Reading a note or hearing a recording saying “You’re chatting with our AI assistant…” is an increasingly common part of customer service interactions. This disclosure is becoming normalized and it might read as clarity to some. But just the existence of disclosure isn’t enough to ensure a customer truly understands the role of the AI in the interaction.
For example, the example above seems straightforward, but several scenarios are possible. The AI-powered chatbot could be fully autonomous, or a human agent might be receiving and sharing AI-generated responses. The agent could be a voice or text agent. A human agent could be handing tasks to an AI agent, or AI might be handling an interaction before transferring it to a human.
As artificial intelligence continues to move from being confined to an identifiable chatbot towards being embedded throughout customer interactions, enterprises should consider their disclosure policies and practices as part of their customer-experience design. For the disclosure of an AI’s role in an interaction to be not just technically compliant but meaningful, what should a customer actually be told?
Disclosure isn’t a binary question
In this CX environment, binary questions about disclosure of AI use aren’t sufficient any longer. Rather than asking “Did we disclose AI?” and providing a yes-or-no answer, the better practice is to think about disclosure related to levels of AI involvement.
These levels will look different depending on where and how AI is used in customer service workflows. For example, the involvement could be AI-facing, where a customer is directly interacting with an AI agent. Or it could be AI-mediated, where a customer believes they are interacting with a person but the responses they’re receiving are generated or shaped by AI. The involvement may be AI-assisted, where a human agent holds the responsibility for the interaction, but AI plays a role – for example, AI is used in information retrieval, summary generation, language translation, etc. It’s also possible a hybrid model, where the responsibility moves between human and AI agents, is used for a single interaction.
In any of these potential levels of involvement, the important question to answer for disclosure is “At what point does the involvement of AI in this interaction become material enough that the customer should be informed?”
This goes beyond simple disclosure of AI involvement to consider the customer’s understanding of that involvement and its impact.
Good disclosure practices depend on the channel
In addition to considering the levels of AI involvement, good disclosure practices should also adjust for the channel of communication between the AI and the customer.
For example, if the communication is happening via chat, communicating the use of AI in the interaction via a persistent visual indicator, such as a banner, is simple to implement. However, there are still important questions to consider around disclosure. How prominent should the visual indicator be, for example? Should it remain visible after the customer is handed off to a human agent and does that depend on if the human may use AI as part of their workflow? Is the visual indicator accessible to all users, including those who are visually impaired and may be using assistive technology?
For voice communication between AI and customers, disclosure of the use of AI should be provided verbally. However, this raises important questions about the timing of the disclosure, its length and whether or not it was heard and understood by the customer. For example, a long script provided for the customer’s information and to ensure compliance can provide necessary information but might also make for a worse customer experience.
When the interaction isn’t live but is happening over messaging or email, clarity around disclosure is even more difficult. For example, should it be provided just at the beginning of the interaction or in every new message? If the response contains a mix of AI-generated and human-generated content, should the labelling get more granular and identify which parts are which? If the responses are human-generated with AI assistance, such as translation, do customers need to know about that?
One common mistake is treating disclosure as a compliance statement that can be replicated across all customer-service channels, says Srinivas Chippagiri, a senior technical staff member at Salesforce. For chat, a persistent visual indicator can work, but voice disclosure should be spoken early, clearly and slowly enough to ensure comprehension, Chippagiri advises. And in asynchronous messaging, one interaction could last several hours and move between humans and AI multiple times, which makes it even more important to signal when the agent’s identity changes.
Even when an AI disclosure meets technical requirements, customers can still be inadequately informed. Chippagiri points to examples like AI disclosures buried in terms of service, AI labels that lack explanation and ambiguous language about whether responses are generated or assisted. “The requirement is met; the person is no better informed,” he said.
For any of these channels of communication, equivalent transparency doesn’t necessarily mean that disclosure practices should be identical. In a multichannel customer-service center, disclosure strategies should work across the channels in use while considering the unique challenges and circumstances of each.
After ensuring customers are adequately aware of the use of AI in a customer-service interaction, it’s important to consider what, if anything, the customer can then do with that information. This ensures your disclosure moves beyond simple compliance into providing a truly positive customer experience.
The relationship between trust and disclosure isn’t necessarily straightforward, according to Dr. Eva Wolf, a researcher and consultant with Big Plans Media. In her study of 394 B2B decision-makers, Wolf found an interaction between message relevance and its perceived persuasiveness that helped to explain participants’ intentions to click through. The stakes of an interaction can also affect the response to visible AI involvement, she says. When AI use is made apparent in lower-stakes interactions, greater trust can be produced, but that disclosure can lead to increased skepticism in higher-stakes interactions, she found. That distinction makes human involvement in interactions especially important. “Disclosure only increases trust when you can tell a human is in the loop,” she said.
The questions to ask here vary by circumstance, AI involvement level and communications channel. To start, consider if a customer can request a human during an AI-agent interaction and how they would do so. Are customers told upfront they can request a human agent? Should that customeralwayshave access to a human? If not, when should that option become available? It’s also valuable to consider how an AI system will recognize a customer’s request for human help and how quickly it should (and can) act on that request.
Customers should have an explicit path to a human when the stakes or emotions of an interaction are high, AI can’t resolve the issue, the interaction involves a dispute or complaint, or the customer simply asks for one, Chippagiri says. That said, it’s not enough to make a human theoretically available if the handoff itself leads to more frustration.
Chippagiri points to systems that promise to connect customers with an agent but then return them to the same AI bot. Other examples include handoffs that force customers to repeat information, leave them waiting without an update, or transfer them to another automated system when they expect a human. “Disclosing that AI is involved does nothing if the escape hatch does not actually open,” Chippagiri said.
Organizations should also consider the stakes of a particular customer interaction. A simple information request and a complaint involving a consequential decision might warrant different approaches for disclosure and human involvement, for example. Should higher-risk or sensitive customer interactions have different requirements for AI disclosure? If so, how do those look? Wolf’s research suggests that visible AI involvement can produce skepticism as the stakes increase. That reinforces the case for ensuring disclosure strategies account for what AI is being asked to do in an interaction, not only if it is or isn’t present.
These questions and the many others that might become relevant for a particular customer/AI interaction help organizations make important distinctions between transparency and agency. Both are important, but for different reasons. Disclosure gives the customer information – in this case, that AI is involved in a customer-service interaction. Choice, however, gives the customer agency to act on that information. Meaningful customer trust regarding AI use will provide some balance of the two.
Disclosure is trickiest with hybrid interactions
Individual interactions where responsibility moves between AI and human agents – hybrid interactions – are the most complicated for disclosure. Consider a customer-service interaction that begins with the customer chatting with an AI agent, then transferring to a human agent who uses an AI copilot, then triggers an AI agent to perform an action in another system while the human agent communicates the result of that action. Who (or what) did that customer just interact with and did that change at various points in the interaction?
From a customer’s perspective, a running technical audit trail in scenarios like these is not necessarily helpful. They won’t have enough context to understand why certain information is or is not being disclosed – and even if they do, they likely won’t have the agency to act on it effectively. At the same time, hiding this transfer of responsibility between AI and human agents is also inadequate for disclosure.
Providing a customer with an effective balance of disclosure and agency in these hybrid CX situations means determining which information is material to that customer. The ideal balance will vary by circumstance, but it’s often important for disclosure and trust for customers to be aware when AI is used to make a decision of consequence, take action on their behalf, access or use sensitive information, generate information that’s presented as authoritative, or replace human judgment.
Chippagiri suggests drawing the line based on if AI materially shapes an outcome or the customer’s decision. When AI makes or recommends eligibility, pricing, or denial decisions, autonomously acts for a customer, or generates the substance of a response without meaningful human review, a line is crossed, he says. On the other hand, behind-the-scenes uses of AI like summarization, routing, spell-checking, or helping to phrase a human agent’s response might not warrant a disclosure.
When disclosureiswarranted, customers still might not need technical details about every system involved in AI’s use. Layered disclosuretion for the customer to get more information if needed, is best in these situations, Chippagiri says. “Describe the role, not the technology stack,” he said
With personalization, those calculations become more complicated. In her dissertation, Wolf examined AI-generated messaging aligned with personality traits of its recipients. Human gatekeeping is still needed because AI lacks the ability to judge if a personality-aligned message is appropriate in a given situation, she concluded. This raises another potential disclosure question for customer-service interactions – do customers need to know if AI is shaping an interaction based on information or inferences about who they are?
Disclosure should be designed around what customers should understand
For Chippagiri, a disclosure banner about AI use can establish transparency, but what happens after that disclosure ultimately determines whether the interaction builds trust or erodes it. Simpler disclosure practices for CX organizations like those aren’t necessarily bad. They are, however, only the beginning of a robust AI disclosure strategy – not the strategy itself.
As AI becomes both less visible and more embedded in customer-service workflows, it will become increasingly difficult for organizations to provide transparency to customersir customers reasonably should know for a given type of interaction, then determine when they need to be provided that information and how it should be given to them
The answers to those three questions will be different across organizations, interaction types, AI involvement levels and communication channels. In considering all these factors and developing disclosure policies, the goal isn’t to disclose every single time AI touches an interaction. Instead, it’s to prevent customers from having a false understanding of who or what is making decisions and acting on their behalf during an interaction, whether that’s being done by an AI or a human agent.
