Case study · Legal · New York
Intake that grades itself, without the AI deciding whether a case has merit.
A personal injury firm in New York City, practising across the five boroughs and Long Island. The firm is not named here at their preference. Everything described below was built and is running.
What this case study does not contain
There are no performance numbers on this page. The engagement has not yet reached its first formal measurement point and paid acquisition has not gone live, so any cost-per-case or lead-volume figure I printed today would be a projection dressed as a result. When there are real numbers and permission to publish them, they will appear here. In the meantime this is a description of what was built and why it is built that way, which is the part that is actually transferable.
The situation
The firm had a live website, a real caseload, and an intention to start advertising. What it did not have was a reliable path from an enquiry to a decision. Enquiries arrived by phone and by form, were triaged from memory, and left no record of why anyone had said yes or no.
Before any of the automation work, an audit of the existing site turned up several problems that would have quietly wasted the entire advertising budget.
Almost nowhere to convert
A single phone link per page, sitting near the very bottom of the document, and a homepage with no enquiry form at all. Paid traffic would have arrived and left.
Results claims with nothing behind them
Several prominent figures about past recoveries and case volume that could not be substantiated on request. In a jurisdiction with strict attorney advertising rules, that is a live risk, not a copywriting quibble. They came down before any spend started.
A repealed statute quoted as current law
A practice-area page described a serious-injury threshold using language that was no longer accurate. A prospective client could have relied on it.
The site itself was unreliable on a phone
A structured audit found a long list of layout defects across tablet and mobile, which traced back to about a dozen underlying causes once grouped properly.
None of that is glamorous, and all of it had to be fixed before automation was worth building. This is a fairly typical finding: the constraint is rarely the absence of AI.
The decision the rest depends on
The model extracts facts. Code decides what they mean.
A law firm cannot have software forming a view on whether someone has a case. That is a judgment reserved to lawyers, and a system that blurs it creates a problem far more expensive than the one it solves. So the grading engine is split in two, and the split is architectural rather than a policy written in a document.
The language model
Reads the message. Extracts facts. Stops.
What kind of incident, when it happened, where, whether there was medical treatment, who else was involved. Structured fields out of unstructured text. It is never asked whether the matter is good, strong, or worth taking.
Ordinary deterministic code
Applies the firm's own rules to those facts.
Limitation periods calculated from dates. Statutory thresholds checked. Conflicts matched against existing matters. Same facts in, same grade out, every single time, and a lawyer can read the rule and change it.
The practical consequence is that the firm can describe exactly what its software does and does not do, in a sentence, to a regulator or an insurer. The grade is a sorting mechanism over facts, not an opinion about a case. Everything else in the build sits on top of that distinction.
What was built
One intake path
Website enquiries and calls captured into a single pipeline, so nothing depends on who happened to pick up.
Response clocks with real teeth
Urgent matters carry a thirty minute clock, standard matters a day. The clock is visible, and an unmet one escalates rather than expiring quietly.
Escalation for the serious categories
Certain matters route immediately to a named human regardless of what any grade says. Some things should never wait in a queue.
Fail toward a person, always
If the model is unavailable or unsure, the enquiry goes to a human rather than being guessed at. An outage cannot silently drop work.
An append-only audit trail
Retained for ten years. What arrived, what was extracted, which rule applied, who acted, when. Written once and never edited.
Acquisition connected to outcomes
Paid search, profile listings and call tracking wired so that spend can eventually be judged against signed work rather than clicks.
The findability layer
Structured data, canonical fixes, practice-area content, and the plain-text files that AI answer engines read when someone asks them for a lawyer.
Compliance handled as scope
Attorney advertising requirements, business messaging registration, disclaimers, and sign-off from the firm's own counsel where it was warranted.
Reporting with an approval gate
Internal digest daily. Anything going to the client is drafted for review and published only on approval. Nothing goes out unreviewed.
Everything in the firm's name
Hosting, database, analytics and ad accounts all owned by the firm. If they replace me tomorrow they change the passwords and keep operating.
What carries over to other engagements
Fix the leaks before you turn on the tap
Sending paid traffic to a site with one phone link and no form converts budget into nothing. The unglamorous fixes came first, and they were the highest-return work in the project.
Put the compliance boundary in the architecture
A policy saying the AI will not decide merit is a promise. A system where the model physically cannot reach the decision is a guarantee. Only one of those survives a hard question.
An audit trail is much cheaper to build than to retrofit
It went in on day one. Adding one after the fact means reconstructing history that was never recorded, which usually means it cannot be done.
Client ownership is a feature, not a concession
Building on accounts in the firm's name removed the entire question of what happens if this does not work out, which made every other decision easier to agree.
Most of this started as a two week audit.
The conversion holes, the compliance problems and the build sequence all came out of mapping the workflow before writing any code.
Questions people ask
Which firm is this?
A New York personal injury firm, and that is as specific as this page gets. The agreed level of disclosure does not include the firm name, the attorneys, or the commercial terms. Those are the client's to disclose, not mine.
Why are there no results numbers?
Because there are none yet that would be honest. The engagement has not reached its first formal measurement point and paid acquisition has not gone live, so any cost-per-case or lead-volume figure printed today would be a projection dressed as a result. When there are real numbers and permission to publish them, they will appear here.
Does the AI decide which cases the firm takes?
No, and the system is built specifically so it cannot. The model reads an inbound message and extracts facts, then stops. Rules the firm owns decide what those facts mean, and a person makes the actual call. Anything uncertain routes to a human rather than being resolved automatically.
How does that stay inside the attorney advertising rules?
By keeping the machine away from anything that constitutes legal judgment. Grades describe facts, never the merit or value of a claim, every A-grade lead is reviewed by a person before contact, and the whole path is inbound only. Every step writes an audit record, so the firm can show what happened and when.
Would this work for a practice that is not a law firm?
The shape carries over: extract facts with a model, decide with rules the client owns, fail toward a person, and log everything. What does not carry over is the rule set, which is specific to how a given firm or practice actually decides. That is what the audit stage is for.