AI & trust
Don’t trust it. Verify it.
The point isn't that the AI is always right — no one can promise that. The point is that you can check it: every claim carries its page, every action waits for your confirmation, and every approval goes on the record.
01Citations
Citations from the model itself.
Attest uses native page-level citations: the model marks which page of which document each claim came from as it writes — not a search bolted on afterward. Hover to see the passage; click to open the page.
Draft response — I-130, bona fide marriage
AI draft · pending review“The couple has maintained a joint residence since March 2019 and has commingled finances across four joint accounts .”
Hover a citation to preview the passage. In the app, clicking opens the source document at that page.
Anything tinted violet came from the model — on this site and in the product.
02Approval
Draft until approved. Everywhere.
Every AI output — a memo, a response outline, a task, a comment, a client message — is a draft or a staged proposal until an attorney confirms it. Clients never see unreviewed AI content, and the same gate applies to every surface: there is no side door where the AI acts on its own.
Review queue
RFE response outline — Chen, Wei
AI draftDraft addresses both asks; Ask 2 flagged as a gap pending the 2024 compensation letter.
Approved by S. Rivera · logged
03Deadlines
A deadline is docketed by you, not by the model.
Miss an RFE deadline and no brilliant draft can save the case — so this is the one place we made the AI slowest. An extracted deadline arrives as a suggestion pinned to the exact notice page it came from, and it reaches your calendar only after you confirm it against that page yourself.
RFE_Notice_Chen.pdf · page 1
…the evidence submitted is insufficient to establish that the beneficiary meets at least three of the regulatory criteria. You must submit the requested evidence within 87 days, no later than September 12, 2026, or the petition may be decided on the existing record.
Suggested deadlineFrom notice p. 1
Sep 12, 2026
RFE response — Chen, Wei · O-1A
Not on your calendar until you confirm it against the page.
04Isolation
Your firm's walls are load-bearing.
Every record carries your firm's id, and every query is scoped to it on the server. The id comes from your session — never from the model — so there is no prompt, however clever, that reaches another firm's files. No cross-firm search, no cross-firm analytics.
Scoped on every row
Every case, document, and message is stamped with your firm's id at write time.
Never from the model
The AI can't supply a firm id. The server takes it from your signed-in session, always.
No shared corpus
Your documents are never pooled, indexed across firms, or used to train models.
05The record
Everything on the record. Especially the AI.
An append-only activity log records every upload, every AI suggestion, and the human who approved it — entries can't be edited or deleted. If a filing is ever questioned, your record of human review is one page, not an email archaeology project.
Activity · Chen, Wei
- Uploaded RFE_Notice_Chen.pdfS. Rivera · 2:14 pm
- Generated an AI draft — RFE response outlineAssistant · 2:16 pm
- Approved an AI draftS. Rivera · 2:31 pm
- Confirmed an RFE deadline — Sep 12S. Rivera · 2:32 pm
Append-only — entries can’t be edited or deleted.
06The models
Your case files never train anything.
We run established third-party models under enterprise API terms that forbid training on customer data. Which model handles which task is an engineering decision we revisit as they improve — what never changes is the boundary around them: your documents are processed to answer the question in front of them, and nothing you upload is retained to make a model better.
Never training data
Prompts and documents go out under API terms that prohibit training, fine-tuning, or model improvement on customer content. Nothing from your cases is absorbed into a model — ours or anyone else's.
Processed, then done
Content is sent for the request that needs it and retained only as long as that request takes. There is no shared corpus, no cross-firm analysis, and no reuse of one firm's files to answer another's question.
Assistive by design
The models read, extract, draft, and cite. They don't file, don't send, and don't decide — every output waits behind an approval gate, and the professional judgment stays with the attorney.
The providers we route to are named in the Privacy Policy, where the subprocessor list belongs.
07AI usage
What the AI cost, and who spent it.
Every model call lands in a ledger: the member behind it, the feature that made it, the model, and the tokens it burned. A firm-wide daily ceiling sits on top of that, so "we've hit the limit" is followed by a screen rather than a support ticket. Work with no person behind it — a client's autofill in the portal, a scheduled workflow, an inbound email — is recorded as unattributed rather than dropped, so the breakdown adds up to its own total.
Everyone in the firm can open it. A member without firm-wide audit permission sees their own rows rather than a locked door.
AI usage · last 30 daysWhole firm
Assistant41%
Inbox — reading what arrives27%
Evidence analysis & RFE drafting19%
Unattributed — portal & scheduled runs13%
Illustrative figures. Every row traces back to a case, a member, or the run that had no person behind it.
08Hard limits
What the AI will not do.
Some of the most important engineering here is refusal. These aren't policies we ask the model to follow — they're gates enforced by the application around it.
- File anything with USCIS, or send anything to a client, without an attorney's approval
- Docket a deadline you haven't confirmed against the notice page
- Present an uncited claim as sourced — uncited text is flagged as unsourced
- Touch another firm's data, change a case's stage, or manage your team or billing
Questions