When the backlog is the bottleneck: applied AI for high-volume legal services

Problem — the work that never reaches the lawyer

Walk into any high-volume legal-services organization and the constraint is rarely the law. It's everything before the law: intake forms, eligibility screening, document collection, deadline tracking, translating a frightened person's story into a structured matter a lawyer can act on. In immigration and legal-aid work this is acute — a single intake can eat 45 minutes of an attorney's time, and the people who most need help are the ones least able to navigate a form-driven system. When intake is the bottleneck, adding lawyers doesn't clear the backlog; the backlog is upstream of the lawyers.

Meanwhile the funding ground is shifting under these organizations — federal legal-orientation dollars pulled back, replaced unevenly by county and philanthropic pilots. Leaders are being asked to serve more people with less certainty about next year's budget. "Hire more staff" is not the answer on offer. "Do more with the staff you have" is the only answer left.

Insight — the judgment is expensive, but most of the work isn't judgment

Here's the trap: because the stakes are so high (getting an asylum deadline wrong can end a case), organizations treat every step as if it requires a lawyer's judgment. It doesn't. The work splits cleanly into two engines:

• A rules engine — deterministic, checkable work: which form applies, what documents a filing requires, what the deadline is, whether an intake is even in scope. This is where a person is spending expensive hours on unexpensive decisions.
• A judgment engine — the irreducibly human calls: is this story credible, what's the theory of the case, is this the hill to fight on. Being wrong here is costly, so a human must stay in the loop.

Applied AI belongs on the first engine, not the second. A well-scoped system turns an unstructured intake conversation — in the client's own language — into a structured, pre-formatted case summary the attorney reviews rather than builds. Intake time drops from 45 minutes to 8. Nobody's judgment was automated; the expensive, repetitive setup work around the judgment was. The lawyer still decides. They just stop doing data entry to get there.

The reason this works is the same reason it's safe: you separate the rules from the judgment, you start where judgment is expensive and repetitive, and you keep a human on every consequential decision. Explainability isn't a nice-to-have here — a system that can show why it flagged a deadline or routed a matter is a system a supervising attorney can actually trust and sign off on.

Path — what to do in two days

You don't need a year-long transformation or a data-science team. You need to find the one place where expensive judgment is being spent on unexpensive, repetitive work, and prove a narrow fix.

1. Map the intake-to-matter path. Every step from first contact to "a lawyer can act on this." Mark each step rules or judgment. The rules steps eating the most attorney minutes are your target.
2. Pick one high-volume, high-repetition step — usually intake structuring or document/eligibility triage — and design a human-in-the-loop assist for it: AI drafts, a person reviews and approves. Never AI decides-and-files.
3. Instrument the before. Minutes per intake, backlog age, matters-per-staff-week. You can't prove you helped without the baseline.
4. Build the thinnest version that touches real cases and put a supervising attorney's approval in the loop from day one.

That's a two-day workshop, not a moonshot. You leave with a scoped build, a baseline, and a human-review gate — not a slide deck about AI.

If this is your situation, spend two days with us. We call it a Foundation Sprint.