Security & privacy

Everything you send to Decidi travels over TLS, so it is encrypted in transit. At rest, your account records, decision history and uploaded files are stored encrypted by our hosting provider. There is no point in the flow where your brief crosses the network in the clear.

Account data, debates and credit ledgers are stored in Supabase, hosted in the European Union region. We deliberately keep what we store narrow — your email, the decisions you run, the resulting transcripts, and the usage needed to meter credits accurately.

Files you attach — photos, screenshots, documents — are used only to inform the active debate you attached them to. They are not repurposed, not shared, and not used for anything beyond generating that council’s response. You can delete uploaded files, and we will remove them on request independently of the rest of your account.

Decidi does not use your prompts, files, transcripts or verdicts to train any model. We do not sell your content. When your debate is sent to the underlying providers, we instruct them not to retain it for training where their API allows that instruction, and provider handling is governed by their enterprise/API data terms — which, for API traffic, process your content to return a response rather than to improve a consumer-facing model.

We keep our supplier list deliberately short and use established providers under their data-protection terms. The services that may process your data, and the only reason each one does:

• Frontier-model providers— OpenAI, Anthropic, Google and xAI, reached through the FastRouter gateway. They receive the brief and work you submit only to return each council member’s response, and process it to answer rather than to train (see above).
• Supabase — stores your account, decisions, transcripts and uploaded files, in the European Union region.
• Vercel — hosts and serves the application, and provides privacy-friendly, cookieless usage analytics (no tracking cookies).
• Paystack — processes payments; it handles your card and billing details so we never do.
• Resend — delivers transactional email (receipts, sign-in codes and account notices) to your email address.

We do not sell or share your content with anyone else, and we use no third-party advertising or cross-site tracking networks.

Access to production data is restricted to the small number of people who operate the service, and only when needed to run or support it. Authentication protects your account, and your data is isolated from other users at the data layer.

Payments are processed by Paystack, a PCI-DSS Level 1 compliant payment provider. Your card details are entered with Paystack, not with us — full card numbers, CVV and PIN never touch Decidi’s servers. We receive only confirmation that a payment succeeded and the associated reference and amount.

You can delete individual decisions from your history at any time. To delete your whole account, your decisions and your uploaded files, email support@decidi.ai. We retain only the minimal billing records the law requires.

Your decisions and their transcripts are kept in your account so you can reopen and download them until you delete them — they are yours to keep or remove. Uploaded files are retained only as long as needed for the debate you attached them to and are removable on request. We hold the minimal billing records required by law, and nothing more than we need to run the service.

We’d rather be honest than impressive. Decidi is an early-stage product and does not currently hold formal security certifications — no SOC 2, no ISO 27001, no HIPAA. We will not imply compliance we don’t have. Formal certification and features like customer-managed retention and audit logging are on the roadmap as the product and customer base grow. If your use genuinely requires a specific certification today, tell us and we’ll be straight with you about where we are.

A council can include several different underlying models — real, live API models, not one model wearing different hats. They are accessed through the FastRouter gateway and assigned round-robin, so an N-mind debate genuinely spans different providers: OpenAI, Anthropic, Google and xAI. When four minds argue, you are typically hearing four different frontier models reason independently.

Each contribution shows which model produced it. Nothing is anonymised into a single voice — when GPT, Claude, Gemini or Grok speaks, the debate names it, so you can weigh a point knowing where it came from.

Personas are distinct expert system-prompts — a security red-teamer, a CFO, a devil’s advocate — layered onto a model to give it a deliberate point of view. The persona shapes the lens; the model does the reasoning. Both are shown together, so you always know that a given argument is, for example, the CFO lens running on one specific model.

The depth you choose selects the calibre of models in the pool:

• Quick — fast, lively, low-cost models for brainstorms and first passes.
• Standard — strong reasoning models for everyday rigour, the sensible default.
• Deep — the flagship frontier models at full depth, for decisions that genuinely matter.

Higher levels draw from more capable (and more expensive) models, which is why a Deep council costs more than a Quick one.

We do not invent a price. Cost is metered from the real token usage of the models that actually contributed, with a transparent markup applied, and it is shown live as the debate runs. You see an estimate before you commit, and the final charge reflects what was genuinely consumed — failed model calls are not charged.

If a model is briefly unavailable, the debate says so and continues with the rest of the council, rather than fabricating a response on that model’s behalf. A missing voice is shown as missing — we would rather give you an honest four-mind debate than a faked five-mind one.