Contract & offer extraction — when you upload a Residential Purchase Agreement (executed contract or an inbound offer), Claude reads the PDF and extracts ~25 structured fields: property address, parties, price, financing type, contingencies, earnest money, closing date, possession, escalation, special terms. You review every value with confidence scores and tap to apply.

Task checklist generation — based on the extracted contract + your transaction template, ClosingDay generates a state-specific ~50-task contract-to-close checklist with deadlines, dependencies, and responsible-party assignments. Tasks that don't apply to your deal (e.g., HOA tasks on a non-HOA property) are filtered out automatically.

Document compliance — Claude infers seller attributes (foreign person, 1031 exchange, trust, corporate seller) from the contract; our rules engine then determines which state-specific compliance documents are required (e.g., in California: TDS, NHD, Lead Paint Disclosure, AB 38 fire-hardening, HOA package, FIRPTA — each state has its own hand-built rule set) with legal citations and deadlines.

Draft email generation — for each upcoming task, ClosingDay drafts the follow-up email in your voice (per your trained tone profile), addressed to the right contact, scheduled at 5/2/1/0-day thresholds before each deadline. Drafts wait in your inbox for review.

Tone profile analysis — when you paste 2-5 of your past emails into Settings → Tone, Claude analyzes them and produces a short tone profile (formality, sentence length, greetings/sign-offs, personality) used as a system prompt for future drafts.

Inspection-summary drafts — when you commit inspection findings as repair tasks on a listing, Claude drafts a summary email to the seller grouping findings by severity.

Scout (in-app assistant) — natural-language commands ("show me overdue tasks", "mark inspection complete") routed through Claude with tool-use to query and update your workspace.

Offer comparison narratives — when you compare offers side-by-side, Claude generates a brief written summary of the trade-offs.

Inspection finding extraction — Claude reads pre-listing inspection PDFs and extracts individual findings with severity rankings, which you can then turn into repair tasks.

All AI features are powered by Anthropic's Claude API. We do not use OpenAI, Google, Meta, or any other LLM provider for any customer data.

We use specific pinned model versions per feature: Claude Opus 4.7 for deep reasoning (Scout, complex extractions), Claude Sonnet 4.6 for structured tool-use (most contract / offer extractions), and Claude Haiku 4.5 for fast / cheap tasks (draft generation, tone analysis). Model selection is recorded in the audit log for every invocation.

When Anthropic deprecates a model, we test the successor on our golden-set evaluations before switching production traffic — a model upgrade is a deliberate ClosingDay deploy, not an automatic Anthropic rollout. We do not advance to a new model without an internal regression check.

Anthropic's own trust posture is published at trust.anthropic.com. Their API operates under SOC 2 Type 2 and ISO 27001 attestations, and their default API tier is configured for "no training on API data." Both certifications and the no-training default are independently verifiable from their public documentation.

Anthropic does not retain or train on data submitted via their API by default. This is Anthropic's standard API policy, not a ClosingDay-specific arrangement — it applies to every customer of theirs at the API tier.

ClosingDay does not opt into any optional data-sharing program with Anthropic. We have not, and will not without notice, enroll the account in any "feedback for training" or "model improvement" arrangement that would let your data influence future models.

We do not train our own models. ClosingDay does not run any in-house LLM, does not have a fine-tuned model derived from your data, and does not produce embeddings from your data for any retrieval-augmented system that persists beyond a single request.

Your contracts, emails, and communications are not used to train any AI model — ours or our providers'.

If Anthropic's public no-training default ever changes, we will notify all active accounts in advance and document the change here in this policy. We treat the no-training default as a material contractual expectation.

Contract PDFs — the executed Residential Purchase Agreement or inbound offer PDF you upload, sent as a base64-encoded document attachment to the Claude API for extraction.

Contract metadata after extraction — when you ask Scout to act on a transaction (e.g., "list overdue tasks"), the relevant transaction record (property address, parties, dates, current task statuses) is included in the prompt context.

Email bodies you paste into tone training — the 2-5 sample emails you paste into Settings → Tone are sent to Claude with a prompt asking for a tone description. The samples themselves are not stored beyond the tone-profile generation step.

Task context for draft generation — the task title, deadline, transaction property address, your tone profile, and the relevant contact name + role are included in the prompt for each follow-up draft.

Inspection PDFs — pre-listing inspection reports you upload are sent to Claude for finding extraction.

Scout queries — your natural-language question or command, plus the workspace context (transactions, tasks, drafts) needed to answer it. Anthropic's tool-use feature lets Claude request specific records only when needed; the entire workspace is not bulk-uploaded on every query.

Your password — bcrypt-hashed in our database; never exposed to any system, AI or otherwise.

Your Stripe payment details — held by Stripe, never crossing our backend boundary, never sent to Claude.

Your Gmail OAuth refresh token / SMTP password — stored encrypted; never sent to Claude.

Other agents' data — cross-tenant data is filtered at the database layer before any prompt is assembled. The visibility rules that gate UI access also gate AI context.

PRIVATE notes on offers — when an agent marks a comparison note as PRIVATE, that note never appears in the seller-portal renderer and never appears in any AI prompt assembled for sharing.

Documents marked NOT visible to seller — the documents-tab visibility filter applies to the AI prompt context the same way it applies to the seller portal.

PII from buyer-name-obscured listings — when a seller has the "obscure buyer names" toggle on, the obscuring applies to AI-generated comparison summaries shown to that seller as well.

Other accounts' data, ever — we do not cross-blend prompt context across agents. Even when Scout is asked a broad question, the assembled context is scoped to the asking agent's visibility rules.

At Anthropic: data submitted via the API is processed in-flight and not retained for training. Anthropic's own retention policy for API logs is published in their trust documentation (currently up to 30 days for abuse-monitoring purposes, with abuse-only access). Refer to their privacy documentation for the precise retention window.

At ClosingDay: the AI outputs (extracted contract fields, generated drafts, tone profiles, Scout responses) are stored in our database alongside the transaction or listing they belong to, for the lifetime of your account.

The raw input prompts to Claude are NOT persisted by ClosingDay — only the structured output is stored. (Contract / offer extractions persist `Extraction.extractionRaw` which is the structured Claude response, not the prompt that produced it.) If you delete your account or a specific transaction, the associated extractions, drafts, and tone-profile data are deleted with the parent record.

Extraction schema versions are tracked (`EXTRACTION_SCHEMA_VERSION` and `OFFER_EXTRACTION_SCHEMA_VERSION` in our codebase) so when we update the extraction prompts, cached results re-process on the next upload. This is an internal cache-busting mechanism — no impact on your data retention.

Every meaningful AI invocation is recorded in our audit log alongside the action it produced: the agent who triggered it, the feature (e.g., DRAFT_GENERATED, CONTRACT_EXTRACTED, ASSISTANT_QUERY), the model used, the timestamp, and the entity it operated on. The log is the same one that records non-AI actions like task status changes and document uploads.

Brokers can request a full audit-log export for any agent in their roster through the broker dashboard; the export includes every AI invocation. This is how we make it possible for a broker compliance officer to answer "what has the AI been used for on this transaction" in a defensible, time-ordered way.

AI calls that fail (Claude rate-limited, model deprecated, network error) are recorded too. Failure paths are visible in the dashboard so a broker can confirm whether a missing draft is "AI was supposed to draft this but errored" vs "AI was never asked to."

California Information Privacy Act and federal GLBA define non-public personal information (NPI) — financial account numbers, social security numbers, driver's license numbers, etc.

NPI you upload (e.g., a contract that contains a buyer's SSN in the signature block) IS sent to Claude as part of the document extraction call — we cannot pre-process a PDF to redact NPI before extraction since detecting it requires reading the document, and reading the document is what Claude is doing.

Anthropic's no-training-default policy applies to this NPI the same as to any other content in the call.

NPI is never used by ClosingDay for advertising or model-training purposes. Any redaction we apply downstream of extraction (e.g., not displaying full SSNs in the UI) is a separate UI policy, not an AI policy claim.

If you need a contractually-stronger NPI guarantee than Anthropic's default policy provides (e.g., an enterprise-tier zero-retention API contract directly with Anthropic), email hello@getclosingday.ai. We can help facilitate a direct contract for enterprise customers with that requirement.

You can clear your tone profile from Settings → Tone at any time — the stored profile description is deleted and future drafts revert to the default friendly-professional tone.

You can dismiss any specific AI draft without sending it. Dismissed drafts don't generate replacements at the same deadline threshold.

You can edit any AI-extracted contract field before saving. The audit log records both the AI's original value and your edit, so the broker can see the AI's accuracy on this specific extraction.

You can disable AI-drafted emails entirely by not approving any of them. By default, nothing ships without your approval — the Pending Review queue holds every draft until you approve it. Auto-send is a separate opt-in you enable and control.

You can request a full export of every AI-generated artifact on your account (drafts, extractions, tone profile, Scout transcripts) by emailing privacy@getclosingday.ai. We turn these around in 5 business days.

You can request deletion of your account and all AI-generated content via the same email. Once deleted, the data is purged from our database within 30 days and from Anthropic's in-flight logs per their own retention policy.

If we add a new AI feature, switch to a new AI provider, or change the data we send to AI, we will update this page and email all active accounts at least 14 days before the change takes effect.

Material changes (e.g., enrolling in any training program, switching providers) require explicit re-consent at next sign-in — you will not be silently moved onto a different AI provider.

Non-material changes (model version bumps within the same provider, minor prompt edits) are documented in our changelog.

The "last updated" date below is updated for every meaningful revision to this page.