What a Legal Intelligence Layer Should Do for an Immigration Firm

Most immigration software is good at storing a receipt number, form, deadline, document, and task. The next useful layer is not another place to chat with a model. It is software that notices when the record requires work, shows why, and carries the attorney’s decision back into the matter.

For a small firm, that can mean an RFE notice becomes a deadline and evidence matrix; a new visa bulletin creates a review queue rather than a client promise; a replacement form edition holds an outdated packet; or a client upload clears one requirement while exposing a different identity conflict.

The standard should be demanding. The system must be able to show its source, preserve uncertainty, and stay quiet when the trigger or authority is not reliable.

Begin with matter events the firm can verify

Useful intelligence starts from a defined event:

• a notice enters the matter;
• a case-status result changes;
• a filing or hearing date is added;
• a client submits a required document;
• a current form or source used by the workflow changes;
• two sources disagree on a material fact; or
• a required evidence row remains open at an attorney-set checkpoint.

The event should be visible. “AI found a risk” is not an adequate explanation. The attorney should see the notice page, status response, source document, form edition, or matter field that caused the review to appear.

Produce a review record, not an answer bubble

A useful review record contains:

This structure is more valuable than a long summary because it gives the firm something it can act on and later audit.

Keep official sources beside volatile work

Immigration work contains sources with different jobs and cadences. USCIS form pages control current form editions and instructions. The USCIS Policy Manual provides agency policy. The Department of State publishes the Visa Bulletin. USCIS separately identifies which chart may be used for adjustment filings. EOIR maintains its current Policy Manual, including filing and deadline guidance.

The technology should record which source was used, when it was verified, and which workflow depends on it. A general nationwide jurisdiction does not prove that a particular workflow is current or correctly wired. If the source, classification, implementation, or freshness check fails, the feature should hold rather than present an old result as available.

Give client-correctable work a closed loop

The strongest automation opportunity is often administrative. If the firm asked for the July statement and the client uploads June, the system can identify the date range, explain the mismatch, send a firm-approved correction, remind within the firm’s rule, and check the replacement.

The loop should stop when the issue becomes legal or factually ambiguous. A missing notice page can go back to the client. Whether the existing evidence satisfies the notice, how to explain a discrepancy, or whether a different filing strategy is appropriate remains with counsel.

The closing receipt should show the original requirement, source fact, issue caught, client message, replacement result, and any attorney exception. That is how the firm knows the automation finished the administrative job rather than creating another task.

Use the same discipline for generative work

The ABA’s Formal Opinion 512 discusses competence, confidentiality, communication, candor, supervisory duties, and fees when lawyers use generative AI. It does not prescribe one product architecture, but it gives firms a practical reason to require review, data-handling clarity, and supervision.

For a generated RFE outline, declaration draft, or client explanation, preserve:

• the matter sources made available to the system;
• the version generated;
• unsupported or conflicting propositions;
• attorney edits and approval;
• client-facing status; and
• the final version used or sent.

The NIST Generative AI Profile is also useful when building a vendor review. It frames generative-AI risk as something to govern, map, measure, and manage, which translates well into a small-firm pilot with named owners and observable failure conditions.

Know when the system should stay quiet

A mature product needs quiet rules. Do not surface a recommendation when:

• the complete notice or source set is missing;
• the matter identity is inconsistent;
• the official source is stale or unavailable;
• the practice or jurisdiction workflow has not been verified;
• the finding depends on a legal characterization counsel has not made;
• the confidence comes from a model score without inspectable evidence; or
• the same issue is already owned and no material event occurred.

Silence with a visible reason is better than a confident but ungrounded card.

Run a closed-matter pilot

Choose five closed or synthetic matters that represent ordinary work:

1. an RFE with several distinct requests;
2. a filing with a changed form edition;
3. a matter with inconsistent identity or date facts;
4. a case-status change that requires obtaining the actual notice; and
5. a client upload that is complete as a file but wrong for the requirement.

For each, score trigger accuracy, source visibility, false alarms, missed issues, client-request quality, attorney time, and whether the decision receipt is usable. Include a matter where the correct behavior is to hold.

Do not evaluate the product on the polish of its summary. Evaluate whether it reduces reconstruction and follow-up while leaving counsel in control of the legal work.

Where DocketBuddy fits

DocketBuddy calls this connected layer Case Intelligence within PracticeOS. It links matter events, source facts, proof gaps, exact client requests, rule freshness, attorney decisions, and a clean record. Ready handles client-fixable collection work within firm-approved rules. Coverage shows the provenance and freshness of supported jurisdictional workflows.

The product is not meant to make independent immigration judgments. Its job is to make the file’s next administrative action and next attorney decision easier to see, complete, and reconstruct.