What changed
Most chatbot transparency debates focus on the customer side: should people be told when they are talking to AI, and should they also be told that a human might step in? This paper asked a narrower operational question that matters for real service teams. When customers know that unseen employees are working behind a chatbot, do they start talking differently, and does that change how much work falls back to people?
The authors studied a hybrid service agent used by a multinational telecommunications company on its website. The system handled common questions about contracts, Internet problems, mobile plans, and related service issues. Customers were randomly assigned across four disclosure conditions: no disclosure, up-front disclosure that humans might step in, disclosure only when a human actually stepped in, or both. Across 8,966 chats, both kinds of disclosure significantly pushed customers toward a more human-oriented writing style: longer, denser, more natural messages instead of short keyword-like requests.
That change mattered because the chatbot was worse at handling those more human-style messages automatically. In the descriptive comparisons, mean human-response frequency rose from 0.74 in the no-disclosure condition to 1.26 in the condition with both disclosures. Mean employee handling time rose from 0.64 to 1.05 minutes, and mean response intensity rose from 1.76 to 3.02. In the structural model, customer communication style had a strong positive association with employee workload.
The surprising detail is that disclosure before the conversation started had at least as much effect as disclosure at the moment a human stepped in. The paper's explanation is psychological rather than technical. Once customers think a real person may read the exchange, they start managing impressions. The follow-up controlled online experiment with 288 participants supported that mechanism and replicated the main direction of the effect.
What this could change for you
For ordinary users, this helps explain why some chat tools feel oddly contradictory. A company may promise transparency and human backup while also nudging you toward short messages, limited text boxes, or highly structured prompts. That is not necessarily bad design. It may be a way to keep the chatbot usable once customers stop treating it like a search box and start writing to it like a person.
For service teams, the result is a warning against treating disclosure as a free ethical upgrade. If regulation, policy, or company standards push firms to disclose more about human involvement, some of the expected automation gains can shrink. More chats get kicked to employees, employees spend longer on them, and staffing assumptions can break if that transition is not planned.
The broader AI lesson is that human-AI systems are shaped by the people interacting with them, not only by the model and the workflow. A chatbot can look technically unchanged while its effective automation rate drops because disclosure altered customer behavior upstream.
What it does not prove
This was one hybrid chatbot at one multinational telecommunications company. The paper does not prove the same workload increase would transfer unchanged to banking, healthcare, travel, ecommerce checkout, or consumer chat tools with stronger language understanding or narrower task design.
The field experiment measures communication style and employee workload, not end-to-end customer welfare. The paper does not show that customers were less satisfied overall, that transparency should be avoided, or that the firm's best policy is secrecy. In fact, the authors report that many participants wanted transparency and said they would feel misled if they discovered human involvement later.
The effect is also about a specific kind of disclosure: human involvement behind a predominantly AI-based service interface. It does not settle the broader legal or ethical debate over chatbot identity disclosure, nor does it prove that better modern language models would erase the workload trade-off entirely.
The bottom line
AI did not solve customer-service automation as a pure engineering problem. It also created a disclosure problem. In this randomized field experiment, telling customers that humans were behind the chatbot changed how they wrote, and that in turn pushed more work back to employees. The honest takeaway is not that transparency is wrong. It is that honest AI service design can cost more human labor than firms expect.
Primary research
More Than a Bot? The Impact of Disclosing Human Involvement on Customer Interactions with Hybrid Service Agents
Information Systems Research · 2024 · DOI 10.1287/isre.2022.0152
