Responsible AI. AI-Native.
How we use AI.
Last updated May 29, 2026
CallSherpa is built for open calls (grants, fellowships, proposals) where decisions affect real careers and real projects. We use AI to help reviewers work faster, not to replace their judgment.
This page sets out exactly what AI does on CallSherpa, what it doesn't do, and how we handle the data that runs through it. If any of this changes, this page changes with it.
Responsible AI is the constraint set below: what AI does and doesn't do on this platform, and how the data is handled. AI-Native is the architecture. CallSherpa was designed around AI assistance from day one, not retrofitted onto a pre-LLM workflow.
1. AI never judges submissions.
CallSherpa's AI will not score, rank, or evaluate a submission. It will not say “this is strong” or “this is weak.” Judgments come from your reviewers.
This is the clearest difference from competitors. Submittable's Review AI and OpenWater's prompt-driven Intelligence both allow AI to score submissions against rubrics. CallSherpa does not, by design.
2. AI never auto-rejects.
No submission is rejected, shortlisted, or filtered by AI. AI may suggest pass / flag / fail during eligibility screening, with an explanation, but a human screener confirms the decision before any applicant is filtered out. Every submission that meets your eligibility criteria reaches a human reviewer.
3. Every decision is a human decision.
Reviewers form their own assessments. AI drafts text (a summary, a translation, a decision letter) only when a reviewer directs it to. AI output never appears unprompted in a reviewer's queue.
4. Submissions are never used to train AI.
We send submissions only to model providers that contractually do not train on customer data. We enforce this on every API call, including when an organizer brings their own OpenRouter key.
5. AI use is disclosed.
Applicants and organizers can see where AI was used, what it did, and which model was used.
In product, AI-drafted reviewer notes and AI-drafted decision emails are labeled. Organizers see per-call AI activity in the call dashboard.
6. Organizers control AI.
AI features are optional. You can run a call on CallSherpa with AI fully off. AI can also be toggled per feature (summaries, translations, decision letter drafting) independently.
The models we use
CallSherpa uses a small, vetted set of model providers. Each is listed below with what it does in our platform and its data policy.
| Model | Used for | Trains on inputs? | Retention |
|---|---|---|---|
| OpenAI GPT-5 nano (via OpenRouter) | AI eligibility pre-screening; lighter reasoning | No | Per OpenAI API policy |
| OpenAI GPT-5.4 nano (via OpenRouter) | Reviewer chat, call control center, navigation, data agent | No | Per OpenAI API policy |
| OpenAI GPT-5 mini (via OpenRouter) | Translation; call extraction from PDF and text | No | Per OpenAI API policy |
| Google Gemini Live | Hands-free voice mode (AI Max tier). Both voice in and voice out are processed by a Gemini multimodal LLM. | No | Streamed only; not stored by CallSherpa or Google beyond the session |
OpenAI does not train on API inputs by default (effective March 2023). On every OpenRouter request, CallSherpa sets provider.data_collection: "deny" to exclude any provider endpoint that would log or train on inputs.
Voice synthesis
On the AI Included tier, when Sherpa speaks, the reply text is sent to Inworld TTS (model inworld-tts-1.5-mini) for speech synthesis. Inworld receives Sherpa's reply text only; applicant submission content does not reach Inworld.
Per Inworld's Terms of Service, Inworld does not train its generally available models on customer materials, with narrow exceptions for explicit feedback and trust-and-safety review. CallSherpa does not submit either.
On the AI MAX tier, voice (both directions) is handled by Google Gemini Live, listed above.
BYOK (Bring Your Own OpenRouter Key)
With BYOK, you supply your OpenRouter API key and usage bills to your account. CallSherpa still chooses which models are called. The platform routes through the same vetted no-training models we use on AI Included, listed above. BYOK changes who pays for usage. It does not change which models CallSherpa uses or how submission data is handled.
Free OpenRouter models that train on inputs are not used by CallSherpa on any plan.
The model list is reviewed quarterly and updated whenever a provider's policy changes.
What this means for applicants
- No AI decides anything about your submission. A human reviewer reads it and makes the decision.
- AI may be used to summarize or translate your submission, or to draft a response to you. When this happens, you'll see it labeled, including which model was used.
- Your submission is not used to train AI models. It's sent only to providers that contractually agree not to train on it.
- You can ask the organizer how AI was used on your application. Organizers using CallSherpa have access to that information.
What this means for organizers
- You can turn AI off entirely. Per call, in your settings.
- You see which models are being used and for what.
- You decide what AI helps your reviewers with. Summaries, translations, decision letter drafting. Toggle each one.
- You can adopt reviewer guidelines aligned with federal grant standards. NIH prohibits peer reviewers from using AI for critiques; NSF says current generative AI does not meet their confidentiality bar. CallSherpa's six commitments map cleanly onto a posture a federal-grant-aligned program can adopt.
- If you bring your own OpenRouter key, CallSherpa still picks the models. You pay for usage on your account; the platform routes through the same vetted no-training models we use on AI Included. This keeps commitment #4 intact even on BYOK.
Questions about how we use AI? Email help@callsherpa.ai.