I. The problem

Agents answer. That is not the same as doing the job.

You have been through it. The agent replies fast, politely, and beside the point. You explain again. It offers the same three links. So you stop explaining and start guessing: agent, human, representative, complaint — which word gets you out of here?

Nobody wanted this. The company pays for the agent and gets angrier customers. The customer wanted one thing done. The agent understood every word and missed what was happening.

That gap is the whole problem, and it is not a problem of politeness or tone. The model has every word for people and no working model of what a person is — so it has no way to hold on to what this one is doing: complaining, checking, asking for a decision, trying for the third time, deciding whether to stay a customer.

Your agent recognises the intent of every message.
It misses the macro-intent of the person.

Intent is what this sentence asks for. Macro-intent is what the person came for and is still after five messages later — the thing they will judge the whole encounter by. An agent that tracks only the first answers each message correctly and loses the person.

II. What we sell

One thing. Licensed, not published.

[ Watch the ontology in practice ] →

In practice it gives an agent the one thing it has no way to represent on its own: the macro-intent of the person — what they are after across the whole encounter, what it is costing them, and what would have to change for them to leave satisfied. That is the difference between a question and a complaint, between the first attempt and the fourth, between someone who wants information and someone who wants a decision — and so between answering and getting the job done.

It is model-agnostic and vendor-agnostic; it sits above whatever you are running. Presence Mechanics is the platform: the ontology, an API engine (in preparation), and the applications built on it. AI'm Director™ — the second tab — is the first application and the working proof. Further applications are in preparation.

Here is our strongest claim, and it is testable. Given nothing but the ontology and a single take, AI'm Director produces deep readings of what is going on inside the person performing — and they are startlingly true. Our founder has verified them. The demo is public, so you can judge that for yourself before you talk to us. Watch it →

We are a startup, and a particular kind of one. We take what acting, directing and dramaturgy have worked out about the human being — and carry it over to AI. Early stage: patent applications filed, one product live, looking for the first licensing partners.

The specification is not on this page. The full structure, the documentation and any verification you need before you buy are shared under NDA.

Use cases

• Customer support agents
• Tutoring and negotiation bots
• Acting and creative AI tools
• Enterprise agents — banking, telecom

III. What Presence Mechanics is not

Four things people assume we do, and we do not.

• NOWe do not recognise emotions from facial expressions. No face reading. Nothing in the system looks at a face and decides what the person is feeling.
• NOWe do not measure micro-expressions or body language. No landmarks, no gesture inventory, no posture score. We make no claim that a twitch means a thought.
• NOWe do not match vocal tone or tempo. No prosody matching, no mirroring, no warmth dial. An agent that copies your speed has learned nothing about you.
• NOWe do not label the conversation with adjectives. Not satisfied, not angry, not frustrated. An adjective closes the case — it is the moment a system stops finding out what is going on.

It is also not a dataset, not a prompt pack and not an empathy layer. There is no corpus here, and nothing that tells a model to sound warmer than it is.

Everything on that list measures a surface. We describe what a person is doing. Those are two different things, and the next section is about what happens when they get confused.

IV. The manual of faces

This is the road the industry is taking with AI. The theatre took it first, and it was absurd then too.

Le Brun gave painters a method: one page per passion, each with its own prescribed brow, eye and mouth. The plates were reprinted for a century and a half, and the method crossed into the theatre. Eighteenth-century acting manuals taught passions the same way: the actor learned the face that jealousy was supposed to have, and put it on when the line came.

The theatre fell for it completely, and paid for it. It worked, in a sense: the audience could name what it was being shown. Nobody believed a person was there. It took the theatre the better part of a century to get out of that habit — you cannot reach a human being by assembling the parts of a face, and in the end everybody in the room could see it.

Put landmarks on a face, score the geometry, output ANGER 0.91 — and you have rebuilt Le Brun's manual with a tensor instead of an engraving tool. Same inventory of faces. Same premise. Faster. Absurd.

This is the direction most of the industry is taking. Affect detection, sentiment scores, emotion APIs bolted onto agents — a seventeenth-century idea with better hardware, sold as understanding people. It is the wrong road, and the theatre already walked to the end of it and turned back.

It is also why agents built that way break exactly where it costs money: they can name a face and still have no idea what the person wants, what they have already tried, or what would make them leave.

V. What we took from the rehearsal room

Why a theatre practice produced a better structure than an annotation schema did.

Rehearsal has never had access to inner states. A director cannot ask for a feeling and be handed one; an actor who tries to manufacture one produces the Le Brun face. So the craft had to be built on what can actually be given, repeated tomorrow, and understood correctly by a thousand strangers at once: what someone is doing, to whom, under what conditions, against what resistance, with what at stake, and how they change course when it does not work.

That is a structure you can compute with. A feeling-word never was. Practitioners tested it every night against live audiences for centuries — a brutal evaluation loop. We formalised what survived, and that is what we license.

VI. A note on the AI Act

Short version.

VII. Who is behind this

Bartłomiej Malik — founder of Presence Mechanics and inventor named in both patent applications. Actor and director by training, which is the point: this structure comes from the room where it was tested, not from a literature review.

The quickest way to judge the work is to watch it run: the AI'm Director demo →

VIII. Protection

What covers the work.

IX. Contact

One address, one next step.