Answer · Legal
Can a law firm use AI to run a conflicts check?
As a widener of the search, yes. As the system that decides there is no conflict, no: the failure is silent.
Use it to widen the search, never to close it. A probabilistic matcher that misses a party returns a clean result, and a clean result is indistinguishable from a correct one until the disqualification motion arrives.
A conflicts check is a completeness question, and completeness is the property probabilistic systems are worst at demonstrating. The check asks whether any name in a proposed engagement touches any name in the firm's history, and the answer that matters is the one that says no. A deterministic index that over-returns produces a pile of near-matches somebody has to clear, which is tedious and safe. A matcher that scores similarity produces a short, clean list, which is pleasant and carries no evidence that the omitted entries were properly omitted.
The useful work is upstream of the decision, and there is a lot of it. Conflicts fail on the party list far more often than on the matching: the subsidiary nobody named, the parent that acquired the counterparty last year, the individual who signs as one entity and is sued as another, the trading name that appears on the invoice and nowhere in the file. Reading an engagement pack, a contract bundle or a corporate filing to extract every entity mentioned is exactly the task these tools are good at, and every name it adds is a name the deterministic index then searches properly.
Name variance is the second job worth giving it. Transliterations, married and maiden names, abbreviations, punctuation, and the difference between a company's legal name and the one everybody uses are the reason a literal index misses real conflicts. Generating the variant set and searching all of them is a widening operation: it can only add hits, so its errors cost review time rather than exposure.
The data going into the tool is itself the constraint most firms notice late. A conflicts check runs on prospective client information — names, adverse parties, the shape of the dispute — and information from a prospective client is protected whether or not the firm takes the matter. That makes the conflicts system one of the more sensitive things a firm operates, and connecting a general-purpose tool to it is a confidentiality decision before it is a workflow decision.
The construction that works keeps the register deterministic and puts the model on both sides of it. Extract entities, expand variants, search the index literally, and then let the model summarise and cluster what came back so a person can clear it faster. Nothing the model produces removes an entry from the result set, and no result set is closed by anything other than a person looking at it. That arrangement gets the speed without moving the decision.
What a firm should be able to show afterwards is the search, not the conclusion. If a conflict later emerges, the defensible position is a record of which names were searched, against which index, on what date — and the useful contribution of an AI layer is that the list of names searched is longer than it would have been. A record that says a tool found nothing is not a record of a search.
Every other AI failure in a firm announces itself eventually; a missed conflict returns exactly what a clean file returns, which is nothing at all.
Siddharth Sharma, Context Theory
Related questions
What about using it to review the results rather than produce them?
That is the safer half and it is genuinely valuable. Clearing a long list of near-matches is slow, repetitive work where the model's job is to group, summarise and explain why each hit is or is not the same party. It still cannot be the thing that removes an entry, but it can be the thing that makes a person's decision on each entry take seconds instead of minutes.
Does this change for a small firm with no conflicts database?
It gets more important, not less. A firm whose history lives in matter folders and email has no index to search literally, and a model reading those folders is the only practical way to search them at all. The correct response is to treat what it returns as a set of leads to confirm by opening the file, and to use the exercise to build the register the firm did not have.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Attorneys & legal cost per lead | $131.63 | Category-wide |
| Attorneys & legal services CPC | $9.87 | Category-wide |
LocaliQ / WordStream Search Advertising Benchmarks 2026 · Google + Microsoft Ads, 20 industries · Apr 2025–Mar 2026 · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | A conflicts check is answered by the absence of a match, so a system whose recall cannot be demonstrated returns a clean result that is indistinguishable from a correct one until an adverse party raises it, which is after the engagement has begun. | Re-running a sample of historically cleared engagements through the proposed matcher and checking whether every known related party is returned. |
| Constraint | Information learned from a prospective client is protected whether or not the firm goes on to act, which makes the conflicts intake one of the most sensitive datasets a firm holds and makes any connection to a general-purpose tool a confidentiality decision first. | Model Rule 1.18 on duties to prospective clients, read against the data-handling terms of the tool the intake would flow through. |
| Software | Conflicts most often fail on the completeness of the party list rather than on the matching, because subsidiaries, recent acquisitions, trading names and individuals appearing under different capacities are frequently absent from the names the firm searched at all. | Comparing the entities named on an engagement form against those named anywhere in the underlying contract bundle or corporate filings. |
| Procurement | An arrangement in which the model expands the search and a deterministic index performs it keeps every model error on the side of extra review, since a widening operation can add candidate matches but never remove one from the result set. | Auditing whether any code path lets a model score, rank or filter entries out of the conflicts result before a person sees them. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one