AI, honestly

MADProducts is an AI-assisted marketing OS, so trust here means specifics: which models answer, what happens when they are not configured, what is stored, and what is never claimed. Everything on this page describes what the platform's code actually does — the same refusals and stamps you would hit in the product.

What the AI does

The models

Every writing task walks an ordered chain of writers, first success wins, each step fails over to the next. Which writers are on the chain depends on what the deployment has configured — a writer without its key or install is not on the chain, and an engine without its key is named-unavailable rather than silently skipped.

The chains, per task quality

cli:* stands for the coding-agent CLI writers — on the chain only when installed on the host. The deterministic placeholder is the last step of the standard and economy tiers only; the higher tiers refuse rather than answer with it.

Where a human decides

• Every AI reply to a review starts as a draft a person uses or dismisses; the autopilot is opt-in, star-thresholded (default 4 stars) and waits your window (default 60 minutes) before posting — re-checked at post time against the current rule.
• Conversation threads pause to staff on your handoff keywords, and web-chat placeholder replies wait for a staff release.
• The agent runtime’s action vocabulary is drafts-only — social-post and email drafts, tasks, notes; publishing, sending and paid pipelines are named-unavailable in that vocabulary, enforced by the database.
• Community comment replies post only when the operator turns auto-reply on; the default is a draft for review.
• Every agent run records its model, token counts and cost, and the audit trail stamps each act with its actor’s kind — human, agent, automation or system.

Your data

• Your workspace content is not used to train models. The platform calls model providers through their inference APIs only — there is no dataset upload, no fine-tuning call and no training pipeline in its code. What each provider does with a prompt it receives is governed by that provider’s own terms and policies; we don’t restate their commitments as ours. The promise above is about this platform's own code and configuration — it is not a summary of any provider's policy, and it is not one.
• What a generation sends is the task’s instructions plus the content you asked about — a review’s text, a thread excerpt, your brief. Generated results come back to your workspace like any other content.
• Local options exist: an Ollama endpoint keeps inference on the host, and brand-voice’s local-only mode refuses cloud writers outright.
• Your workspace data — conversations, drafts, agent knowledge, review replies, audit records — lives in the platform’s Supabase Postgres database (this deployment: AWS eu-central-1, Frankfurt), with media in Cloudflare R2. Every AI-surface table is scoped to its workspace’s staff by row-level security.
• Deleting a workspace removes its content. For account deletion and retention details, see the privacy policy.

What we won't pretend

• No review-platform or social reply API is integrated yet: AI replies and conversation sends are RECORDED with what would have been posted, or REFUSED by name — never faked as sent.
• Placeholder output is stamped as a stub wherever it appears, and the automation refuses to ship it: the review-reply autopilot will not post a stub, conversation AI will not send one to a customer, and agent runs served entirely by it are stamped heuristic-only.
• Brand-voice learning is a deterministic statistical extractor over approved content you bound — each source sampled to its first 50,000 characters — never an AI model call.
• Answer-engine citation parsing is a fixed URL regex: the same answer text in, the same domains out — no model reads the answer for links.
• Paid media (Kie stills and clips) and narration (ElevenLabs) run in their own gated pipelines: the autopilot’s paid stage is skipped unless the run was explicitly authorised, and it stops at its budget.
• Without a real writer configured, the surfaces say so by name and refuse or hold — the placeholder is never dressed up as a generation.

Questions & escalation

Questions about AI use: [email protected]. Inside the product, every agent run's model, tokens and cost are in the workspace's AI cost view, and every act — human, agent, automation or system — is in the audit trail with its actor named.

See also the privacy policy and the terms of service.