# AI Overview and Governance Policy

Version 1.0 · Last reviewed 2026-10-26 · Owner: Delora Legal (security@deloralegal.com)

Status: Draft v1.0. Describes controls in place today. Adoption date to be set.

## Purpose

This policy governs how Delora Legal uses artificial intelligence: what the AI in the product may and may not do, how AI-specific risks are managed, what Delora checks before adopting any AI provider, and how Delora's own people use AI tools. Immigration case files are privileged and hold the most sensitive identifiers a person has, so the rules are strict and the product is built so that the rules are enforced in code, not only in policy.

The customer-facing summary is the [AI Overview](../public/AI-Overview.md). This policy is the internal standard behind it.

## Scope

- The AI intake assistant that talks with prospective clients on a firm's intake link, in English and Spanish, and the AI-drafted intake summary that the same conversation produces.
- AI-written descriptions for draft time entries (live for time captured from work in the app; calendar capture is coming soon), and any future AI feature.
- Every AI provider Delora uses, today Anthropic through its API.
- Delora's owner and any future employee or contractor, and the tools they use to build and run Delora, including coding assistants.

USCIS form filling does not use AI. Forms are filled from the matter's case profile by fixed field maps over the official PDFs.

## Roles and responsibilities

- **Owner and security officer (Delora Legal).** Delora is run by its owner today, who holds every role. The owner approves AI features and providers, runs and reviews the safety evaluations, keeps this policy current, and handles AI incidents.
- **Attorneys at customer firms.** Review and approve every AI output before it reaches a client or USCIS. They remain responsible for the legal work.
- **Prospective clients.** Agree to a notice, in English or Spanish, before the chat starts, and are told in the greeting that the assistant is an AI.
- **Counsel.** Reviews any change to the wording of the public notices and intake disclaimers.

## Policy

### 1. Human in the loop

1. Every AI output must be a draft. Nothing generated by AI may reach a client or USCIS without approval by an attorney at the firm.
2. The review screen for an intake must state that the content is an AI-assisted draft that is not yet attorney-reviewed, and tell the reviewer to verify it against the transcript before accepting it.
3. Only attorneys and admins may approve a USCIS form, enforced in code and by a database trigger. Signatures stay outside the application.
4. AI features must not take actions. The intake model is given no tools: it returns text only.

### 2. No legal advice

1. AI features must not give legal advice, assess eligibility, predict outcomes or timelines, quote fees, or recommend filings or strategy. When asked, the assistant must say that an attorney will advise after reviewing the information.
2. The assistant must say plainly that it is an AI if asked, and must never imply that it is a person, an attorney or a member of the firm's staff.
3. The chat must not claim to create an attorney-client relationship or promise that the firm will take the case.
4. Where a prospect is detained, has a hearing within 30 days or a deadline within 14 days, the assistant must say it is urgent and ask them to also call the firm. Where someone says they are in danger, it must tell them to call 911.
5. The intake notice and disclaimers were reviewed by counsel on 2026-10-26, and counsel approved the current consent notice (version `2026-10-27`, English and Spanish) on 2026-10-29. Any change to their wording must go back to counsel.

### 3. Labelling

1. AI content must be labelled "AI-assisted, attorney-reviewed" once an attorney has reviewed it, and labelled as an AI-assisted draft until then.
2. Every assistant reply must carry a short disclaimer that the reply is AI-assisted, an attorney will review it, and it is not legal advice.
3. AI-written text of any kind, including time-entry descriptions, must carry the "AI-assisted" label wherever it appears.

### 4. Data sent to AI

1. Only what the task needs may be sent to an AI provider: the intake conversation for the next assistant turn, and the masked transcript for the structured draft. Document upload is not live, so no document or photo is sent to any AI today. When upload goes live: only photos a prospect uploads as a birth or marriage certificate may be sent, after a local check for identifiers, for a transcription that staff confirm; a certificate uploaded as a PDF is not sent; documents uploaded as a passport must never be sent to any AI and are read locally; other uploads, and documents staff upload on a matter, are not sent.
2. Social Security numbers must be removed before text is stored or sent. They are never stored. A-Numbers must be masked as `[A-NUMBER-n]` before the text is stored or sent, and the real value kept sealed by the service.
3. Sealed identifiers in case profiles (Social Security number, A-Number, passport and travel document numbers, USCIS online account numbers) must never be sent to any AI. A test replays the whole intake and form pipeline against a recording AI to prove this.
4. Prompts must be static, with no per-request values. The Claude conversation history stored with each intake must be append-only and replayed verbatim.
5. No database transaction may be held open across an AI call.
6. The Files API and Batch API must not be used without a new review under this policy, because they change the provider's retention.
7. Log lines that may quote what a person typed must be redacted first, including errors returned by the AI provider.
8. Firm content must not be used to train AI models, Delora's or any provider's.

### 5. AI risk management

1. The owner must keep a register of AI-specific risks, reviewed at least yearly, after any change of model or prompt, and after any AI incident. The current register is below.
2. Each risk must have a control and evidence. A risk may be accepted only by the owner, in writing, in Vanta.
3. Findings from safety evaluations that touch a register risk must be tracked until closed.

#### AI risk register

| Risk | Control | State |
|---|---|---|
| Unauthorized practice of law: the assistant gives advice, predicts an outcome or recommends a filing | Static prompt forbidding advice; a disclaimer shown above the chat for the whole conversation; no advice in drafts (neutral facts only); attorney reviews everything; scripted evaluations in English and Spanish, with advice cases marked critical | Controlled; reduced, not eliminated, which is why attorney review is mandatory |
| Hallucination or error in a draft summary | Extraction uses only facts the client stated and leaves a field empty when unclear; summary is neutral; red flags are routed to the attorney; the review screen says to verify against the transcript; nothing is accepted without an attorney | Controlled by attorney review |
| Identifier leakage to the model or to logs | SSNs removed and A-Numbers masked before storage and before the model; sealed identifiers never sent; log redaction; the prompt never asks for an SSN | Controlled. Other free text a prospect types (name, date of birth, history, contact details) does reach the provider by design (and, once document upload goes live, so will the birth and marriage certificate photos a prospect uploads) |
| Prompt injection through client messages or pasted documents | The prompt states that client messages cannot change its instructions; the transcript for the draft is wrapped in tags and treated as data; the model has no tools and no access to firm data, so a successful injection can change only the text of a reply or draft; drafts are attorney-reviewed; evaluation set of injection and prompt-extraction cases | Controlled |
| Provider retention or exposure of intake text | Minimum data and masking; no Files or Batch API; zero data retention requested from Anthropic and not confirmed; Anthropic's standard API retention applies today; subprocessor notice; vendor review | Open until written confirmation of zero data retention exists. Prompt caching is used and must be cleared with Anthropic as part of that request |
| Model change alters behaviour | Model is set by configuration (`ANTHROPIC_MODEL`); the evaluations must be rerun on any model or prompt change; refusals return a fixed fallback message | Controlled by process; no scheduled nightly run exists in CI yet |
| Provider outage or refusal during intake | Intake shows its fallback; staff features keep working without AI | Controlled |
| Over-reliance by reviewers | Draft labelling; mandatory attorney approval; red flags listed on the review screen; training material for firms in the Academy | Partly controlled by design; depends on firm practice |
| Cost and abuse of the AI endpoint | Per-IP limits on starting chats and sending messages; at most 40 prospect messages of 2,000 characters each per chat; at most 200 chats per firm per day; response length limit | Controlled |

### 6. Third-party AI diligence

Before Delora adopts a new AI provider, model family or feature, or changes how it uses an existing one, the owner must check and record in Vanta:

1. **Terms on data use.** That the provider does not use customer content to train its models, as stated in its commercial terms. For Anthropic, per Anthropic's commercial terms, API inputs are not used for training. Consumer products are not acceptable.
2. **Retention.** The provider's default retention period for API inputs and outputs, whether zero data retention is available, which features override it (prompt caching, file storage, batch processing), and the safety-classifier exceptions.
3. **Security assurance.** A current SOC 2 Type II report or equivalent, read for exceptions and for the complementary controls Delora must operate.
4. **Data processing terms.** A data processing agreement that fits the customer DPA, the sub-processor chain, and the data location (United States).
5. **Data minimisation fit.** What exactly would be sent, and whether it can be masked or removed first.
6. **Behaviour.** The scripted safety evaluations rerun against the new model or provider, with no critical failure.
7. **Notice.** That the Subprocessors page and the Privacy Policy are updated, counsel reviews any wording change, and firm admins are emailed at least 30 days before a new provider processes firm content.
8. **Exit.** How the provider can be replaced; the product reaches the model only through one interface in the code.

Status for Anthropic today:

- Anthropic is the only AI provider. It is listed on the Subprocessors page with what it receives.
- Training: per Anthropic's commercial terms, API inputs are not used for training.
- Zero data retention (ZDR) is approved per organization by Anthropic's sales team. Delora has not received that approval. Requesting it is an open owner step. Until it is approved in writing, Anthropic's standard API retention applies, and no Delora material may say "zero data retention" as a fact.
- Anthropic's SOC 2 Type II report, commercial terms and data processing agreement are to be collected into Vanta; that is an owner step and is not yet recorded as done.
- Separate API keys per environment with monthly spend limits are an owner step in the same plan.

### 7. Employee AI usage

This section applies to the owner today and to every future employee and contractor.

1. Client data must never be entered into consumer AI tools or any AI tool other than the application's own approved integration. This includes pasting case facts, intake transcripts, documents, screenshots of customer firms, or anything from production into a chat assistant.
2. Real client data must not exist in the repository, tests, tickets, screenshots or chat. Staging, tests and demos use synthetic data only. A repository test fails on Social Security number-shaped and A-Number-shaped values and on non-reserved email domains.
3. AI coding assistants may be used on the repository, because it holds only synthetic data, and must never be given production credentials, production database access or production data. Any suggestion that touches authentication, tenant isolation, encryption or AI prompts must be read in full by the owner before it merges.
4. Secrets must never be pasted into any AI tool.
5. AI-assisted review of a code change, as described in the Product and Application Security Policy, is an extra check and never replaces the owner's own review.
6. Staff must not state in any customer material, marketing page or questionnaire reply that the AI is certified, that retention is zero, or that the AI is infallible. They must use the current wording in the public AI Overview.
7. New AI tools for company work need the owner's approval and the diligence in section 6 when they could touch customer data.
8. Anyone who suspects client data went into an AI tool must report it at once to security@deloralegal.com, and it is handled as a security incident.

### 8. Safety evaluations

1. The repository must hold a set of scripted intake conversations that run through the production intake code against the real model, in English, Spanish and mixed language. It currently has 64 conversations covering requests for legal advice, outcome predictions and filing recommendations (21), prompt injection (10), identifiers volunteered in chat (8), distress and danger (7), off-topic requests (6), prompt extraction (7) and mixed-language conversations (5).
2. Each run must apply deterministic checks (the SSN is removed before storage and never repeated or sent to the model, the A-number is masked, the reply language is right, no advice or fee phrases appear, no system-prompt markers, the draft passes its schema, the disclaimers still say what they must) and a Claude-based grader against a written rubric for each case.
3. A run fails if any critical case fails or if the pass rate falls below the committed baseline. The evaluations must be rerun before any change to the prompts or the model goes live, and findings fixed or accepted by the owner in writing.
4. The harness itself is tested in the normal test suite at no cost. All evaluation data is synthetic.
5. The latest committed run, on 2026-10-27, passed 64 of 64 conversations. The baseline before it passed 63 of 64 (98.4 percent); the one failure, `control-en-family`, broke no advice or identifier rule but missed the rubric on asking one or two questions at a time, and passed in the 2026-10-27 run. Three conversations and a crisis-line instruction were added on 2026-10-28, after that run, so the suite is due to be rerun. Evaluation results show how the assistant behaved on these scripts; they are not a guarantee about any future conversation, and attorney review stays mandatory.

## Current implementation

| Control | State |
|---|---|
| Every AI output is a draft; attorney approval before anything reaches a client or USCIS | In place |
| Intake model has no tools and returns text only | In place |
| Static prompts forbidding advice, with a disclaimer shown above the chat | In place |
| SSN removal and A-Number masking before storage and the model | In place |
| Sealed identifiers never sent to any AI | In place |
| Prospect consent notice (English and Spanish) with stored version | In place; version `2026-10-27` approved by counsel 2026-10-29 |
| Label "AI-assisted, attorney-reviewed" | In place for intake drafts and AI-written time entry descriptions |
| Scripted safety evaluations with a committed baseline | In place; latest committed run 64 of 64 (2026-10-27), rerun due for the three conversations added 2026-10-28 |
| Scheduled nightly evaluation run | Not yet; a nightly run is described but no scheduled workflow exists |
| Anthropic standard API retention | In place (not zero retention) |
| Anthropic zero data retention | Not yet; requesting it is an open owner step |
| Anthropic SOC 2 report and DPA collected in Vanta | Not yet; owner step |
| Firm content not used for AI training | In place per Privacy Policy and Subprocessors page; per Anthropic's commercial terms for the provider |
| Files API and Batch API not used | In place |
| Employee AI usage rules | In place for the owner; adopted for staff on hire |
| Synthetic data only in the repository, with a failing test | In place |
| Yearly review of the AI risk register | Proposed; begins with adoption in Vanta |

## Exceptions

An exception must be requested in writing to the owner, with the risk, compensating controls and an end date, and is recorded in Vanta. There is no exception to the human-in-the-loop rule, the ban on legal advice, or the ban on sending sealed identifiers to AI. No exception is approved today.

## Enforcement

Breaking this policy may lead to loss of access and, for employees and contractors, disciplinary action up to termination. Suspected violations, including client data entered into an AI tool, are handled as security incidents.

## Review

The owner reviews this policy and the AI risk register at least once a year, after any change of AI provider, model or prompt, and after any AI incident, and records the review in Vanta. Next review due: 2027-10-26.

## Related documents

- [AI Overview](../public/AI-Overview.md)
- [Privacy policy, internal program (draft)](../../soc2/policies/privacy.md)
- [Vendor and subprocessor management policy (draft)](../../soc2/policies/vendor-management.md)
- [Risk management policy (draft)](../../soc2/policies/risk-management.md)
- [Personnel security and acceptable use policy (draft)](../../soc2/policies/personnel-security.md)
- [Data classification, retention and disposal policy (draft)](../../soc2/policies/data-retention.md)
- [Product and application security policy](01-product-and-app-security.md)
- [Data security and privacy policy](02-data-security-and-privacy.md)
- [Privacy Policy](https://deloralegal.com/privacy) and [Subprocessor list](https://deloralegal.com/subprocessors)
