AI usage transparency

Why we use AI at all

The hard part of retail outreach isn't sending email. It's writing 200 pitches that each feel hand-written for that specific buyer at that specific store with that specific stocking pattern. Doing that by hand is ~5 hours per 50 pitches. AI gets it to 5 minutes — but only if a human reviews and edits each one before send.

That's the trade-off Pitch2Retail makes: AI for the first draft, human for the final word. We don't pretend AI alone is enough.

What models we use

We standardize on Anthropic for one reason: their data usage policy for API customers is the strictest in the industry — no training on API inputs or outputs by default. We do not multi-vendor across OpenAI / Google because each new provider is a new privacy surface.

What data we send to Claude

For each AI operation, exactly this and no more:

Pitch generation

• Your brand profile (your inputs: brand name, products, positioning, mission)
• Recipient business context: company name, role, industry, store location, public store info from Google Places
• Recipient first name (for the salutation only)
• Selected pitch template + tone instructions
• Compliance footer + sender block

Reply classification

• The reply text (with email signature stripped before send)
• Original subject line and pitch summary (for context)
• No prior thread history is included beyond the immediate reply

Account research

• Public website content fetched at research time
• Company name, public LinkedIn description (if any)
• Selected news / press release URLs you authorize

Vision audience filter

• Up to 3 storefront photos from Google Places per candidate retailer
• Your filter criteria ("does this store stock specialty kitchen tools?")
• Photos are not retained after the filter step — held in memory for the API call only

What we don't send to Claude

Anthropic does not train on your data

Per Anthropic's commercial terms:

"Anthropic will not train its models on Customer Content from API services."

This means your pitches, your audience data, your replies — none of it ever becomes training data for future Claude models. Anthropic retains API request logs for up to 30 days for abuse detection only, then deletes them.

For Growth+ tier customers, we can request zero-retention mode from Anthropic where logs are deleted within 24 hours instead of 30 days. Email hello@pitch2retail.com to enable.

Human-in-the-loop review

This is non-negotiable in Pitch2Retail's design:

1. AI drafts a pitch → goes to your Review queue
2. You see the draft side-by-side with the recipient profile
3. You can edit, regenerate, skip, or approve each pitch
4. Only approved pitches go through the send pipeline
5. "Approve all" is available for batches you've spot-checked, but every individual pitch is still rendered for your review first

There is no autonomous-send mode. Even on follow-up sequences, the next-touch drafts queue for review before they go.

This makes a difference for GDPR Art. 22 (no purely automated decisions affecting recipients) — see below.

How to disable AI features

You can run Pitch2Retail with AI fully off:

• Settings → AI features → Off — disables Claude on your tenant
• Pitches are drafted from your templates only with mail-merge variables ({{first_name}}, {{company}}, etc.)
• Reply classification is replaced with a flat "needs review" state
• No price reduction (we still pay platform fixed costs) but full feature parity for non-AI flows

Should you disclose AI usage to recipients?

Currently no jurisdiction requires it for B2B email outreach. But several may soon:

• EU AI Act (Aug 2025+): requires disclosure when AI generates "synthetic content" intended to inform/persuade. Likely covers AI-drafted marketing email.
• California SB 1047 (proposed): AI labeling rules for commercial communications
• FTC guidance (US, 2025): material misrepresentation rules cover undisclosed AI-generated content claiming to be human

Our pitch templates default to a neutral framing — they don't pretend to be 100% hand-written, and they don't fake "I was just thinking about you." We recommend (not require) you add a footer line like:

"This outreach was drafted with AI assistance and reviewed by [your name] before sending."

Toggle this on in Settings → Compliance → AI disclosure footer. We'll add it automatically to every pitch.

Hallucination and factual accuracy

Claude can occasionally invent facts — a fake program name, a misattributed quote, a company history detail. We mitigate but don't eliminate this:

• Grounded prompts: we feed Claude only verified facts about your brand and the recipient, and instruct it not to invent unstated information
• No claim escalation: we strip phrases like "voted #1", "industry-leading", "guaranteed" that the AI sometimes adds
• Compliance lint: banned-phrase scanner runs on the AI output before it's queued for your review
• Your review: you're the last line of defense — read every pitch before approving

If you spot a hallucinated fact, click "Report" on the pitch — it goes to our quality queue and feeds the prompt-tuning loop.

Regulatory note: EU AI Act + GDPR Art. 22

For EU/UK recipients, two additional considerations:

• EU AI Act: Pitch2Retail's pitch generation is classified as "limited risk" (generates synthetic text) — disclosure obligations apply but no high-risk classification or conformity assessment is required. We comply via the optional AI disclosure footer above.
• GDPR Art. 22: recipients have the right not to be subject to "purely automated decisions" with legal or similarly significant effects. Receiving an AI-drafted business outreach email isn't covered — sending a pitch isn't a legal decision. But we still keep human-in-the-loop review as a defense-in-depth measure.