Which AI agents we use, what data is shared, and what a human always does.

Which models we use

Model	Provider	Used for
Claude Opus / Sonnet	Anthropic	Strategy, long-form writing, code review, the bulk of client and studio thinking
Gemini Flash / Pro	Google	Bulk content variants, batch translation, Vinted listing generation in VintSnap, manual parsing in BuildBuddy, tariff classification in TariffShield
Perplexity Sonar Pro	Perplexity	Real-time research with cited sources
OpenAI GPT models	OpenAI	Selectively, as a second opinion or for code-focused tasks
Veo / Sora	Google / OpenAI	Short-form marketing video
ElevenLabs / Fish Audio	ElevenLabs / Fish	Voiceover for marketing video, in 24+ locales
Midjourney / Flux Pro	Midjourney / Black Forest Labs	Image generation for marketing

We rotate models as the market moves. The current routing matrix is published on the operations page.

What data goes to which model

Client work

Strategic context, briefs, internal documents, draft copy, code. Always under the model provider's API terms — not consumer ChatGPT/Gemini/Claude apps. We do not share client data with consumer chat interfaces.

For sensitive client engagements (regulated industries, NDA-bound), we use enterprise-tier endpoints with zero-retention configurations where the provider supports them.

Product end-user data

Each product handles user data differently. In summary:

• Titra: By default, all logs stay on the device. titra Pro adds optional encrypted cross-device sync. Anonymous usage analytics only. Health data never goes to analytics.
• VintSnap: Photos sent to Gemini for analysis are not stored after the analysis. Listing data and history live on the device with optional cloud backup.
• TariffShield: Product data is sent to Anthropic for classification under Anthropic's API terms. We retain only what's needed to serve the merchant. No personal customer data leaves the merchant's Shopify store.
• SitePrice / Sibliny / BuildBuddy: Per-product privacy disclosures will be added when each product launches publicly.

Product-specific privacy policies live on each app's own privacy page. The master AgentM privacy policy covers studio engagements and the public agentm.co.uk website.

What humans always do

The hard rules from our operating contract:

• A human reviews every email before it sends
• A human reviews every social media post before it publishes
• A human reviews and approves every paid ad before it spends money
• A human reviews and approves every public app store change
• A human approves every customer-facing communication, including app review responses
• A human is named on every piece of long-form content — bylines aren't ambiguous
• A human handles every escalation, dispute, or sensitive customer interaction
• A human signs every contract and authorises every payment

The full operating cadence is at /operations.

What we will never do

• Auto-publish to social media without human approval
• Auto-send emails to your customers based solely on AI judgment
• Sell or share product end-user data with third parties for marketing
• Train AI models on client confidential data without explicit written consent
• Pretend AI-generated text is human-written when bylines, attribution, or context call for clarity
• Use AI to impersonate real people, generate misleading testimonials, or fabricate reviews

Subprocessors

The AI providers above are subprocessors for both client work and product operations. We rely on each provider's published security and data handling commitments. The full subprocessor list — including infrastructure (Vercel, Supabase, Cloudflare), analytics (PostHog, GA4), error tracking (Sentry), and customer comms — lives at /trust/privacy#subprocessors and is updated when it changes.

Got a specific question about how we use AI?

We answer real questions on the record. Email hello@agentm.co.uk with the subject "AI disclosure" and we'll respond within two business days.