Doctoral Position in Customer-Facing AI in Marketing and Service

At the Center for AI Value, part of the Technology Marketing Group at ETH Zurich, we investigate how organisations can turn generative and multimodal AI into measurable value. Our current focus is one of the most consequential open questions in this space: how customer-facing AI should be designed when it speaks to real customers on behalf of a real firm. Generative voice agents now handle substantive service conversations, hand customers over to human employees, and generate content that reaches large professional audiences — yet the design decisions behind them are made largely by intuition.

We are looking for a doctoralcandidate to help change that. The position is built around three projects, all of them conducted with industry partners and all of them designed to move from controlled experiments into live operations. You will study how tightly a firm should prescribe what its AI agent says, what happens emotionally and informationally when a conversation is handed from an AI to a human colleague, and how multimodal models can predict the performance of marketing content before it is published.

Research at the Center builds on the group's track record in data science, conversational AI, and marketing analytics and on substantial investments in on-premises GPU infrastructure that allows us to work with confidential partner data under full data control. The focus of the position is on research, aiming at publications in leading international management journals. The starting date is flexible. Applications will be accepted until the position is filled.

For further information please refer to Sebastian Tillmanns.

The position spans three projects, each carried out in close collaboration with an industry partner and each combining controlled experimentation with access to real operational settings.

The projects build on established collaborations, ongoing research, and technical infrastructure already in place at the Center.

Prompt granularity in AI voice agents. Firms deploying generative voice agents must decide how tightly to prescribe what the agent says. Tight specification buys reliability, completeness and policy compliance; open specification buys adaptivity to cases nobody anticipated. This project treats that decision as the machine analogue of the classic scripting-versus-empowerment trade-off in service operations — with the crucial difference that script strictness, an aspiration with human employees, becomes an enforceable parameter with a large language model.

The programme comprises laboratory experiments with live voice interaction, a comparison of adaptive prompting policies, and a field experiment in the live call operations of a partner in the insurance sector. Outcomes are deliberately behavioural: verified task success, conversational repair derived from transcripts, and compliance failures such as unauthorised commitments.

Emotional grounding in AI-to-human handovers. When a customer in a negative emotional state is transferred from a voice agent to a human employee, does the emotional acknowledgement offered by the AI carry over — or does it retrospectively read as a script if the human does not build on it? This project examines whether the two agents jointly establish common ground about the customer's situation, and how that perceived continuity shapes whether customers feel heard and go on to recommend the provider.

It progresses from a factorial experiment through studies with trained actors simulating the handover to a field implementation with the same insurance partner.

Multimodal prediction of marketing content performance. Building on an ongoing doctoral project and an established data collaboration in industrial technology, this project develops and benchmarks multimodal embedding approaches, text and image, encoded separately or jointly, as inputs to models that predict the engagement a social media post will generate before it is published. The successful candidate will work closely with the doctoral researcher already active in this research stream.

Together, the projects give the successful candidate something rare: theoretically grounded…