How the agent layer is bounded.

Which AI we use

Mallín uses Anthropic's Claude family of large language models, accessed via Anthropic's commercial API. We currently use Claude Opus 4.7 for high-judgment tasks (brief generation, stakeholder analysis) and Claude Haiku for faster lower-stakes inference (chat replies, theme generation).

Anthropic is listed as a subprocessor on our Subprocessor List.

What goes to the AI

When you use Mallín, the following may be sent to the AI model as part of a request:

• Deal substrate: account information, opportunity records, stakeholder names and titles, call transcripts you've loaded, notes you've written.
• Public-web research: press releases, SEC filings, LinkedIn profile data we've gathered about the companies and people in your deals.
• Your prompt: the question or instruction you give the chat surface.
• Cross-deal patterns within your tenant: notes you've tagged as cross-cutting wisdom are surfaced to the model on related future deals (your tenant only — never another customer's patterns).

What does NOT go to the AI

• Authentication credentials, MFA secrets, OAuth tokens, API keys.
• Payment information (we don't collect any; if we did it would never reach the model).
• Personal information from people not associated with your deals.
• Data from other Mallín customers — full stop. Cross-tenant context never crosses the model boundary.
• Data you've marked private or that lives in private notes in your CRM (where your CRM marks it as private).

Anthropic's data handling

Under Anthropic's standard commercial API terms, Anthropic does not use API inputs or outputs to train their models. API requests are retained by Anthropic for up to 30 days for trust and safety review, then deleted. Anthropic is a subprocessor — your data crosses to them when you use Mallín features that require AI inference.

For the authoritative source on Anthropic's commercial data handling, see Anthropic's Commercial Terms and their Privacy Policy.

The cognition contract

Every recommendation Mallín makes — every brief section, every stakeholder read, every suggested next move — must satisfy a five-element structure:

• Risk: what could break the deal, named clearly.
• Move: the prescriptive next action, not a list of considerations.
• Evidence: the quoted source the reasoning rests on (call transcript, public filing, LinkedIn profile, past rep note).
• Temporal proof: when the evidence was generated (artifact timestamp, not wall-clock guess).
• Decision: what the rep is being asked to approve or reject.

The model is constrained to produce this structure or nothing. We do not ship recommendations that don't have linked evidence — that's the anti-hallucination guarantee. If the model can't cite the source, it doesn't make the claim.

Human-approval gates

Not everything Mallín generates is auto-applied. Outputs are tiered by reversibility:

• Auto-write tier: notes, logged activities, last-contact timestamps. Safe to write directly to CRM; easy to undo if wrong.
• Approve-then-write tier: MEDDPICC field changes, role assignments, competition flags. Mallín proposes; the rep approves before anything writes.
• Never-auto tier: deal stage, amount, close date, forecast category. Mallín does not write these even with rep approval — these are the rep's explicit decision in their CRM. Mallín can suggest, never write.

Write-through to CRM

Mallín's outputs land in your CRM as the system of record. Notes you save in Mallín become CRM notes; drafted-and-sent emails become CRM activities; action items become CRM tasks. We do not maintain a parallel data plane where rep contributions live only in Mallín.

CRM permissions govern visibility. If your CRM has a private-note flag, Mallín passes it through. If a rep doesn't have access to a record in your CRM, they don't see it in Mallín either.

Pattern learning (within your tenant only)

When a rep tags a note as a "cross-deal pattern" in Mallín, that note is surfaced to the model on future similar deals within the same tenant. This is how Mallín learns each customer's operating style over time.

Pattern learning is strictly tenant-bounded:

• Patterns from Customer A never inform briefs for Customer B.
• We do not train shared models on patterns from any tenant.
• Patterns are stored alongside the deal note they came from; deleting the note deletes the pattern.

Provenance & audit trail

Every AI-generated artifact carries provenance metadata: which model produced it, when, against which substrate, with which prompt template version. The rep can see this in the cockpit; the audit log captures it for compliance review.

When the model regenerates an artifact, the prior version is retained — not overwritten. The rep can review the diff between versions to understand what changed and why.

What Mallín will not do

• Generate or send customer-facing communication without rep approval.
• Modify CRM stage, amount, close date, or forecast category — even with rep approval. Those are the rep's explicit decision in their CRM.
• Surface cross-tenant context. Your patterns inform your briefs; Customer A's never inform Customer B's.
• Train shared models on your data.
• Make claims without linked evidence (cognition contract).
• Quietly overwrite rep contributions. If the model would change something the rep wrote, the rep approves the change first.

Reporting AI errors

If Mallín produces a recommendation you believe is wrong, harmful, or out of scope, please report it: hello@mallin.io. We review every report and use it to improve our prompts, evaluation harness, and guardrails. Reports are not attributed back to the model provider.

Updates to this policy

We'll update this policy as Mallín's AI usage evolves (new models, new providers, new guardrails). The "Last updated" date at the top reflects the most recent change. Material changes are communicated by email in advance.