7 min read
TL;DR
Clementine AI sells behavioural science as a layer on top of an AI phone agent. The only clear public number is an operational one, and the harder claim, that the agent handles hesitation well, is still unproven.
Imagine a woman calling a hearing clinic at 2 in the morning. She isn’t ready to book; she wants to know whether the ringing in her ears means what she fears, and whether asking commits her to anything. How should an AI agent respond to that hesitation?
Matthew is an AI agent from Maastricht startup Clementine AI, founded in 2024. Founder Vince van de Weijer says it reads behavioural science cues and responds to hesitation the way an experienced receptionist might, with the aim of reducing the stigma that keeps people from seeking hearing care.
One Number, Honestly
During a promotional webinar, a Virsono executive said Clementine cut missed inbound calls to zero for the hearing aid brand Earlens. I’ll give it its due, because answering every call is a real result, but it only tells us the phone gets picked up. It’s attributable, not independently audited, and the other advertised gains rest on industry benchmarks, not a company-specific dataset.
Whether the behavioural side changed bookings, disclosure or treatment uptake is a separate question, and nobody has answered it publicly.
A €1.7 million seed round in August 2026, led by Amsterdam-based Healthy.Capital, funds the push. The investor calls the approach “AI-driven activation.” The commercial question is whether that layer improves outcomes over a responsive agent with well-written scripts.
Empathy, Still Manual
Call centres automated routing decades ago, and scheduling software took over most of the calendar work. Conversation was the stubborn part. A good receptionist hears hesitation before the caller admits to it, slows down, rewords a question, and sociologists call that skill emotional labour. Two receptionists can follow the same script and leave a hesitant caller with very different impressions, which is why it’s so hard to bottle.
Healthcare call agents are mostly sold on calls answered, appointments booked and cost per call. Clementine’s pitch reaches further, into why a person hesitates before seeking care at all.
It wouldn’t stop at hearing care either. Debt collection, legal intake and sales all have moments where people dodge disclosure or put off a decision. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, and if more of those systems handle real conversations, behavioural design becomes something companies can buy rather than an informal quality of a good employee.
A fair test would put the agent and human staff on comparable calls, then sort callers ready to book from those still deciding whether to seek care. IC found the same gap between headline accuracy and consequential performance in contract review.
Nice Pays, Apparently
There is some evidence for the broader idea outside hearing care. A debt collection study found AI agents pulled in more money than human staff when their tone fit the borrower’s situation, warm early on and firm after repeated delays, with reported gains running from 49% to 94%. Hearing care is a different setting, so I wouldn’t lean on that result for the same mechanism here.
Behaviour change research offers a sturdier way to ask what an agent is doing. The BCT taxonomy, first published in 2013, lists methods such as social support, credible sources and self-monitoring. Buyers can ask which of those an agent uses, what prompts them and how the responses were tested.
Emotion recognition is on shakier ground. A 2025 preprint looks at whether language models can spot emotional cues and adjust a conversation, though it doesn’t show they’d hold up on live hearing care calls. A Frontiers paper separates signals of care from the human experience behind them, which is the part I keep coming back to.
Persuasion complicates things further. Research across five experiments showed people reacted less favourably to AI sales recommendations once the agent’s profit motive was obvious, and a related paper asks whether customers notice they’re being persuaded. Neither covers hearing care, but together they show how technique, disclosure and caller state affect a conversation, as with any AI help that serves the deployer as well as the user.
Brussels Has Notes
Here the public record gets thin. The material IC reviewed doesn’t name the behavioural techniques, the signals that trigger them, or any behavioural scientist, design specification or academic partner behind the product, so it’s hard to compare Clementine with other agents.
Article 5 of the EU AI Act bans systems that exploit vulnerability tied to age, disability or economic circumstances when that materially distorts behaviour and causes harm. Commission guidance offers the example of a therapeutic chatbot promoting expensive medical products to people with cognitive disabilities.
Hearing aids can be expensive, and hearing loss is more common among older people, but that overlap alone doesn’t bring a system under the ban. Article 5 also requires exploitation, material distortion and harm, and the Commission example is illustrative, not a ruling on hearing care. Still, any company adapting health conversations around vulnerability should be able to say which signals the system reads and what they’re allowed to change. The same question runs through the debate over AI explanations.
Prove The Rest
Hearing care is a useful test because the conversation can affect the decision to seek care. Research from debt collection and sales suggests related techniques can change outcomes elsewhere, but nobody has publicly measured the bridge to this product. Proving the behavioural claim takes a comparison of conversations with and without the adaptive layer, controlling for caller intent and human involvement.
The outcomes to track:
- Completed appointments
- Symptom disclosure
- Treatment follow-through
- Caller trust
The evidence should also show which technique made the difference, and whether the benefit holds once callers are clearly told they’re speaking with AI.
Answering every call points to real demand for hearing care conversations at any hour. Whether behavioural science becomes an AI product category in its own right is still open, and I’d rather leave it open than pretend one webinar number closes it.
For the operating cost of answering those calls, compare Vapi, Retell, Bland and Synthflow. That comparison covers billing and telephony, not the clinical or behavioural performance of a hearing-care agent.
Fine Print
Company-reported results are labelled as such and haven’t been independently audited.
Fair Questions
What would make behavioural science an AI product category?
A problem customers recognise, plus outcomes that can be compared across vendors. Without both, it’s a feature every company describes differently.
Where does Clementine fit in this emerging category?
It couples hearing care call automation with behavioural science language and a stated aim of lowering barriers like stigma. The clearest public result is fewer missed inbound calls, and the public material doesn’t separate out what the behavioural layer added.
Can behavioural techniques improve AI outcomes?
Debt collection research found better results when an AI system matched its tone to the borrower’s situation. That backs the broader idea in one setting, though not that it holds in every industry.
How does European regulation affect the category?
Article 5 doesn’t automatically ban emotionally adaptive systems. The test is whether a company exploits someone’s vulnerability in a way that materially distorts their behaviour and causes harm. Anyone working in sensitive settings will likely need clearer documentation of the techniques and safeguards involved.
Spotted an error? Contact the desk. Our corrections policy explains how we handle corrections and updates.