_source_fit.v1	ReadyMind AI	R1	Data sufficiency, use-case fit, and readiness limitations
governancemind.apply_evidence_policy.v1	GovernanceMind AI	R2	Evidence class, confidence, policy, owner, and review routing

Read-only Industrial Evidence Agent

Test seven synthetic MRO evidence scenarios.

The deterministic demonstration chooses only the scenario-required versioned skills, preserves source references and legitimate exceptions, enforces an execution budget, abstains when context is insufficient, creates a tamper-evident sample event chain, and routes the result to human review.

Sample readiness report

A decision pack, not a production approval.

A generated result includes use-case fit, ten readiness scores, autonomy recommendation, decision rights, human approval requirements, agent security risks, an evaluation plan, a first-agent recommendation, and a remediation backlog.

Evidence classEstimated for self-declared diagnostic inputs

Default outcomeFoundation remediation, read-only prototype, controlled pilot, or do not use an agent

Owner decisionProceed, constrain, remediate, investigate, or defer

Agent Risk Register

Threats, controls, and residual gaps remain visible.

Risk	Current control	Remaining requirement
Direct or indirect prompt injection	Separate instructions from data; quarantine prompt-like source content	External models and retrieved content require further red-team testing
Excessive agency or unauthorized tools	Fixed skill allowlist, schema validation, R2 ceiling, deny by default	External tool gateway and credentials are not implemented
Cross-tenant disclosure	No tenant or customer data in the synthetic demo	Tenant isolation must pass before any customer pilot
Unsafe HTML or spreadsheet output	Escaped HTML and neutralized formula prefixes	File export paths require continued regression tests
Missing context or false confidence	Source references, evidence classes, exceptions, limitations, and abstention	Domain-owner validation remains mandatory
Runaway planning, latency, or cost	Deterministic fixed plan and no external model calls	Metering and loop limits are required for a future model-backed pilot

Agent evaluation plan

Evaluate the complete trajectory, not only the final answer.

Dimension	Review evidence	Minimum gate
Use-case classification	Fit result and non-agent alternative	Can recommend analytics, workflow, RPA, copilot, read-only agent, or no agent
Tool trajectory	Selected skill IDs, versions, sequence, and schema status	Every tool is allowlisted and R2 or lower
Evidence quality	Source-reference coverage, reason, confidence, limitation, and exception	Every material conclusion has references
Safety	Injection, formula, authority, output, and tenant tests	No source content changes policy or permissions
Abstention	Insufficient-data and missing-context scenarios	Escalate rather than fabricate a conclusion
Review usefulness	Accept, reject, defer, assign, and request-more-data paths	Named human owner retains the decision
Operations	Repeatability, latency, cost, version, and incident records	Baseline first; no public performance percentage yet

Quantitative targets will be defined only after a reproducible baseline exists. No unvalidated evaluation percentage is published.

Interoperability boundary

MCP, A2A, and OPC UA are readiness questions, not current integrations.

MCP may later expose narrowly authorized tools. A2A is deferred until a single orchestrator proves a real collaboration need. OPC UA Companion Specifications may support future semantic context. None is implemented as a live connector in this controlled demo.

What AI2COE does now: methodology, sample scoring, deterministic synthetic skill orchestration, source-referenced evidence, R0-R2 policy, and sample human review.

What AI2COE does not do: operational execution, source-system write-back, external-agent collaboration, live industrial control, or verified customer outcomes for this capability.

FAQ

Questions enterprise teams should resolve before an agent pilot.

What is Industrial Agentic AI Readiness?

It is the ability to prove that a use case, operational data, industrial context, tools, identity, authority, human oversight, security, evaluation, and value model are ready for a bounded agent pilot.

Does every industrial AI use case need an agent?

No. The fit test can recommend analytics, deterministic workflow, optimization, RPA, a copilot, a read-only evidence agent, or no agent.

Is the readiness score a benchmark?

No. It is an equal-weight methodology score from self-declared planning inputs. It is not a certification, market benchmark, or production approval.

Does AI2COE deploy operational agents today?

No. AI2COE currently provides a controlled synthetic demonstration and readiness methodology. It does not execute operational actions or write to source systems.

What does the Evidence Agent demonstration do?

It runs controlled synthetic MRO scenarios through only the required versioned read-only skill contracts, preserves source references and exceptions, abstains when context is insufficient, and produces a human-review decision pack.

Can source records instruct the demonstration?

No. Source text is treated as untrusted data. Prompt-like content is quarantined and cannot alter policy, permissions, authority, or the skill allowlist.

What authority is allowed?

The initial ceiling is R2: retrieve approved information, analyse evidence, and recommend a human-reviewed action.

What is required before a controlled pilot?

Tenant isolation, approved data handling, unique identity, scoped credentials, tool authorization, evaluation datasets, security testing, human acceptance criteria, monitoring, incident response, and an exit plan.