Summary

• A NiCE survey collected 118 responses from practitioners at more than 70 customer organisations across several markets.
• Approximately six in ten respondents described multi-agent environments, including systems coordinating service requests and administrative tasks.
• Respondents identified identity verification, human oversight, and auditable records as important controls, but the selected sample does not represent the wider industry.

Customer service departments are beginning to handle requests initiated by consumers’ artificial intelligence assistants, creating new requirements for businesses to establish who has authorised a transaction when software communicates directly with another automated system.

A survey conducted by NiCE among 118 customer service practitioners at more than 70 organisations found that approximately six in ten respondents were using environments containing multiple cooperating AI agents. Participants also described personal assistants contacting companies to arrange appointments and coordinate routine service requests.

The research was published on 29 October and drew on responses from members of the NiCE User Group, collected between 22 September and 10 October. Organisations represented included businesses in the UK, Europe, North America, India, and Australia.

Because the participants were already NiCE customers and the survey used a single open-ended question whose responses were coded by theme, its findings cannot establish how widespread agent-to-agent customer service has become across the wider economy.

Customers delegate contact to automated assistants

Traditional customer service automation generally begins with a person contacting a business through a website, telephone line, or messaging application. A chatbot or voice assistant may answer common questions before passing more complicated requests to an employee.

Agent-to-agent interactions introduce a different arrangement. The party initiating a request may be an automated assistant acting on instructions that its user supplied earlier, while the receiving organisation may also use software to interpret the request and complete the associated work.

A personal assistant might contact a business to rearrange a service appointment, request information about an existing account, or confirm delivery details. The company’s agent must then determine which information it can disclose and whether the external assistant has authority to make the requested change.

NiCE’s respondents described both external interactions and cooperation between AI systems within their own organisations. One agent might interpret a customer’s intention, another retrieve account information, and a third initiate an administrative action or route the request to an appropriate service.

Such arrangements may reduce repetitive work, particularly where the underlying task involves retrieving structured information and updating established business systems.

However, dividing a process between several automated components also creates more opportunities for information to be misunderstood or an action performed without sufficient authority.

Businesses need to know which system initiated a request, what information it received, which permissions applied, and whether a person approved any consequential decision.

Operational examples remain self-reported

The survey includes examples from transportation, healthcare, energy, and building technology organisations, although most participating companies are not named in the published material.

An engineer at a global transportation business described personal AI assistants contacting its virtual agents to negotiate appointments and clarify account details without a person actively participating in either side of the exchange.

The respondent reported reduced handling times for routine queries and described identity checks and interaction logs as part of the controls used to manage those conversations.

At a UK energy utility, a head of customer experience reported that the proportion of interactions handled through end-to-end automation had increased from zero to almost 35%.

That figure was provided by an unnamed respondent and was not accompanied by an independently verified performance assessment or a standard definition of successful resolution.

An interaction completed without escalation to a human employee does not necessarily establish that the customer’s underlying problem was resolved correctly. Systems can complete an automated process while still producing an inaccurate answer or inappropriate outcome.

Another UK respondent, working for an urgent and integrated care provider, described the potential for separate AI agents to coordinate entire administrative tasks rather than merely providing information to staff.

Healthcare environments introduce additional responsibilities because service requests may involve sensitive personal information, changes to appointments, and decisions whose consequences differ from those of a routine commercial enquiry.

These examples indicate how selected organisations are experimenting with the technology, but they provide insufficient evidence to calculate typical cost savings, resolution improvements, or adoption rates across customer service operations.

Identity becomes a separate technical problem

When a person contacts a business, the organisation may use passwords, security questions, identity documents, or account-specific information to establish that the request is legitimate.

An AI assistant acting on that person’s behalf complicates the process because the software needs a way to demonstrate both its identity and the authority delegated to it.

Permission to retrieve an appointment time does not necessarily include permission to cancel the appointment. Similarly, an assistant authorised to check an order’s delivery status should not automatically gain access to payment details or unrelated customer records.

Organisations will therefore need methods for expressing delegated authority, verifying permissions, and determining when additional confirmation from the customer is required.

NiCE identifies Agent2Agent, Model Context Protocol, and other emerging integration mechanisms as relevant to its customer service products.

Agent2Agent supports communication and task coordination between agents, while Model Context Protocol provides a structured mechanism for connecting AI applications with tools and information sources. These technologies address different aspects of interoperability.

Neither protocol independently establishes that an automated system has lawful authority to disclose personal information or complete a financial transaction. That responsibility remains with the applications and organisations implementing the controls.

NiCE is also collaborating with Meta and other organisations on the proposed Personal Agent Protocol, intended to support secure communication between consumer assistants and enterprise agents.

The development of common protocols could reduce the need for individual businesses to build separate integrations for every assistant provider. However, the benefits depend on adoption, compatible implementations, and agreement about how identity and authorisation are handled.

Oversight remains part of automation

More than half the survey respondents identified measures such as oversight, auditability, human intervention, or unified interaction logs as factors supporting wider use of automated agents.

These controls can help organisations establish a record of what happened when a service request crosses several systems, particularly where an external assistant triggers an action inside an enterprise application.

Sascha Poggemann, chief operating officer of NiCE Cognigy, said: “The organizations scaling fastest are the ones that can see and govern every agent interaction.”

Monitoring provides useful evidence for investigating failures, but an audit record cannot guarantee that an original instruction was authorised or that the resulting business decision was correct.

Customers also need a dependable process for reversing mistakes. If an assistant changes an appointment incorrectly, discloses information to the wrong party, or makes a request beyond its authority, the affected person must be able to establish what occurred and obtain a correction.

Where automated interactions involve financial commitments or sensitive records, those arrangements become more important because the consequences extend beyond an inconvenient customer service experience.

NiCE’s survey provides examples of early adoption among businesses already engaged with its technology. Wider evidence on completed transactions, error rates, customer satisfaction, and security incidents will be needed to establish whether agent-to-agent operations consistently improve service delivery.

The reported use cases nevertheless demonstrate that some customer service systems are being adapted to accept requests from software acting on behalf of people. The practical requirements now extend beyond answering questions to verifying authority, controlling actions, and retaining reliable records when several automated systems participate in the same transaction.