Enterprise GenAI adoption is stalled by governance gaps

At a glance

What this is: Knostic argues that enterprise GenAI adoption is being held back less by model capability than by governance gaps, trust deficits, and regulatory complexity.

Why it matters: IAM, security, and governance teams need to treat GenAI access, output controls, and monitoring as part of identity governance, not as a separate AI pilot concern.

By the numbers:

• Only 7% of enterprises had fully embedded governance programmes with continuous KPI monitoring by mid-2025.
• Enterprise adoption rates jumped from 55% in 2023 to 75% in 2024.

Context

Enterprise GenAI adoption is the point at which generative AI moves from experimentation into business operations that affect decisions, speed, and value creation. The central governance gap is that many organisations are trying to scale AI before they have the controls to understand who can use it, what it can reveal, and how to monitor it safely.

The article frames adoption as a governance problem, not a model-quality problem. It points to sandboxes, AI steering committees, continuous monitoring, and benchmarking as the practical structures that make AI adoption governable across compliance, DevOps, and business teams.

Key questions

Q: How should enterprises govern GenAI adoption before moving from pilots to production?

A: Enterprises should treat GenAI governance as an operating model, not a policy add-on. That means defining ownership across security, compliance, DevOps, and the business, then using sandboxes, access controls, and continuous monitoring to validate use cases before production exposure.

Q: Why do AI programmes become fragile when governance is added too late?

A: Because AI systems inherit whatever data, context and process gaps exist at launch, and those gaps get amplified as the use case spreads. Late governance usually means retrospective fixes, inconsistent controls and unclear responsibility. The longer the delay, the harder it becomes to standardise or prove trust.

Q: What are the signs that AI governance controls are not keeping pace with adoption?

A: Common warning signs include unclear ownership for AI use cases, inconsistent approval processes, limited visibility into where sensitive data enters models, and weak evidence for audits or assessments. Teams also struggle when privacy, security, and legal review happen late or manually, because that usually means governance is reactive rather than embedded in the AI delivery process.

Q: Should organisations use internal sandboxes or direct production testing for GenAI?

A: Organisations should use internal sandboxes first. Direct production testing increases exposure before teams understand leakage, role boundaries, and compliance impact, while sandboxes with synthetic or masked data let practitioners validate use cases and governance controls with far less risk.

Technical breakdown

Why GenAI adoption stalls without governance

Generative AI changes the control surface because output risk is not limited to explicit data access. Prompts, retrieved context, model responses, and inferred knowledge can all create exposure even when the underlying source data remains protected. That means traditional approval-only governance breaks down once the system is asked to answer in natural language across business contexts. The article's focus on trust, regulation, and oversight reflects a broader reality: adoption fails when organisations cannot explain what the model can see, what it can infer, and who is accountable for that behaviour. Practical implication: governance must cover prompt flow, response visibility, and access scope together.

Practical implication: govern prompts, outputs, and inference as one control surface, not as separate tooling concerns.

What internal sandboxes change for enterprise AI testing

A regulatory-style sandbox is a controlled environment where teams validate GenAI use cases without exposing production data, real customer records, or regulated workflows. In enterprise terms, this usually means synthetic or masked data, limited personas, defined test boundaries, and traceable experiments before broad rollout. The value is not just privacy protection. It is the ability to prove whether the use case is viable, where it leaks, and which access patterns are unsafe before business users depend on it. Practical implication: internal sandboxes should be treated as a governance gate, not as a temporary lab for experimentation.

Practical implication: use sandboxes to validate access patterns and leakage risks before any production-facing GenAI rollout.

Why persona-based access controls matter for AI outputs

Persona-based access control is an output governance model that tailors what an AI system can reveal based on the user's role, task, and context. That is different from standard application access, because the risk is not only whether a user can reach data directly, but whether the model can surface sensitive knowledge indirectly through synthesis and inference. This becomes especially important in copilots and search-style experiences where users may never query the source system itself. Practical implication: organisations need controls that evaluate context at response time, not only at login or data retrieval.

Practical implication: apply response-time access decisions so AI output is constrained by task, role, and context.

Threat narrative

Attacker objective: The objective is to obtain sensitive knowledge or operational advantage through AI-mediated oversharing rather than through direct system compromise.

1. Entry occurs through broadly enabled GenAI use cases where users can query copilots, search tools, or internal assistants without sufficient governance guardrails.
2. Credential or entitlement abuse is less the issue than overbroad context access, which allows the model to retrieve or infer knowledge beyond intended user scope.
3. Impact appears as oversharing, hallucinated confidence, compliance exposure, and loss of trust when AI outputs reveal sensitive business information.

Breaches seen in the wild

• McKinsey AI platform breach: McKinsey AI platform hack exposed 46M chats and sensitive data.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Governance, not model quality, is now the binding constraint on GenAI adoption. The article is right to centre trust and control rather than raw capability. Enterprises do not fail because they lack access to GenAI features; they fail because they cannot make those features safe enough to approve for normal business use. That shifts the discussion from innovation theatre to identity, access, and monitoring discipline.

Only 7% embedded governance is not a maturity gap, it is an operating model gap. A governance programme that is not embedded cannot reliably govern prompt logging, output review, or context-aware access decisions. The problem is not a missing policy statement but a missing control plane that connects compliance, DevOps, and business ownership. Practitioners should read that 7% figure as evidence that most organisations are still trying to scale AI on procedural promises.

Internal sandboxes are the right pattern because they separate adoption speed from production exposure. Controlled test environments let teams validate use cases, surface leakage paths, and test persona boundaries before full deployment. That matters because AI risk emerges in the interaction between data, context, and output, not in the model alone. The practical lesson is to treat every pilot as a governance proof point, not just a technical demo.

Persona-based access is becoming the AI analogue of need-to-know governance. GenAI systems can reveal more than the user explicitly asks for, so controlling only source-data access is insufficient. The governance model has to include what the model may infer, summarise, or disclose based on role and task. That is a change in identity thinking, because the risk now sits in the response layer as much as in the entitlement layer.

Continuous monitoring is the only scalable answer once AI moves into production. Static approvals cannot keep up with prompt drift, workflow drift, or changing business context. The article's emphasis on KPI tracking, traceability, and dynamic evaluation reflects the right direction for enterprise governance. Teams should assume that GenAI adoption will accelerate faster than their review cycle unless they instrument the control plane accordingly.

From our research library:

• Only 13% of organisations feel extremely prepared for the reality of agentic AI despite the majority racing toward autonomous adoption, according to the 2026 Infrastructure Identity Survey.
• Read next: Agentic AI Identity Risk Board Briefing

What this signals

GenAI adoption now depends on whether governance can move at the speed of experimentation. Static review cycles are too slow for prompt drift, output drift, and changing business context, so organisations need controls that evaluate the response layer as part of the identity layer. The practical shift is from one-time approval to continuous governance.

Only 7% of enterprises had fully embedded governance programmes with continuous KPI monitoring by mid-2025, according to the 2026 Infrastructure Identity Survey. That gap shows why many organisations can launch pilots but cannot safely convert them into routine business capability. The issue is not model availability, it is control maturity.

Persona-bound disclosure becomes the decisive boundary once copilots and search tools enter production. If access controls stop at the source system, the AI layer can still surface sensitive knowledge through inference and summarisation. Teams should assume the model is now part of the access path, not just a consumer of it.

For practitioners

• Define an AI governance operating model Establish joint ownership across security, compliance, DevOps, and business leads so GenAI approvals, monitoring, and escalation paths are explicit before rollout.
• Build internal sandboxes for GenAI testing Use synthetic or masked data, limited personas, and traceable experiments to validate use cases before they touch production workflows.
• Instrument continuous monitoring for AI workflows Track prompt safety, oversharing alerts, hallucination frequency, and use-case drift so governance can detect problems after deployment, not only at approval time.
• Adopt persona-based access control for outputs Restrict what the model can disclose by user role, task context, and business scope, especially in copilots and search-style interfaces.
• Use net assessments to benchmark adoption Compare pilot-to-production conversion, usage rates, and control effectiveness across business units so leaders can see where governance is slowing or enabling scale.

Key takeaways

• Enterprise GenAI adoption is constrained by governance maturity more than by model capability or user enthusiasm.
• The biggest operational risk is not direct system compromise but oversharing, inference, and weak control over AI outputs.
• Internal sandboxes, continuous monitoring, and persona-based access decisions are the controls that make scaled adoption governable.

Key terms

• GenAI Governance: The set of policies, controls, and accountability rules that determine how generative AI is approved, used, monitored, and retired. Effective governance connects purpose, data classification, access boundaries, logging, and ownership so AI programs do not scale faster than control maturity.
• Internal AI Sandbox: An internal AI sandbox is a controlled testing environment where teams validate generative AI use cases before production exposure. It typically uses masked or synthetic data, limited personas, and traceable experiments so organisations can study leakage, compliance impact, and workflow fit without broad business risk.
• Persona-Based Access Control: Persona-based access control groups users by job function and expected responsibilities rather than by broad application entitlement. In AI governance, it helps limit which copilots, agents, or embedded features can access sensitive data for a given role.
• Continuous AI Governance Monitoring: An ongoing process for observing how AI systems behave after deployment and using that evidence to guide accountable action. It combines behavior, identity and access, threat context, and agent activity so teams can detect meaningful change, preserve human review, and keep the system operating within approved purpose and risk boundaries.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity lifecycle are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are responsible for identity security strategy or NHI governance in your organisation, it is worth exploring.