Identity technology company iProov has launched a service intended to verify that a genuine, authorised person has approved a specific action requested by an artificial intelligence agent. The service is designed for enterprise workflows in which automated systems can initiate consequential actions but companies still require traceable human approval.
London-based iProov says its hosted service implements the Human Approval and Presence Specification, or HAPS, an open technical approach released in September 2026. Its initial integrations connect to agentgateway, developed by solo.io, and Cerbos, a platform used to manage authorisation policies. The arrangement is designed to produce proof that an identified individual approved the exact transaction an AI system is seeking to perform.
Under the described workflow, an AI agent requests an action. If a Cerbos policy requires human approval, that action is paused while the named approver is shown its details and scope. The person then completes an iProov presence check. A signed, single-use credential is produced and linked to the specific action before the agent is permitted to proceed.
Binding approval to the transaction details is a critical part of the design. If an agent were to change a payment amount or destination after approval, the credential would no longer match the request and the action would be blocked. The aim is to prevent a general confirmation from being reused to authorise an altered or unrelated instruction.
Agentic AI differs from a conventional chatbot because the software can connect to tools and take steps on behalf of a user. That might involve changing customer records, approving an operational workflow, requesting information from another system or initiating a payment. Many such actions are relatively low risk, but others require a stronger record of who permitted them and on what basis.
Enterprise security teams commonly separate authentication, authorisation and audit. Authentication addresses identity; authorisation defines what a person or system can do; and audit creates a record of what happened. An automated agent complicates those boundaries because it may operate under delegated permissions, interpret instructions dynamically and interact with multiple applications before completing a task.
The HAPS approach attempts to join the three functions at the point where an important action needs specific human approval. Cerbos evaluates the rules, iProov supplies a human-presence check, and the resulting record can be retained for subsequent review. Businesses would still need to define which actions require approval, who is eligible to give it and how exceptions are handled.
Andrew Bud, iProov’s founder and chief executive, said increasingly consequential tasks must remain subject to human control and accountability. The company describes its capability as proof of genuine human approval, rather than simply a message indicating that an approval button has been clicked. The mechanism is more specific than a routine two-factor authentication prompt, but its effectiveness in different environments will depend on implementation.
The specification is open and supported by publicly available technical material, including a partial reference implementation and test vectors. Openness may allow other gateways and security tools to implement compatible approval checks, although the availability of a technical specification does not itself establish industry-wide acceptance or interoperability at production scale.
Governance policies will need to reflect the level of consequence. An AI agent summarising a meeting does not require the same approval process as one initiating a supplier payment or changing bank details. Requiring biometric checks too frequently could encourage workarounds, while applying them only at the final stage may miss earlier risky activity. Organisations will need to define transaction thresholds, escalation rights and acceptable fallback processes in advance.
Important questions remain around accessibility, false rejections, privacy and the resilience of identity checks when devices or networks are unavailable. Companies handling sensitive transactions will also need to consider how long evidence is retained and which parties can inspect it. The launch supplies a specified human-approval mechanism, although deployment figures, interoperability across products and independently measured performance have not yet been established.
You must be logged in to post a comment.