# AI Overview

Version 1.2 · Last reviewed 2026-11-09 · Owner: Delora Legal (security@deloralegal.com)

Status: Public overview. Reflects the product as of the review date.

This page is for the attorneys and partners who decide whether their firm can use Delora. It says what the AI in Delora does, what it never does, what data it sees, and what is still being put in place. It describes the product as it is today, not as we hope it will be.

## The short version

- The AI is an intake assistant. It talks with your prospective clients, in English or Spanish, and writes a draft summary for your team.
- For your attorneys, admins and accredited representatives only, it also answers research questions from the public sources in Delora's research library and summarizes immigration decisions and agency notices. Identifiers are masked before it sees anything, and every sentence cites its source.
- Document upload is not live yet, so the AI reads no documents today, apart from the decisions and notices an attorney gives the research summarizer (below). In intake it sees only the text of the conversation. The plan for when upload goes live is under "When document upload goes live" below.
- It does not give legal advice, predict outcomes or recommend filings.
- Nothing it produces reaches a client or USCIS until an attorney at your firm has approved it.
- It runs on Anthropic's Claude models through Anthropic's API. Your firm's content is not used to train AI models, ours or our provider's.
- Anthropic's standard data-retention terms apply today. We are putting zero-data-retention terms in place and will say so here only once Anthropic confirms them in writing.

## What the AI does in Delora

1. **Intake assistant.** When a prospect opens your firm's intake link, they first read and agree to a notice in English or Spanish. The chat then greets them, says plainly that it is an AI and not a person, and asks for the facts an attorney needs: who they are, what help they want, their immigration history, relevant family, notices received, any court dates, deadlines or arrests, and how to reach them. It asks at most two questions at a time.
2. **Draft summary.** When the prospect presses "Send to the firm", a second AI step turns the conversation into a structured draft for your intake inbox: applicant details the prospect stated, the kind of matter they appear to be asking about (for routing only, not a judgment about eligibility), neutral red flags for an attorney to look at first, whether the matter looks urgent, and a short factual summary. Anything the prospect did not clearly say is left blank rather than guessed.
3. **Certificate reading: not live.** Document upload is not live, so the AI reads no certificates or other documents today. What is planned for when it goes live is set out under "When document upload goes live" below.
4. **Time entry descriptions.** When Delora drafts a time entry from work done in the app, the AI writes its one-line description from the case type, the time spent and the actions taken. SSNs are removed and A-Numbers masked before anything is sent. The description is labelled AI-assisted, and the entry stays a draft until an attorney or admin approves it. Time capture from calendars is coming soon and will work the same way.
5. **Research Q&A and summaries (attorneys, admins and accredited representatives only).** An attorney can ask a question of Delora's research library (the INA, 8 CFR, the USCIS Policy Manual, BIA and Attorney General precedents and adopted AAO decisions). Delora's own search finds the passages; the AI answers only from them, and every sentence must cite one or Delora removes it. An attorney can also have a decision or an agency notice summarized, with every sentence pinned to a page and paragraph. The question, the passages and the document's text are masked first: Social Security, A-, passport, travel document, USCIS account and card or account numbers are removed or replaced, and receipt numbers cut to their last four. The AI is given nothing else from the matter. Documents from VAWA, T and U cases are never sent. An uploaded PDF is not kept. Every answer and summary is labelled an AI-assisted draft until an attorney marks it reviewed; it never predicts outcomes, recommends filings or says whether a source still applies.

The AI does not fill USCIS forms. Forms are filled from the matter's case profile into the official PDFs by fixed field maps, and an attorney or admin approves each one.

## What the AI never does

- It does not give legal advice, assess eligibility, predict outcomes or timelines, quote fees, or recommend a form, filing or strategy. If asked, it says an attorney will advise after reviewing the information.
- It does not claim to be a person, an attorney or a member of your staff, and it says it is an AI if asked.
- It does not say or imply that the chat creates an attorney-client relationship or that your firm will take the case.
- It does not ask for a Social Security number. If a prospect types one anyway, it is removed before the message is stored or sent anywhere, and the assistant asks them not to share it.
- It takes no actions. It has no tools and no access to your firm's records or other firms' data. In intake it sees no documents, because document upload is not live; the research summarizer reads only the decision or notice an attorney picks. It produces text that a prospect or your team reads.
- It does not send anything to a client or to USCIS. Only people at your firm do that.

## Human review

Drafted work is a draft until your firm reviews it. The intake assistant itself talks with prospects in real time, so its replies are not reviewed one by one before the prospect sees them: it works under fixed rules, and a filter holds back replies that read like legal advice (your staff see what was held back). Your attorneys see the summary labelled as an AI-assisted draft that is not yet attorney-reviewed, with an instruction to verify it against the transcript before accepting. Accepting or declining the intake is a decision for an attorney or admin; paralegals can read intakes but cannot accept or decline them. Where AI content appears after review, it is labelled **AI-assisted, attorney-reviewed**.

When a prospect is detained, has a hearing within 30 days or a deadline within 14 days, the assistant tells them it is urgent and asks them to call your firm directly; fixed words such as "detained" or a court date soon also show a fixed urgent notice, and your admins and attorneys get an email when the intake is sent ("Urgent" when flagged). If someone says they are in danger, it tells them to call 911; if they mention suicide or self-harm, it gives the 988 Suicide & Crisis Lifeline, and a fixed 988/911 line is added to any fixed reply they see.

## Provider and model

Delora uses Anthropic, a United States company, as its AI provider. The assistant runs on Claude models accessed through Anthropic's API. Anthropic is the only AI provider and is listed, with what it receives, on our [Subprocessors page](https://deloralegal.com/subprocessors). The exact model version is set by configuration and changes only after the safety tests below are rerun.

## What data the model sees, and what it does not

**The model sees:**

- The text of the intake conversation, so it can write the next reply: what the prospect types, plus the instructions that tell it how to behave. The instructions are fixed text and do not contain your firm's data.
- The language the prospect chose.
- At the end, the same conversation as a transcript, to write the draft summary.
- For the research Q&A and summarizer (attorneys only): the masked question and the library passages found for it, or the masked text of one decision or notice.
- Whatever else a prospect chooses to type, such as their name, date of birth, country, immigration history, email or phone number, because that is what an intake collects.

**Before any text reaches the model:**

- Social Security numbers are removed. They are never stored.
- A-Numbers are replaced with a placeholder such as `[A-NUMBER-1]`. The real value is kept sealed by Delora and shown masked on screen.

**The model never sees:**

- The sensitive identifiers saved in case profiles: Social Security numbers, A-Numbers, passport and travel document numbers, and USCIS online account numbers. These are encrypted in our database and are never sent to any AI. An automated test runs the whole intake and form pipeline against a recording stand-in for the AI to check this.
- Any intake document. Document upload is not live, so no file is sent to the AI from intake today. The research summarizer sends the masked text of the one decision or notice an attorney picks, never the file itself. We use no file-storage or batch features of the provider's API.
- Your firm's clients, matters, billing or other firms' data. The research features send no matter field, not even the case type.

Our application logs do not contain personal data, and text that may quote what someone typed is redacted before it is logged.

## When document upload goes live

Document upload is not live yet. This section describes the design for when it is, and none of it applies today. Before upload opens, this page will be updated and the notice prospects agree to will be checked against it.

- A prospect can add photos or PDFs of their documents during the chat and say what each one is. Staff can also upload documents on a matter.
- A photo added as a birth or marriage certificate is first checked on our servers for anything that looks like an identification number or a passport, and is then sent to the model, which transcribes the names, dates, places and issuer printed on it into a draft. It is told not to record identification numbers. Nothing from it is used until a member of your staff confirms it. The file is sent with the request itself, under the same retention terms as the conversation.
- A certificate added as a PDF is never sent: staff enter its details by hand.
- A document added as a passport is read on our servers from its machine-readable zone and is never sent to the AI.
- Any other document is stored for your firm and not read. Documents your staff upload on a matter are not sent to the AI either.

## Training

Firm content is not used to train AI models, ours or our provider's. This is stated in our [Privacy Policy](https://deloralegal.com/privacy). For Anthropic, it is also the position under Anthropic's commercial terms, which say that API inputs are not used for training.

## Retention

Anthropic's standard API data-retention terms apply today. We have not yet obtained zero data retention from Anthropic. We are asking for it, and Anthropic approves it organization by organization. We are not saying it is in place. One feature we use, prompt caching, is one Anthropic lists as something that can affect zero-retention agreements, so we are asking Anthropic to confirm how it applies. We will update this page only when the terms are confirmed in writing.

On our side, an intake conversation that a prospect never sends to your firm is deleted after 30 days, and one your firm did not take on is deleted after 180 days. Accepted intakes are firm content and stay while your account is open. Details are in the [Privacy Policy](https://deloralegal.com/privacy) and the [Trust Center](https://deloralegal.com/trust).

## Safety testing

We test the intake assistant with 143 scripted conversations in English, Spanish and mixed language. They cover requests for legal advice, outcome predictions and filing recommendations; attempts to override or extract its instructions, including inside pasted documents; Social Security numbers and A-Numbers typed into the chat; people who are detained, frightened or in danger; and off-topic requests. Each conversation runs through the same code as the live product, against the real model. Fixed checks verify, for example, that an SSN is removed and never repeated, that A-Numbers are masked, that the reply is in the right language and that no advice or fee wording appears. A second Claude-based reviewer grades each conversation against written criteria. The conversations use synthetic people only.

The latest committed results (2026-10-27) show every conversation then in the suite passing (64 of 64). The earlier 63 of 64 baseline and its one failure, which broke no advice or identifier rule, are recorded in the AI governance policy. On 2026-10-28 we added a crisis-line instruction and three conversations; the committed results predate both, so they are being rerun. Our rule is to rerun the tests whenever we change the assistant's instructions or the model. These tests show how the assistant behaved on the scripts we wrote. They do not guarantee every future conversation, which is why attorney review is required.

## Labelling and the prospect's experience

Prospects are told at the start that they are talking with an AI. A note above the chat, shown for the whole conversation, says that the assistant is an AI and not a person, that an attorney reviews everything, and that it is not legal advice. The notice they agree to before the chat explains that your firm receives what they share and that an AI assistant processes it. The version they agreed to is stored with their session.

## What we have not done yet

- Zero data retention with Anthropic: requested, not confirmed.
- An independent security assessment of the application: planned, not yet done.
- A SOC 2 report: Delora does not hold one. A Type I report is planned.
- A scheduled nightly run of the safety tests: the tests run on demand today.

## Questions

For a security or AI questionnaire, or any question about this page, email [support@deloralegal.com](mailto:support@deloralegal.com). To report a vulnerability, email [security@deloralegal.com](mailto:security@deloralegal.com). Our policies are on the Trust Center, and the [Trust Center](https://deloralegal.com/trust) lists what is in place and what is still ahead.

This page is general information about the product and is not legal advice. Delora is software, not a law firm, and your attorneys remain responsible for their clients' matters.
