Science1 distinct publisher3 min readPublished
Two studies from one team suggest frustrated customers gain something from the slowness of typing, awkward timing for the banks and postal services scaling humanlike voice agents. Both measure self-reported feeling rather than resolved tickets.
The Scientist · Science desk
Compiled by The ScientistSomething wrong?How this is made
The channel-preference figure usually cited in voice's favour was collected under the one condition that removes the problem. In the Australian Contact Centre Industry Benchmarking Report, phone led at 27% against 12% for web chat among people who were told their issue would be resolved [6]. That is the easy branch. The European Journal of Marketing studies deliberately sit in the other one, where something has already gone wrong and the customer arrives carrying it [1]. A preference expressed about a solved problem is weak evidence about an unsolved one. The two percentages also sum to 39, which leaves 61% of that sample picking channels the write-up never lists [12].
The satisfaction gap needs the same care. Of 1,031 Australians surveyed, 78% said they had used automated customer service and 39% said they were satisfied [5]. Read as whole-sample figures, that is roughly 804 users and roughly 402 satisfied people, or about half of users content rather than two in five [13]. Read as a statement about users only, the satisfied count falls to about 314 [14]. The article does not say which, and the gap between those two readings is the gap between a mediocre channel and a poor one.
The proposed mechanism for the relief effect is mundane, and to me the most interesting part: writing is slower and more deliberate than speaking, so a customer who has to type gets a chance to redirect attention while some of the heat comes off the complaint [7]. Note where the causal weight sits. Modality was assigned by the researchers, so the relief difference between chatbot and voicebot is an experimental result. The step from relief to better evaluations is reported as an association [3], which is the ordinary limit of a measured mediator: relief may be doing the work, or both may follow from something the design did not isolate.
What this does not tell you is whether relief reaches anything a contact centre books. The write-up gives directions and self-reports without sample sizes, effect sizes, or field outcomes such as repeat-contact or resolution rates [16], and the chatbot advantage is qualified as holding in some situations, with the moderators unstated [2]. It is written by one of the study's co-authors [11], which is normal for the format and worth knowing. The deployment numbers cannot be lined up against the research either: the bank counts voice and messaging in a single total [8], and Australia Post's reported 55% self-service AI share for 2025 does not identify a channel [9].
My view, with its conditions. The cheaper of the two findings to act on is the emotion-labelling one, since saying "I understand you are feeling angry" rather than "feeling bad" raised perceptions of emotional competence and satisfaction [4], and that is a script change rather than a model change. The cooling-off finding deserves a production test before anyone rebuilds a routing tree around it, because self-reported relief in a study and a repeat contact avoided in a live queue are different measurements.
Ranked by verification strength, evidence, and original report placement.
In one study, customers interacted with a text-based chatbot rather than a voicebot after experiencing a service problem; in some situations the less humanlike option produced more favourable customer responses.
Participants who interacted with a chatbot reported greater feelings of relief than those who interacted with a voicebot, and those feelings of relief were in turn associated with more positive evaluations of the service experience.
Research published in the European Journal of Marketing by Mai Nguyen and colleagues involved several studies examining how people respond to different forms of AI customer service following service failures.
In separate research published in the Journal of Business Research, a robot that said "I understand you are feeling angry" was compared with one that said "I understand you are feeling bad"; customers responded more positively to the accurate identification, and more precise emotional recognition increased perceptions of the agent's emotional competence and improved customer satisfaction.
The Australian Contact Centre Industry Benchmarking Report found that, when people were told their issue would be resolved, a phone call was the preferred channel at 27%, compared with 12% for web chat.
The Commonwealth Bank recently revealed its AI-powered customer service platform handles more than 2 million conversations every month across voice and messaging channels.
Distinct publishers with included, body-backed reporting in this cluster.
phys.org
1 article · September 2, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
leadership
Uber spent a year of AI coding budget in four months. The cap is not the fix.1 distinct publisher
product
Mastercard's Australian outage came from a scheduled update, which is the part worth auditing1 distinct publisher
leadership
Your Face Is Now Inventory: ASIC Logged 19,400 Scam Reports and $7.4m Lost to Ten Deepfaked Figures1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Two DOIs, not one number
Both papers are named with journals, years and DOIs, which is more provenance than most AI-and-customers stories offer. But every quantity a reader would need to judge the relief effect is absent — how many participants, how large the gap, whether it held outside the hedge 'in some situations' — and the outcomes are self-reported feeling rather than a fixed problem. The satisfaction gap propping up the argument arrives as 'one recent survey', unnamed and untraceable.
The deployments run the other way
The only real-world numbers here belong to the trend the research questions, not to the research: a bank at two million AI conversations a month across voice and messaging, a postal service at 55% of inquiries self-served. Both are company disclosures passed along, and neither is broken out by channel, so they establish that humanlike automation is scaling in Australia and nothing about anyone rethinking text versus voice on the strength of these findings.
Careful piece, travelling claim
phys.org's own prose is disciplined — 'not always the answer', 'may occur', 'in some situations' — and it ends on a design question rather than a verdict. The overshoot happens in transit: 'chatbot users reported more relief in a lab-style comparison' turns into 'typing is better for upset customers' the moment the hedges drop, and the cooling-down mechanism is presented as explanation when it was never tested. Two of the three supporting statistics also lean harder than they can bear.
The advocate is the only witness
The byline belongs to a co-author of both papers, and to its credit the piece says so outright. That is the normal shape of phys.org's academic explainers and it is also why nothing here pushes back: the summary of the studies, the choice of which Australian statistics to quote, and the decision to leave out sample sizes are all made by the people whose work benefits from the finding landing.
One voice, nothing checked
A single publisher, a single author, and that author is inside the research. The papers themselves are locatable by DOI, which keeps this above guesswork, but no independent reporter has read them against the summary, no deployer has been asked what its own channel data shows, and an unresolved ambiguity in the satisfaction figure sits unexamined in the middle of the argument.