AI Disclosure

AI is involved in the following features:

• Reply drafts. When a new email arrives, we generate a draft reply in your voice. The draft sits in the inbox for you to edit, approve, or discard.
• Inbound classification. Every email is tagged as a lead, question, objection, promotion, social notice, update, forum post, or other category. This controls what shows up in your CRM.
• Lead scoring. Each contact gets a 0–100 intent score that updates on every reply, based on language signals (timing words, urgency, specific questions, dollar amounts, etc.).
• Intent and stage detection. We infer when a contact has moved from new inquiry → active → under contract → closed, based on the language of the thread.
• Thread summaries. Long threads get condensed to a few bullet points so you can catch up quickly.
• Calendar slot suggestion. When an inbound email mentions scheduling, we read your free/busy windows and propose meeting times in the draft reply.
• Sequence and follow-up generation. Multi-step drip campaigns and standalone follow-ups are written by AI based on the contact's history.
• Voice profile. When you paste sample emails into the voice trainer, we extract a structured profile of your tone, vocabulary, sign-off, and cadence. That profile seeds every subsequent draft.
• Instagram DM replies. For accounts with Instagram DM automation, the same AI reads incoming Instagram direct messages, classifies and scores them, and drafts — or, on autopilot, sends — replies in your brand voice, exactly as it does for email.

ReplyOP currently calls Google Gemini 2.5 Flash for all AI tasks listed above. We chose it for its long-context window, low latency, and per-token economics that let us pass your full thread history with each request.

We may switch or add models in the future (for example to use a thinking-tier model for harder classification problems). If we do, we'll update this page and, where the change affects what data is sent or where it is sent, give you notice in-app.

When AI runs, ReplyOP sends a prompt to Google's Generative Language API (generativelanguage.googleapis.com) over TLS. The prompt typically contains:

• The full content of the email or Instagram DM being processed (subject + body).
• The most recent prior messages in the same thread.
• A short voice profile derived from your past sent mail.
• Your name, signature, and a few contact details for personalization.
• Structured instructions (the system prompt) describing the task.

Google returns the model output, which we store on your account record (as a draft, a classification, a score, etc.).

Per the Google Gemini API terms for paid usage, the prompts and responses we send are not used by Google to train or improve their generally-available models. Google retains data only as needed to provide the service and for abuse monitoring, and does not retain it long-term for training.

We do not opt your account into any data-sharing or model-improvement program with Google. We do not send your content to any other AI provider.

ReplyOP ships with four autopilot modes, set per account in Settings → Autopilot mode:

• Manual. No AI action is taken without your approval. AI still drafts and classifies in the background, but nothing is sent.
• Assist. AI drafts every reply; you approve before sending. This is the recommended starting mode.
• Rules-Based. AI sends automatically only when one of your explicit automation rules matches. Everything else is flagged for review.
• Full Autopilot. AI sends replies on its own when its confidence score is greater than 80%. Below that, it falls back to a draft for you to review. Sequence steps, follow-ups, stage progression, and calendar bookings also run on their own. Items the AI can't confidently handle are surfaced in the flagged queue.

You can switch modes at any time, instantly. You can also override any individual autopilot decision: drafts can be edited or rejected before send; sent emails can be undone within your email provider's send-undo window; sequence enrollments can be paused per contact.

What autopilot does without you. When autopilot is on, ReplyOP can send emails and Instagram DM replies, propose meeting times, book meetings on your calendar, progress leads through stages, and run multi-step follow-up sequences — all without an explicit click from you on each action. Drafts go out under your name and from your email address. Recipients have no way of telling an autopilot reply from one you typed yourself.

Why mistakes happen. AI is probabilistic. Even with a high confidence score, the model can:

• Invent a detail — a price, an address, a date — that wasn't in the thread.
• Propose a meeting time when you're actually busy (e.g. an external calendar we don't see).
• Reply to the wrong person on a forwarded thread, or confuse two leads with similar names.
• Misread tone — sound formal when the lead is casual, or vice versa.
• Apply a rule or sequence step that's no longer appropriate for that contact.

Your responsibility. When you enable autopilot you accept that anything autopilot sends is sent in your name and is your responsibility — including mistakes the AI makes. The first time you switch to an auto-sending mode we'll ask you to acknowledge this in a one-time dialog. Practical guardrails:

• Review the activity log regularly and pause autopilot if anything looks off.
• For high-stakes communication — contracts, offers, payment terms, financial commitments, regulated messages — review every draft yourself, even if autopilot wrote it confidently.
• Pause autopilot for any contact where the stakes change (contract drafted, attorney involved, dispute opens) using the per-contact pause on the contact card.

If you use ReplyOP for real estate, it scans every outbound draft for language that could violate the Fair Housing Act, state fair-housing rules, or the NAR Code of Ethics. When the scanner trips, autopilot pauses and the draft is flagged for your review. The same scanner runs on drafts you write yourself, as an advisory.

The scanner is a safety net, not a guarantee. It can miss subtle violations and it can flag false positives. It does not constitute legal advice and it does not absolve you of compliance obligations. If you are unsure whether a message complies with applicable fair-housing rules, do not rely on the scanner — consult your broker or counsel.

ReplyOP provides AI assistance "as is." We make no guarantee of accuracy, completeness, or appropriateness of any AI output. Any communication sent through ReplyOP — drafted, suggested, or auto-sent — is your responsibility. You agree to review AI-generated content before relying on it for legal, financial, or commercial decisions.

You have several layers of opt-out, from broadest to narrowest:

• Disable autopilot entirely. Settings → Autopilot mode → Manual. Nothing is sent without explicit click-through.
• Disconnect your inbox. Settings → Email accounts → Disconnect. We stop syncing new mail and stop running AI on your behalf, though your existing data is retained until you delete it.
• Delete a sequence or rule. Stops the specific automation without affecting the rest.
• Delete your account. Settings → Danger zone. Removes all of your data and revokes our connected Google and Meta OAuth tokens.

If at any point you want a copy of every record we hold for you before deletion, the "Download all my data" button in Settings → Account produces a complete JSON export.

AI output is probabilistic and imperfect. In particular:

• Hallucinations. The model can invent facts (a price, an address, a meeting time) that aren't in the source thread. Always read drafts before sending.
• Tone drift. The voice profile is a best-effort approximation. On unusual topics it may sound off-brand.
• Misclassification. Important leads can occasionally land in the promo bucket, or vice versa. The CRM filter is reversible per email.
• Score volatility. Lead scores are signals, not verdicts. Treat them as a sort order, not a decision.

Concerns about how AI is used in your account, or requests to change a default: info@optovo.co.