AI, data & platform governance

One governance route from idea to operation

Use this page to decide what can proceed, who must be involved, what evidence is required, and when work must stop or escalate.

Back to tools

Start here

This is the single governance route for AI Marketplace tools, AI use cases, automations, dashboards, agents, and data-heavy requests. Governance is proportionate to risk, but every request follows the same lifecycle.

1

Classify

Define the outcome, users, data, impact, and whether AI or automation is actually needed.

2

Assign

Name the business, product or app, data, and support owners. Identify required reviewers.

3

Evidence

Complete the relevant data, security, architecture, evaluation, and approval records.

4

Operate

Monitor use, quality, access, cost, incidents, changes, and the eventual retirement decision.

How to use this page: confirm the Marketplace guardrails, select a risk route, then work through the six delivery gates. A high data-readiness score supports a decision; it is not approval by itself.

Marketplace guardrails

These are current rules. They apply to every tool hosted on the AI Marketplace. The final section is explicitly separated because it contains proposals, not approved controls.
RuleWhat it means
You build it, you own itThe app owner maintains the tool, its dependencies, user guidance, and retirement. The central team does not take ownership of the app.
Not an IT-supported serviceThere is no app-level SLA, change-management service, or helpdesk unless a separate service arrangement is agreed.
Internal use onlyMarketplace tools are for internal productivity. Do not share them with clients, use them in front of clients, or reference them in client deliverables.
Experimental by defaultAny tool should be built assuming it could be deprecated within a year.
No work-product retentionUploaded files and results are not stored permanently. Operational access logs may be retained for platform monitoring. Download your outputs before closing.
No critical workflowsDo not make the Marketplace a dependency for a mission-critical process.
Infrastructure is centralThe platform team manages hosting, routing, identity integration, and deployment pipelines; the app owner remains responsible for app behavior.
Onboarding is controlledInitial questions and app registration go through the platform admin before deployment.
Platform cost is centralBase platform infrastructure is not charged back to app owners. Material model, storage, traffic, or third-party costs still need an identified owner and monitoring plan.

Risk and approvals

Choose the highest applicable route. Lower-risk work still requires approved tools, appropriate data, and human review; higher-risk work adds formal review and evidence.

Basic controls

Public or low-risk work

Public or non-sensitive information, individual drafting, summarising, or exploration with no automated action. Use an approved tool and verify the output.

Governed review

Internal or reusable work

Internal data, multi-user tools, stored outputs, connectors, dashboards, or repeatable workflows. Name owners and complete data-readiness and control review before build.

Formal approval

Sensitive or high-impact work

Client confidential, personal, financial, regulated, restricted, external-facing, system write-back, or high-impact decisions. Engage the relevant governance teams before proceeding.

AI security and client data

Any AI use case involving client confidential, personal, financial, regulated, or otherwise sensitive data must go through the AI risk assessment process before proceeding. If there is uncertainty, engage AI Security before entering client data into any AI tool. More guidance is available on the AI Security page.
Review areaRequired questionEvidence or decision
Data and privacyWhat classification, personal data, client restrictions, retention, residency, or cross-border obligations apply?Classification, permitted-use position, access approval, and privacy or DPIA decision where triggered.
Security and platformIs the tool approved for this data and use? Are access, secrets, logging, vendor, and incident controls adequate?Approved platform and control record; AI Security or Cyber Security decision where triggered.
Legal and clientDo contracts, licences, regulation, or account commitments permit the proposed processing and output?Legal, account, or client approval and any conditions.
Automation and impactCould the output affect people, clients, safety, finance, compliance, or write back to a system?Human decision point, audit trail, testing threshold, override, and accountable decision owner.
User accountabilityCan users understand limits, verify output, and report a suspected exposure or unsafe result?User guidance, prohibited uses, validation steps, and incident route.

Six delivery gates

A request only moves forward when the evidence for its current gate is sufficient. “Stop if” conditions are pause points for resolution, not reasons to hide or bypass the issue.
1

Frame the use case

Start with the user need, not the technology. Decide whether the work is self-service, Microsoft Copilot, Copilot Studio, Power BI, Power Platform, a custom mini app, a reusable micro-tool, or a full application.

EvidenceBusiness problem, users, expected output, route recommendation, business owner, support owner.
Stop ifNo accountable owner can accept the operational, legal, security, and quality consequences.
2

Classify data and obligations

Identify whether the workflow uses public, internal, client confidential, personal, financial, regulated, commercially sensitive, or restricted data. Check client contract terms and residency needs before any sensitive data is used.

EvidenceData classification, access approver, client restrictions, privacy or DPIA trigger, residency or cross-border position.
Stop ifThe data cannot be lawfully, contractually, or securely used in the proposed AI workflow.
3

Prove data readiness

For data-heavy requests, the Requirements Agent should not treat the work as build-ready until the source landscape and ongoing ownership are clear.

EvidenceSource inventory, source of truth, data owner, access approver, quality issues, update cadence, pipeline or connector, support owner.
Stop ifThe source of truth, ownership, permissions, quality, refresh path, or ongoing support is materially unclear.
4

Choose platform and controls

Use the lightest approved route that fits the workflow, but do not bypass controls for speed. Sensitive data needs an approved or Trusted platform, least-privilege access, logging, retention clarity, and security review.

EvidenceApproved platform, SSO, RBAC, audit logging, secrets handling, retention, vendor review, architecture or security input.
Stop ifThe design relies on personal AI accounts, public tools, unapproved vendors, or unresolved security and residency controls.
5

Test before reliance

Define what good output looks like before pilot. Test accuracy, edge cases, bias, sensitive information exposure, prompt injection, over-reliance, and any automated action or system write-back.

EvidenceTest cases, pass criteria, failure modes, human review point, prohibited outputs, validation owner, pilot findings.
Stop ifUsers would rely on untested output, unchecked automated actions, or unclear confidence limits.
6

Operate, monitor, and retire

Governance does not end at launch. The owner must monitor use, cost, incidents, access, model or data changes, accuracy drift, support demand, and whether the tool should be changed or retired.

EvidenceIncident route, audit logs, usage and cost monitoring, support process, change review, review date, sunset criteria.
Stop ifThere is no way to detect failure, investigate issues, manage cost, maintain ownership, or retire unsupported use.

Data ownership and data-layer decisions

Do not start by asking which app to build. First decide what governed data capability is missing, who owns it, whether the proposed use is permitted, and whether a reusable data layer should exist before an app, agent, dashboard, or workflow is built.

Three questions for every data-heavy request

Outcome

What depends on the data?

Name the decision, report, workflow, agent response, system action, or client outcome that would rely on it.

Sources

Where does it live?

List systems, APIs, files, SharePoint sites, dashboards, warehouses, exports, and manual steps.

Authority

Who can approve use?

Name each source owner, access approver, quality owner, and the business owner accepting the output.

Ownership patterns

PatternWhat it meansGovernance move
Single accountable ownerOne owner controls the source of truth, access, definitions, quality, and change decisions.Confirm permission, quality limits, refresh, retention, and support before build or AI use.
Multiple domain ownersThe workflow combines functions, regions, systems, projects, or dashboards.Align definitions, permitted use, quality gaps, and who accepts the combined-output risk.
No clear ownerThe request depends on inherited spreadsheets, unmanaged exports, unofficial dashboards, or tribal knowledge.Do not treat it as build-ready. Establish ownership or a governed dataset first.
Client or third partyThe data is client-provided, licensed, regulated, externally sourced, or contractually constrained.Review contract, residency, permitted use, AI restrictions, retention, and security before processing.

Choose the right data-layer response

MoveUse whenMust be clear first
Use the existing sourceOne owner, one source of truth, low sensitivity and reuse, known quality, simple workflow.Owner approval, access, quality caveats, retention, and human review.
Create a governed dataset or viewRecurring reporting, repeatable extracts, stable sources, clear owner, inconsistent manual handling.Fields, definitions, refresh, quality checks, access group, support, and change process.
Create a reusable data product or layerMultiple consumers or owners, conflicting definitions, sensitive data, AI retrieval, apps and dashboards sharing data.Ownership, lineage, controls, service expectations, monitoring, and architecture ownership.
Route to architecture or platform reviewSystem write-back, master data, regulated workflows, audit-heavy operation, cross-border constraints, or long-lived dependency.Architecture, security, legal, data platform, and operational governance decisions.

Data-readiness is a gate

Low readinessDiscovery can continue, but the work is not build-ready. Ask the requester for sources, source of truth, owner, access, quality, refresh, pipeline, and support details.
Medium readinessScope with caveats only. Resolve named gaps before sizing delivery effort, connecting systems, piloting, or using sensitive data.
High readinessContinue route validation and risk review. High readiness is not approval; it means the data evidence is strong enough for the next governance step.

Stop delivery while any of these remain unresolved

Accountability and evidence

Governance needs named decisions, not just consultation. Record one accountable owner for the outcome and make each specialist decision visible.

RoleAccountability to record
Business ownerOutcome, users, value, acceptable business risk, funding where needed, and the decision to continue or stop.
App or product ownerDesign, user guidance, testing, dependencies, support, changes, monitoring, and retirement.
Data ownerSource of truth, definitions, quality limits, access, permitted use, retention, and change approval.
Platform teamMarketplace hosting, routing, identity integration, deployment controls, platform monitoring, and platform standards.
Specialist reviewersAI Security, Cyber Security, Privacy, Legal, Architecture, Data Platform, or account leadership decisions when triggered by the use case.
UsersFollow permitted use, protect data, verify outputs, keep human judgement, and report failures or suspected exposure.

Keep one evidence pack

Ownership record

Business owner, support owner, decision owner, intended users, expected outcome, and escalation route.

Data record

Data classification, source inventory, source of truth, access approver, quality limits, update cadence, and retention need.

Security record

Approved platform, access model, logging, secrets handling, residency, vendor position, and required AI Security review.

Evaluation record

Test cases, accuracy threshold, prohibited uses, human review point, failure modes, bias checks, and pilot results.

Operations record

Monitoring, cost review, incident path, support process, change review, review date, and sunset criteria.

Approval record

Risk level, required reviewers, decision, open conditions, and the date the decision needs to be revisited.

Operate, review, and retire

WhenMinimum controlTrigger for action
Before launchNamed owner, user guidance, passed tests, approvals, support and incident route, cost visibility, review date, retirement plan.Any missing owner, approval, critical test, or safe rollback prevents launch.
In operationMonitor availability, usage, cost, access, errors, output quality, incidents, and user feedback in proportion to risk.Unexpected use, material drift, rising cost, control failure, or recurring support demand prompts review.
On changeReassess material changes to model, prompt, data, vendor, connector, user group, geography, output, or automated action.A change that alters risk or obligations returns to the relevant delivery gate.
On incidentContain, preserve evidence, notify the correct owner and security route, assess impact, correct, and record lessons.Suspected exposure, unsafe output, unauthorised access, or harmful action is escalated immediately.
At review or retirementConfirm the use case still has value, ownership, acceptable quality, valid approvals, affordable operation, and a supported platform.Remove access and integrations, handle retained data, archive the decision record, and communicate retirement.

Review and escalation templates

Add the request context once, then open or copy the relevant email. Replace placeholders and select the actual accountable recipient; a template does not constitute approval.

Request context

These fields may be prefilled with ?request_id=..., &title=..., &requester=..., and &route=....

Data triage

Request a data-readiness review

Use when ownership, sources, quality, access, refresh, pipeline, or support is incomplete.

Source owner

Confirm source of truth

Use when an apparent owner needs to confirm authority, access, quality, refresh, and permitted use.

Multi-owner

Align several owners

Use when several teams, regions, systems, or dashboards need one agreed data position.

Architecture

Request a data-layer decision

Use when reusable datasets, shared definitions, integrations, AI retrieval, or platform ownership may be needed.

AI Security

Request AI Security review

Use for sensitive, client, personal, regulated, external-facing, automated-action, or high-impact AI.

Client / legal

Confirm permitted use

Use for client-provided, contractually constrained, licensed, regionally restricted, or regulated data.

Policy roadmap

Proposals, not current rules. These items need an owner, decision, and implementation plan before they can be treated as policy.
TopicDecision still needed
App reviewWhether all new apps require a recorded review of security, code quality, accessibility, dependencies, and ownership before launch.
Data classificationWhich additional controls apply to each class of data and which classes are prohibited from the Marketplace.
Usage and valueWhich adoption, outcome, quality, and user measures determine continued investment or retirement.
Cost governanceWhen per-app budgets, alerts, chargeback, or approval thresholds are required.
LifecycleReview frequency and formal criteria for unsupported, unmaintained, unused, or superseded tools.
Incident responseA published runbook with severity, contacts, containment, evidence, notification, recovery, and learning steps.
Access controlWhen an app requires a restricted group, privileged access, periodic recertification, or segregation of duties.
Compliance mappingHow applicable legal, regulatory, contractual, AI, privacy, records, and sector obligations map to controls and evidence.

Reference frameworks

These references inform the control model; ERM policy, client obligations, and decisions from accountable governance teams take precedence.