How law firms use AI for client intake and matter triage
Intake is where AI pays for itself first, because it is the one process in a firm that is simultaneously high volume, highly repeatable, and expensive to get wrong. It is also where the worst legal AI products live, so the useful question is not whether to automate it but what to demand before trusting any of it.
Why intake first
Every other candidate workflow in a firm has a scheduling problem. Document review arrives in bursts. Research depends on the matter. Intake is continuous, arrives in a predictable shape, and consumes staff time in direct proportion to volume regardless of whether the matters are any good.
It is also the process where the cost of a mistake compounds fastest. A matter that sits for three weeks before anyone notices the limitations period has nearly run is a different problem from a slow document review. One is inefficiency. The other is a claim.
The three jobs, which are genuinely different
1. Structured capture
A prospective client describes what happened in their own words. Somebody then has to turn that into fields: when it happened, where, what kind of injury, who the defendant is, whether insurance has been identified, what documentation exists. Until that translation happens, nothing downstream can act.
This is the least glamorous job and the highest return. A firm can have excellent judgment and still be slow if the facts arrive as a voicemail transcript. It is also the safest use of AI in the whole process, because extracting a date from a narrative involves no legal judgment at all.
2. Qualification
Grading a matter against consistent criteria so that attention flows to viability rather than to whoever called first. Done properly, this is defensible and useful. Done as an opaque score, it is a liability, because a firm cannot explain a decision it cannot inspect.
3. Routing
Getting the matter to the right practice group, in the right office, with the relevant facts attached. For multi-office or referral-network firms this is where the delay usually lives, and it is almost entirely a data problem rather than a judgment problem.
What a defensible qualification score looks like
The single most important property of an intake score is that a person can take it apart. Here is the structure we publish for ours, offered as a reference for what to require from any vendor.
| Dimension | Scale | Weight | What it measures |
|---|---|---|---|
| Liability strength | 1 to 10 | 30% | Evidence of duty and breach as described |
| Damages severity | 1 to 10 | 30% | Injury severity and documentation supporting causation |
| Bellwether potential | 1 to 10 | 20% | Suitability as a lead case in aggregate litigation |
| Viability | 0 to 1 | 20% | File completeness, explicitly not merits |
Weights are fixed and disclosed. Grade A requires an aggregate at or above 0.70 with liability at or above 6.0. Every number that produced a grade is visible on the record, so a partner reviewing a rejected matter can see exactly which dimension held it down.
Ask every vendor about boundary cases. A matter scoring 0.700 against a 0.70 threshold is an A by exactly nothing. Move any single dimension by one point and it is a B. A scoring system that does not surface that sensitivity is presenting a 1-to-10 human judgment as though it carried three decimal places of precision. We publish the sensitivity next to the grade for this reason.
Two properties matter as much as the weights. Viability measures file completeness rather than merits, so it never expresses a view on whether a claim should succeed. And the record states what was not evaluated. If no incident date was captured, the deadline screen did not run, and "no hard stop recorded" is a materially different claim from "no deadline problem found." Only the first is true, and a system that blurs them is telling a firm it checked something it did not.
The deadline screen, and where AI should stop
This is the highest-value output of intake triage and the place where careless products do real damage.
Identifying the governing limitations provision and its period is retrieval of public law, and AI does it well. Computing a specific date from a client's facts is a different act. It is only safe where the underlying rule has been verified by an attorney for that jurisdiction and that claim type.
Statutes of repose make this sharper. A repose period runs from a different triggering event, often the sale or completion of a product or improvement, and can bar a claim that is perfectly timely under the limitations statute. A product liability matter can pass the limitations screen and still be dead. A tool that surfaces only the limitations period has given a firm false comfort.
Our position is that unverified rules produce no date. Where a repose period has not been attorney-verified for a jurisdiction, the platform reports the period as a lead to check and explicitly declines to compute a deadline from it. That costs a feature. It is the right trade, because a wrong date written into a calendar is worse than an empty field: the empty field gets noticed.
Routing and the handoff into your system
A qualified matter that stays in the vendor's dashboard has not been triaged. It has been graded, in a second place someone has to check.
Routing should resolve on the firm's own territory and practice-area rules, narrowest first: city, then county, then state. The handoff should create the contact and matter records in the firm's system, attach the intake documentation, and write structured values into the firm's own field names. Caseworth integrates directly with Clio and with SmartAdvocate through CaseSync, and field mappings are configured centrally rather than by individual users, so credentials and data flow stay auditable.
Central configuration is a deliberate choice. Integration credentials are the most sensitive thing a firm hands a vendor, and self-serve credential entry spreads them across accounts nobody is tracking. Each firm's credentials are sealed under a data key belonging to that firm alone, stored in a table that denies all access by default.
What AI should not do at intake
The line is not subtle, and vendors cross it constantly.
- It should not decide whether to take the case. Conflicts, capacity, fee structure, and professional judgment about the merits are the firm's. A grade is an input.
- It should not tell the prospective client they have a claim. Applying law to an individual's facts and returning a conclusion is legal advice, and a non-lawyer platform doing it is practicing law. The boundary is covered in our UPL compliance guide.
- It should not produce a deadline it cannot source. See above. This is the one that becomes a malpractice claim.
- It should not route contact details around consent. If a prospective client has not agreed to be contacted, that field is empty by design, and the system should report it as consent withheld rather than as missing data. The remedies are different.
Measuring whether it worked
Run intake AI in observe-only mode first. Let it score and record without changing what actually happens to the matter, and compare its grades against the decisions your staff made. Two to four weeks of that tells you more than any vendor demo, and it gives you a baseline to defend the threshold you eventually enforce.
The metrics worth tracking: time from first contact to a qualified decision, percentage of matters with a complete field set, deadline screens that actually ran, and the disagreement rate between the score and your intake staff. That last one is the most informative. High disagreement means either the model is wrong or your criteria were never written down.
For the governance work that should precede any of this, see how law firms can implement AI safely. For the wider operating model, see what makes a firm AI-native.
Frequently asked questions
How does AI improve law firm client intake?
Structured capture of narrative facts, consistent qualification so attention follows viability rather than arrival order, and routing with the facts attached. The deadline screen is usually the highest-value single output.
Can AI decide whether a firm should take a case?
No. It can grade against disclosed criteria and flag disqualifying conditions. Acceptance involves conflicts, capacity, and professional judgment, and stays with the firm.
What should a firm ask about how a score is calculated?
Dimensions, weights, thresholds, boundary sensitivity, and what the score did not evaluate. An undisclosed score is not reviewable, and a decision made on it is not defensible.
Should AI calculate the statute of limitations?
It can identify the governing provision and period. It should compute a date only from attorney-verified rules, and it should treat statutes of repose separately because they run from a different event and can bar an otherwise timely claim.
How does intake connect to a case management system?
Through a direct integration that creates contact and matter records, attaches documentation, and writes into the firm's own field names. If anyone is retyping, it is not integrated.
This article is general information about legal technology operations. It is not legal advice, not an ethics opinion, and describes no outcome for any particular matter. Caseworth is not a law firm and does not provide legal advice.