04 · How AI on Dilly works

Look behind the curtain.

Most students assume their resume becomes training data. It doesn't. Below is exactly what each AI feature does, what data it sees, and what it never sees. If your mental model of "AI app = your data trains the model" is wrong here, this is where you can verify it.

Fit Narrative
Anthropic · Claude Haiku
Sent to the model
Your published profile facts and the job description text.
Not sent
Your conversations with Dilly. Your draft resumes. Other users' data. Your name, school, or contact info.
Cached
5 minutes per profile, then dropped from cache. Output saved to your account.
Used for training
No. Anthropic processes under zero-data-retention enterprise terms.
Tailored Resume
Anthropic · Claude Sonnet
Sent to the model
Your profile facts, the target job posting, and the resume template you picked.
Not sent
Your conversations. Other resumes you've generated. Other users' content.
Cached
The output is saved to your account. The prompt is dropped after the call.
Used for training
No.
AI Coach
Anthropic · Opus on Pro · Sonnet on Dilly · Haiku on Starter
Sent to the model
Your last few messages in this thread, your profile snapshot, and relevant context (jobs you saved, deadlines).
Not sent
Other users' conversations. The full Dilly content corpus. Anything you marked private.
Cached
Conversation history stays on your account, deleted when you delete your account.
Used for training
No.

Provider status

Verified contract status with each AI vendor we use. Current as of May 4, 2026. Provider terms change; we re-check quarterly and ship a changelog entry if anything moves.

✓
Anthropic
Enterprise · Zero data retention · No training on inputs or outputs
✓
OpenAI
API · No training (OpenAI default since March 2023)

If you've ever wondered "is my resume training the AI?"

Short answer: no. Long answer:

When you talk to Dilly, your message is sent over an HTTPS API call to Anthropic (or, for a few features, OpenAI). Both providers operate API tiers under contracts that explicitly forbid using your data to train models. Anthropic's enterprise plan goes further: zero data retention, meaning the message isn't even stored on their side after the response is generated.

What this looks like in practice: you ask the AI Coach a question, the question and your profile snapshot are sent in a request, the model returns an answer, and the request is discarded on Anthropic's end. We save the answer to your account so you can read it later, but the original prompt and response are not kept anywhere outside your account.

If a future change to a provider's terms ever weakens this, we'll either move providers or ship a changelog entry that says exactly what changed. We're not going to let "we updated our subprocessor terms" turn into a silent shift in how your data is treated.