Effective date: August 27, 2026 Last updated: August 27, 2026
MAFICA is, by design, an AI-native product: nearly everything you see was produced or assisted by an AI model. This Policy tells you plainly which AI systems are involved, what data they see, what they are and are not allowed to do on your behalf, and how we designed around their known limitations. It supplements the Terms of Service and Privacy Policy.
You are interacting with artificial intelligence. Every response from the AI CEO, an AI Employee, or a Capability is AI-generated. We do not use AI to impersonate a human support agent.
MAFICA operates a central Privacy Firewall that blocks personal information, customer and employee data, internal company data, confidential and financial data, credentials, and your Company Memory from ever being sent to any external AI provider (Anthropic, OpenAI, Google, Voyage AI). This check runs immediately before every outbound call to an external AI provider, for every feature described in this Policy, and fails closed: if a piece of content's origin cannot be confirmed safe, it is blocked, not sent on a best-effort basis.
One direct, current consequence of this design: because the AI CEO, AI Employees, Multi-Agent Collaboration, Recommendation Engine, and Social Studio's caption drafting all normally work by grounding a request in your Company Memory, and Company Memory is never sent externally, these features do not currently return an AI-generated response — they surface a plain, provider-neutral message explaining that the request contains protected data, rather than fabricating an answer or silently sending your data anyway. Company Memory itself is unaffected (you can still record and browse facts, meetings, and KPIs); it is specifically the step of sending that memory to an external AI for grounding that is blocked. Section 9 (Current Feature Status) lists this precisely, feature by feature. We are building a private, non-external-AI processing architecture so these features can be restored without changing this Policy's core guarantee.
We classify every AI provider MAFICA's code integrates with into one of three states, so this table never implies a capability is live in production when it isn't:
Regardless of status, every row below is equally subject to Section 1's Privacy Firewall — a provider being "currently used" never means it receives your Company Memory or other protected data.
| Provider | Status | Role | What it sees |
|---|---|---|---|
| Anthropic (Claude models) | Currently used | Powers the AI CEO's conversational responses, AI Employee analysis, Multi-Agent Collaboration, Recommendation Engine, and Social Studio's caption drafting — for the requests the Privacy Firewall allows through (see Section 1); every one of these features' normal request also includes Company Memory, so in practice each currently receives a blocked-request response instead of an AI-generated one, pending Section 1's private-processing work | The text of your prompt/instruction, for a request the Privacy Firewall has confirmed contains no protected data — never Company Memory, customer data, or other confidential business information |
| Voyage AI | Currently used (integration), not currently invoked (blocked) | Generates vector embeddings for Company Memory search | Nothing today — creating and searching Company Memory is, by its nature, entirely about your own business data, so the Privacy Firewall blocks every real call site that would send it to Voyage. Company Memory add/search do not currently work via external embedding for this reason (Section 9) |
| Google (Gemini models) | Available when configured | Text-generation capability (an alternative to Anthropic), and image/video/narration generation for Marketing Capabilities | The text of your prompt/instruction for text generation, for requests the Privacy Firewall allows through; for image/video, only a creative brief and — separately — any reference image you explicitly attach, which the Firewall treats as upload data and blocks the same way |
| OpenAI (GPT models) | Available when configured | Text-generation capability, an alternative to Anthropic | The text of your prompt/instruction, for requests the Privacy Firewall allows through (same basis as Anthropic above) |
| Higgsfield | Not yet functional | Was designed to generate AI images and video for Marketing Capabilities; its integration is not presently wired into any live feature | N/A — not currently invoked |
We do not operate our own foundation models; we call these providers' APIs. Each provider processes data under its own API-usage terms, which (as of this Policy's effective date) do not permit them to train their general-purpose models on API customer data by default. If that changes, if a provider moves between the states above, or if we add or remove a provider, we will update this Policy and the Privacy Policy's sub-processor table together.
MAFICA's AI CEO and Research AI are designed to answer only from Company Memory's actual facts and KPIs, and to say so plainly when the data needed to answer a question doesn't exist yet, rather than estimating or inventing a number to sound helpful — and, per Section 1, to say so plainly rather than fabricate an answer when a request cannot be sent externally at all. This is a system-prompt-level and Privacy-Firewall-level constraint we treat as core to the product, not an incidental behavior. It reduces, but does not eliminate, the risk of an inaccurate AI response when a response is returned — see Section 6.
No AI component in MAFICA autonomously executes an action that affects your money, your data, or a connected third-party account. Every recommendation an AI Employee generates is queued in Decision Gate and requires your explicit approval before the corresponding Plugin executes it. Approved actions can, in most cases, be rolled back afterward. The only exception is an Automation or Workflow that you have specifically configured to run unattended on a schedule — in that case, you are accepting reduced human oversight for that specific, named automation only, not for the Service generally, and you can disable or delete that automation at any time.
Social Studio and Marketing Capabilities draft content — social captions, ad creative, Creative Briefs — that is never published or sent automatically. You review and manually trigger publication (for example, posting to Instagram or Threads). You are responsible for that content once you publish it, including its accuracy, its compliance with the receiving platform's own content rules, and applicable advertising and consumer-protection law. Per Section 9, caption drafting currently returns a blocked-request message rather than an AI-generated draft.
We do not use the business data you provide to MAFICA to train our own models, and our AI sub-processors' API terms (as of this Policy's effective date) similarly do not permit them to train their general-purpose models on your data by default. If this policy from any sub-processor changes in a way that would affect you, we will update this section.
MAFICA's AI CEO, AI Employees, and Social Studio operate across MAFICA's supported interface languages. Style and terminology conventions (for example, hashtag and tone conventions for social captions) are tuned per language, but the underlying reasoning is performed by the same Anthropic models regardless of interface language; translation quality for AI-generated prose is not separately certified per language and should be reviewed by a fluent speaker before external publication, particularly for Marketplace listings and social posts. This applies once a grounded or generated response is actually returned — see Section 9 for which features currently are.
Reflecting Section 1's Privacy Firewall as implemented today, not a target state:
| Feature | Status |
|---|---|
| AI CEO chat | Currently returns a blocked-request message, not an AI-generated reply — every request is grounded in Company Memory, which the Privacy Firewall blocks from external transmission |
| AI Employees (Research, Finance, Sales, Marketing, Integration) — direct analysis | Same as AI CEO chat — blocked |
| Multi-Agent Collaboration | Same as AI CEO chat — blocked |
| Recommendation Engine | Same as AI CEO chat — blocked |
| Company Memory (add a fact/meeting/KPI, or search by meaning) | Currently blocked — both operations use Voyage AI embeddings, an external AI call the Privacy Firewall does not permit for Company Memory content. Browsing and listing existing Company Memory entries is unaffected |
| Social Studio caption drafting | Currently returns a blocked-request message, not a drafted caption — grounded in Company Memory the same way as AI CEO chat |
| Creative Studio script/storyboard/video generation | Unaffected — these do not use Company Memory |
| Onboarding assistant (text) | Unaffected — does not use Company Memory |
| Onboarding assistant (screenshot analysis) | Currently blocked — a screenshot is treated as uploaded data, which the Privacy Firewall does not send externally |
We are building a private, non-external-AI processing architecture so the blocked rows above can be restored without changing Section 1's core guarantee that protected data is never sent externally. We will update this table as that work progresses.
Questions about how MAFICA uses AI can be sent to: support@mafica.jp