New paperModeling the Puppeteer, Not the Puppet: a new thought paper on theory of mind for agent-to-agent systems.Read now
Teleperson

Teleperson Labs

Language, neurobiology, and AI.

Most research on conversational AI studies the machine. We study the person on the other end of it — how speech changes when you know something artificial is listening, what has to be understood about another mind before you hand it your intent, and what survives of language when nobody in the conversation is human.

See the research areasRead the library

Key takeaways

  • Teleperson Labs studies one question from four directions: what happens to human language when the counterpart isn't human.
  • We sit on both sides of the conversation — a consumer agent and a company agent — which is an unusual vantage point for studying the exchange itself.
  • Four areas: language & register, theory of mind & delegation, trust and repair, and machine-to-machine semantics.
  • We publish the open questions alongside the findings, including the ones we can't answer yet.
  • Everything here is public: papers in the research library, shorter pieces on the blog.

The vantage point

We can watch both ends of the same conversation.

Every other platform in this category sees customer service from inside the company — one half of the exchange, instrumented for the side that pays for it. Teleperson runs an agent for the consumer and an agent for the company, which means we can study the conversation as a conversation: what the person meant, what the system heard, where the repair happened, and increasingly, what two agents say to each other when the human steps back.

That is the whole reason this program exists, and it is the one thing a company-side-only vendor cannot replicate.

Research areas

Four directions on one question.

01

Language & register

How does the way people speak change when they know a machine is listening?

Thirty years of phone trees trained a dialect: clipped nouns, over-enunciated digits, the flat repetition of an account number to something that cannot hear frustration. That register outlives the technology that created it — people still use it on systems that would understand them perfectly well in ordinary speech. We study how it forms, what it costs in resolution time, and what actually breaks it.

Open questions

  • How many turns does it take a caller to abandon IVR register once an agent responds naturally?
  • Does register shift transfer across companies, or is it re-learned per brand?
  • Which repair strategies (rephrasing, spelling out, raising volume) predict abandonment?

Published work

02

Theory of mind & delegation

What has to be modeled about another mind before you can hand it your intent?

Delegation is a cognitive act before it is a technical one. To let something act for you, you have to model what it believes, what it will do with ambiguity, and where its understanding of your goal diverges from your own. This is theory of mind — the same faculty that lets people finish each other's sentences — applied to a counterpart that has no biography and no stake in the outcome.

Open questions

  • What signals let a person calibrate how much autonomy to grant an agent, and how fast?
  • When an agent represents a person, whose theory of mind does the other side model — the agent's or the principal's?
  • Does a wrong action by a delegated agent damage trust differently than the same error made by a human?

Published work

03

Trust, load & repair

What makes a conversation feel resolved, rather than merely ended?

Resolution is a felt state, not a status code. It depends on timing, on being understood the first time, on whether the repair — the moment something goes wrong and gets fixed — happens gracefully. Latency, turn-taking and prosody carry more of that weight than the words do. We are interested in the gap between conversations a system logs as successful and conversations a person would describe that way.

Open questions

  • At what response latency does a voice exchange stop feeling conversational?
  • Which failure is more costly to trust: an agent that stalls, or one that answers confidently and wrongly?
  • Can felt resolution be measured without asking, from prosody and turn structure alone?

Published work

04

Machine-to-machine language

What happens to language when nobody in the conversation is human?

When a consumer's agent negotiates with a company's agent, the exchange keeps the shape of language — reference, commitment, repair, grounding — while shedding everything that made it legible to people. Meaning has to survive translation across protocols that were designed independently. We work on the semantics of that hand-off, and on the records it should leave behind.

Open questions

  • What is the minimum shared semantics two agents need to make a commitment that binds?
  • How should meaning degrade — safely — when one protocol has no term for what the other is asking?
  • What does a negotiation transcript have to contain to be evidence after the fact?

Published work

How we work

Published openly, including what we got wrong.

These are working papers and field notes, not peer-reviewed articles — and we say so on the page rather than in a footnote. Every area above lists its open questions next to its findings, because the honest state of this field is that most of it is still open.

Where a claim rests on someone else’s work, we cite it. Where it rests on our own deployments, we say which ones. We would rather publish a narrow result that holds than a broad one that reads well.

Research libraryField notes

FAQ

Teleperson Labs FAQ

Working on the same questions?

We collaborate with researchers in linguistics, cognitive science, and human-computer interaction, and with design partners willing to let the work be grounded in a real deployment.

Get in touchFollow the work