> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trailercast.io/llms.txt
> Use this file to discover all available pages before exploring further.

# How the AI learns

> Plain-English explainer for customers, prospects, and security reviewers.

> A plain-English explainer for customers, prospects, and anyone evaluating TrailerCast. Updated 2026-05-08.

## The short version

TrailerCast's AI gets better in two ways:

1. **You tell it what you sell and how you sell it** (the [AI Personalization form](/setup/ai-personalization), completed at onboarding).
2. **It watches what works**, which moments your prospects rewatch, which trailers convert, which calls progress to closed-won, and uses those patterns to make sharper picks on your future demos.

What it does **not** do: train Anthropic's general AI models. Look at data from other TrailerCast customers. Send your transcripts anywhere outside the AI providers we list in our privacy policy.

## The two layers

<Tabs>
  <Tab title="Layer 1. What you tell us">
    The AI Personalization form (Settings → AI Personalization). Five fields covering:

    * About your **product**
    * About your **buyer**
    * Your **industry**
    * **Value props** you emphasize
    * **Pain points** your buyers have

    This form **defines the rules** the AI follows for everything it generates for your company. Editable any time, versioned, visible to every AE, not hidden in admin settings.

    The form is the floor. It works from day one, even before you've uploaded a single demo.
  </Tab>

  <Tab title="Layer 2. What your data shows us">
    As you use TrailerCast, the system accumulates signal you didn't tell it.

    **Engagement.** When a prospect rewatches a moment 3 times, "this kind of moment matters." When they bounce at the 8-second mark, "this kind of opening doesn't work."

    **AE corrections.** When you "Rebuild with different clips" and remove a moment, that's "this was wrong." When you mark a call private, reclassify a meeting, or @mention a teammate to flag something, every action is signal.

    The system uses **retrieval-augmented generation** (RAG): when picking moments for your *next* demo, it first searches your *past* demos for the 5 most-similar past moments that prospects actually engaged with. The AI imitates the patterns it sees in your high-engagement past work.

    This is why TrailerCast gets noticeably sharper around demo #10–25.
  </Tab>
</Tabs>

## How the layers combine

For every output (a trailer's moment selection, a call summary, a deal brief, an Ask Winter answer), the AI assembles its prompt like this:

```
1. Base instructions (what we hardcode for everyone)
2. YOUR declared preferences (your form)
3. YOUR proven examples (top 5 retrieved from your corpus)
4. The new content to process
```

Layer 1 sets the rules. Layer 2 sharpens within those rules.

## What the AI does NOT do

<Warning>
  Read this section if you're evaluating TrailerCast for security review or due diligence.
</Warning>

* **It does not fine-tune the underlying AI models.** TrailerCast uses Anthropic's Claude, which Anthropic does not allow customers to fine-tune. Your data never modifies the weights of the AI model itself.
* **It does not share your data with other TrailerCast customers.** Each company's corpus is strictly isolated. Customer A's transcripts can never appear in Customer B's prompts. Enforced at the database level (Postgres row-level security) and verified pre-launch via a documented audit runbook.
* **It does not send your transcripts to train AI vendors' general models.** The AI providers we use (Anthropic, Deepgram, ElevenLabs) are bound by enterprise data-processing agreements. Inputs are processed for your request and not retained for training.
* **It does not silently make changes to your rules.** Even when the AI observes patterns and wants to suggest updating your form, nothing changes without your Global Admin explicitly accepting the suggestion.
* **It does not learn from prospects.** Prospects who watch a trailer aren't "training" the AI. Engagement data informs the SELLER's future trailers, not other prospects' experiences.

## Transparency

Because trust matters more than cleverness, TrailerCast is built to show you why the AI made every decision.

| Feature                       | What it does                                                                   |
| ----------------------------- | ------------------------------------------------------------------------------ |
| Form panel on every AI output | See which preferences were active when this output was generated               |
| Form change history           | Versioned edits, see who changed what, when, revert if needed                  |
| Pending suggestions           | RAG-derived form suggestions stay pending until accepted, with sample evidence |
| Audit log                     | Every AI call logged: which prompt, which retrieved examples, who triggered it |

## Improvement curve

| Stage       | Demos uploaded | What's working                                                  |
| ----------- | -------------- | --------------------------------------------------------------- |
| Day 1       | 0              | Personalization form rules apply                                |
| Demo 1–5    | A few          | Form rules + general AI pattern matching                        |
| Demo 5–25   | A handful      | Form + retrieval surfaces your best past examples               |
| Demo 25–100 | Real corpus    | Retrieval finds strong matches; AI imitates your proven winners |
| Demo 100+   | Mature         | System suggests form updates based on patterns                  |

## Common questions

<AccordionGroup>
  <Accordion title="Does my data train other companies' AI?">
    No. Strict tenant isolation, enforced at the database level. Retrieval can only see your company's own corpus.
  </Accordion>

  <Accordion title="If I delete a demo, is it gone from the AI?">
    Yes. Demo deletion removes the source recording, transcript, AI outputs, and the embeddings used for retrieval. Within \~24 hours the AI no longer references that demo.
  </Accordion>

  <Accordion title="Can I export everything the AI has 'learned' about my company?">
    Yes. The Personalization form exports as JSON. Your corpus is just your demos + transcripts + AI outputs, already exportable. The retrieval index is derivative of those, so re-derivable from the exports.
  </Accordion>

  <Accordion title="What if I'm not happy with how the AI is performing?">
    Three levers, in order: edit your Personalization form (often the answer), review and reject any auto-suggestions you don't agree with, or contact support, we'll diagnose with you.
  </Accordion>

  <Accordion title="Are you using my data to train your own proprietary AI?">
    No. TrailerCast doesn't train any proprietary AI models. We use foundation models from Anthropic (Claude) for reasoning, Deepgram for transcription, and ElevenLabs for narration. Your data is input on a per-request basis under enterprise DPAs.
  </Accordion>

  <Accordion title="What happens if Anthropic changes its terms?">
    Our DPA with Anthropic is a binding contract that survives unilateral terms changes. If a future change ever permitted training on customer data, we'd notify you in advance and give you the option to pause AI features on your account before any change took effect.
  </Accordion>
</AccordionGroup>

## Roadmap

| Capability                                   | Status  |
| -------------------------------------------- | ------- |
| Personalization form (Layer 1)               | Shipped |
| Retrieval over your corpus (Layer 2)         | Q3 2026 |
| Engagement-weighted retrieval (RAG + signal) | Q4 2026 |
| AI-suggested form updates with evidence      | Q1 2027 |
| Per-AE personalization within a company      | Backlog |

## TL;DR

* The AI follows **rules you set** (Personalization form) and **patterns it sees in your data** (engagement signal). Both are visible to you, both are editable, neither leaves your tenant.
* It does **not** fine-tune the underlying model, share your data with other customers, or use your data to train Anthropic.
* It gets noticeably sharper around **demo #10–25** and continues improving with volume.
* Every AI output is **traceable**, see exactly what rules and examples produced it.

Questions? [hello@trailercast.io](mailto:hello@trailercast.io)