The prediction: In 2027, buyers of consequential AI-enabled services will increasingly require current, deployment-relevant evidence before activating or expanding their use.

An AI supplier releases a new capability. Development is complete, the feature is available, and customers can enable it.

Yet some customers may hold back.

The question is whether the assurance previously provided still covers what the service can now do. A completed security review or an earlier assessment may establish confidence in the original deployment without answering questions about new functionality, data connections, or operating conditions.

We expect these questions to become more prominent in purchasing, activation, and expansion decisions during 2027.

That changes the commercial significance of AI governance. A supplier may be technically ready to deliver a capability but unable to demonstrate that its controls remain appropriate for the deployment a customer is considering.

OneTrust research highlights the evidence challenge. Only 28% of respondents report governance evidence and audit trails among their current governance activities. Meanwhile, 38% prioritize preparing for regulations, standards, or audits over the next 12 months.

The report also examines emerging customer expectations around traceability, monitoring, and intervention. The direction is toward assurance that helps resolve a specific business decision, rather than relying exclusively on evidence collected during an earlier review.

What changes for businesses?

For AI suppliers, the ability to provide relevant assurance may increasingly influence customer acceptance and the pace of expansion. Commercial, product, security, and governance teams will need to establish what evidence applies to each capability and who can resolve outstanding concerns.

For buyers, assurance becomes part of deciding whether to activate a service, limit its use, or require additional controls.

Financial and business leaders may also need to consider whether the value delivered by an AI capability justifies its operating costs and exposure before expanding investment.

What to look for in 2027

• Customers requesting updated assurance before enabling new AI capabilities.
• Contracts requiring reassessment or new evidence following material changes.
• Greater scrutiny of AI-related exposure in insurance and commercial decisions.

The prediction: In 2027, leading high-consequence AI agent programs will add task-specific authority, tested intervention, and independently protected records to deployment criteria.

An agent is tasked with reducing cloud spending while preserving order-processing deadlines. It scales down unused capacity and postpones batch jobs, with both actions falling within its technical permissions. The operating bill improves, but overnight order processing misses its deadline. Although each action was individually permitted, their combined effect violated a condition of the task the business had authorized.

This illustrates a growing problem in AI agent governance: access to perform individual actions does not establish authority to produce every outcome those actions make possible.

As organizations deploy more connected and autonomous workflows, they will need to define what agents are permitted to accomplish, which trade-offs are unacceptable, and when human intervention is required.

That authority must remain meaningful after deployment.

An agent that gains another connection or responsibility may be able to affect business outcomes that were never considered in its original approval.

In OneTrust research, 48% of respondents report at least one incident involving unapproved actions by AI systems or agents during the past year.

Only 28% report governance evidence and audit trails.

These findings make action authority and incident reconstruction current operating concerns.

The next challenge is establishing whether an agent operated within its approved purpose.

A record of technically permitted actions may help explain what happened, but evaluating the overall outcome requires comparing those actions and their consequences with the task's authorized limits.

What changes for businesses?

Organizations expanding high-consequence agent workflows will need stronger deployment criteria.

Business owners must define unacceptable outcomes, while security and engineering teams test whether restrictions, escalation, and revocation work where actions take effect.

Audit and governance teams will need records that the acting agent cannot independently alter, enabling them to reconstruct significant decisions and investigate failures across connected systems.

If those conditions cannot be demonstrated, businesses may need to limit an agent's authority or retain human checkpoints.

What to look for in 2027

• Agent deployment reviews that explicitly define business purpose and task boundaries.
• Tests showing that unauthorized actions can be stopped or escalated.
• Independent records that support investigation across systems.
• Early evaluations of harmful outcomes arising from individually permitted actions.

The prediction: In 2027, leading enterprises will begin treating material changes in purchased, built, and embedded AI as reasons to reassess existing approvals, even when no new supplier is onboarded.

An enterprise may operate a model internally, purchase access to a hosted AI service, and use agents embedded within existing business applications.

However, those conditions may change without any further purchasing decision. A new connector could give an assistant access to additional customer information, a configuration change could broaden its permissions, or a model update could affect the reliability of a control previously used to approve its deployment.

The important question, therefore, is whether the evidence supporting the original decision remains valid as the system evolves.

OneTrust research highlights the challenges organizations already face in maintaining visibility into third-party AI.

Some 39% identify limited visibility into third-party AI components, vendors, or data sources as a factor delaying AI initiatives. Separately, 28% report that vendor AI tools were adopted before a full assessment had taken place.

These findings support a broader need for oversight that continues after initial approval.

In 2027, we expect leading organizations to define material changes as reassessment events. That means identifying when a change affects the intended use, data access, connected systems, or effectiveness of existing controls.

It doesn't require a complete review after every software update. The response should be proportionate to the changed exposure, the capability's business importance, and the evidence that remains valid.

What changes for businesses?

Approval becomes an ongoing responsibility across the AI estate.

Technology and data teams will need visibility into models, applications, connectors, and consequential information flows. Procurement and legal teams may need clearer contractual notification and intervention rights. Business owners must understand when they can continue using a changed capability and when its use should be restricted.

For critical workflows, these decisions may also require fallback arrangements that preserve business continuity.

The goal is to make reassessment responsive to meaningful changes without creating an unnecessary review burden.

What to look for in 2027

• AI reassessments triggered without a new procurement event.
• Change notifications connected to affected applications and business uses.
• Documented decisions to continue, restrict, remediate, or replace changed capabilities.
• Greater attention to fallback arrangements for critical AI-dependent workflows.

The prediction: In 2027, early adopters of AI-mediated customer experiences will begin connecting explicit data-use choices to the processing those choices permit, making refusal and withdrawal more effective within individual services.

A shopper asks an AI assistant to reserve a coat at a nearby store.

During the conversation, they explain that it is a gift and describe the recipient's preferences.

That information helps the assistant complete the immediate task.

It may also reveal details that a conventional reservation form would never have requested.

What happens if the service later wants to use those details for advertising?

The original conversation doesn't automatically establish permission for a different purpose.

Where an additional choice is required, the organization needs to explain the proposed use and ensure that the customer's response changes what happens next.

AI conversations make this challenge more visible because the information collected can be broad, unstructured, and context-dependent.

Information provided to complete one task may become available for uses beyond those the customer expected.

OneTrust research shows the wider data-governance challenge.

Fifty-two percent of respondents identify data quality, access, privacy, or security concerns as factors delaying AI initiatives.

Thirty-nine percent prioritize better visibility into risk and data use.

These findings don't establish widespread adoption of conversational permission mechanisms, but they reinforce the importance of understanding and governing how information moves through AI workflows.

We expect initial developments in 2027 to concentrate within individual services, where organizations can connect a customer's choice to the subsequent processing it controls.

Reliable enforcement across different agents and providers is a less mature possibility.

What changes for businesses?

Product, marketing, privacy, and engineering teams may need to reconsider how data-use choices are presented during AI-mediated interactions.

They will need to distinguish the information necessary to fulfill a request from additional processing that requires a separate decision.

Where applicable, systems should also demonstrate that refusal or withdrawal changes downstream use.

If a customer declines optional advertising, the service should honor that choice while completing the original task where possible.

Customer preferences, technical permissions, and legal consent must remain distinct.

An instruction to an assistant doesn't automatically constitute legal consent, and recorded permission alone doesn't establish that processing is lawful.

What to look for in 2027

• AI experiences requesting data-use choices at the point of a proposed additional use.
• Records connecting customer choices to their purpose and intended recipient.
• Evidence that refusing or withdrawing permission changes subsequent processing.
• Early pilots addressing choices when AI workflows move between services.