An agent reads the intake notes and documents, lists every party and name variant that needs searching, and summarizes the search results for the responsible attorney, who makes the conflict decision.
Every new matter has to clear a conflict check before the firm accepts it, and in a small firm the check is only as good as the list of names someone thought to search. The ethics rules it serves — ABA Model Rules 1.7 on current clients, 1.9 on former clients, 1.10 on imputation across the firm, and 1.18 on duties to prospective clients (the intake material itself is prospective-client information), as adopted in each state — care about adverse parties, related entities, and people who are not on the intake form: the opposing party's parent company, a spouse, a former business name, the officer who signed the contract. Practice management software does the searching well. Clio, for example, runs conflict checks against the firm's existing records, lets you add several searches to one check for adverse parties, and lets the firm include calendar events, notes, document names, and communication logs in what is searched. What the software cannot do is decide which names to search. That is a reading job — intake call notes, an emailed demand letter, a contract, a complaint — and it is where checks fail: a subsidiary nobody noticed, a maiden name, "ABC Holdings" versus "ABC Holding Co." A language model is good at exactly that extraction, and it is the one step this workflow automates. The ethics constraint shapes the design: ABA Formal Opinion 512 (July 29, 2024) says a client's informed consent is required before inputting information relating to the representation into a self-learning generative AI tool, expects lawyers to assess a tool's Rule 1.6 confidentiality risk even where consent is not triggered, requires firms to set policies and train staff on AI use, and warns against relying on output without appropriate independent verification. So the model runs under commercial terms that do not train on firm data, the firm documents its risk assessment of the tool, the model only produces a list and a summary, and the attorney decides.
Before any prospective-client information goes to a model, decide and write down which tool is approved, under what terms, and who may use it. Use the Claude API (or another model) under a commercial agreement you have read: confirm in the contract that inputs are not used for training and what the retention period is, keep a copy, and write down the firm's assessment of the confidentiality risk of sending information relating to the representation to that tool. Formal Opinion 512 places the duty to set this policy on managerial lawyers. If you cannot confirm the terms, stop here — run the extraction checklist by hand instead.
Use the intake form or notes in Clio (Clio Grow if you use it for intake) plus any documents the prospect sent: demand letters, contracts, prior pleadings, emails. The fuller the material, the better the list. For a corporate party, add its public filings: for a public company, the subsidiaries exhibit (Exhibit 21) to its SEC annual report, which lists significant subsidiaries — smaller ones may be omitted, so it is a starting point, not a complete affiliate list; for a private company, the state business registry entry (registered agent, and officers where the state publishes them).
An n8n flow sends the material to the Claude API with the extraction prompt in the assets. It returns every person and entity mentioned, the role of each (prospective client, adverse party, related entity, witness, counsel, insurer, guarantor, family member), and the name variants worth searching: former names, d/b/a names, abbreviations, punctuation and suffix variants (Inc., LLC, Co.), and nicknames. Each name carries the sentence it came from, so the reviewer can see why it is there.
A paralegal or the intake attorney reviews the list — adding names the model missed, deleting irrelevant ones — and runs the conflict check in Clio with one search per name using Add another search. By default only firm administrators and users with the General Access role can run conflict checks and change their search settings, so grant that permission to whoever runs intake if needed. Configure the conflict check search settings once to include notes, document names, communication logs, and calendar events. The model does not run or interpret the search; the firm's system of record does.
Paste or export the check results back to the agent with the party list. It produces a one-page summary: for every hit, the name searched, the matching contact or matter, open or closed, the firm's role in it, and the attorney responsible — grouped into likely irrelevant (common name, different person), needs review, and potential conflict. The summary never says a matter is clear. It says what was searched and what was found.
The responsible attorney reviews the summary and any underlying matters, decides whether there is a conflict, whether it is consentable, and whether waivers are needed, and records the decision. Record the check, its date, and the decision on the matter in Clio so they live with the file. Save the agent's party list alongside it; it documents what you thought to search.
Add a trigger: when a new party is added to an open matter — an amended complaint, a new defendant, an intervening entity — run steps 3 to 6 for the new names. Conflicts that arise mid-matter are the ones small firms miss, because the intake check was the only one.
Use these templates as-is or customize for your business.
You are preparing a conflict-check search list for a law firm. You do not decide whether a conflict exists. Extract every person and organization mentioned in the materials.
For each, return:
{
"name": "as written",
"type": "person | organization",
"role": "prospective_client | adverse_party | related_entity_of_client | related_entity_of_adverse | officer_or_owner | family_member | witness | opposing_counsel | insurer | guarantor | other",
"related_to": "name of the party this one is connected to, or none",
"source_quote": "the exact sentence where the name appears",
"search_variants": ["variants worth searching"]
}
VARIANT RULES:
- Organizations: with and without suffix (Inc, Inc., Incorporated, LLC, L.L.C., Co., Corp.), with and without 'The', ampersand vs 'and', any d/b/a or former name stated in the materials, and the parent or subsidiary names stated in the materials.
- People: full name, first initial + last name, any former or maiden name stated, common nickname of the first name.
- Do not invent relationships. Only list a parent, subsidiary, spouse, or former name if the materials state it.
After the list, add: "names_mentioned_but_unclear": any partial names or references like 'the landlord' or 'his business partner' that need a follow-up question to the prospect.Using the party list and the conflict check results below, write a summary for the responsible attorney. Do not state or imply that the matter is clear of conflicts. Do not give a conclusion on whether any conflict is waivable. FORMAT: Matter: [prospective matter description] Searched: [count] names, [count] variants, on [date] by [name] POTENTIAL CONFLICT (attorney review required): - Searched '[name]' (role: [role]) -> matches [contact/matter], [open/closed], firm represented [whom], responsible attorney [name]. Why it matters: [one line]. NEEDS REVIEW (could be the same person or entity): - ... LIKELY UNRELATED (common name, different identifiers): - ... (state the identifier that differs, e.g. different city or date of birth) NO HITS: [list of names with no results] OPEN QUESTIONS FOR THE PROSPECT: [from names_mentioned_but_unclear] Decision: ______ Waiver needed: Y/N Decided by: ______ Date: ______
1. APPROVED TOOL. Conflict-prep extraction and summaries use [Claude API via firm automation] only. Consumer chat apps are not approved for client or prospective-client information. 2. TERMS ON FILE. Our agreement with the provider states inputs are not used for training; retention is [period]. Copy stored at [location]. Reviewed by [name] on [date]. Our Rule 1.6 risk assessment of the tool is stored at [location]. 3. WHAT THE TOOL DOES. Lists names and variants from intake material; summarizes search results. It does not run searches, decide conflicts, or recommend waivers. 4. VERIFICATION. A person reviews every list before searching and every summary before the attorney relies on it. The attorney reads the underlying matters for every POTENTIAL CONFLICT and NEEDS REVIEW item. 5. CLIENT DISCLOSURE. If our terms or tool change such that information relating to the representation could be used to train a model, we obtain informed consent before use, with an explanation of the risk (ABA Formal Opinion 512) — not a boilerplate clause. Confirm your state bar's guidance as well. 6. TRAINING. Everyone who handles intake reads this note and completes [training] before use. 7. RECORDS. The extraction list, the check, and the decision are linked to the matter. Signed: __________ (managing partner) Date: ________
Get a new AI workflow every week. Prompts, tool stacks, and ROI math included.
AI does the categorization or first-draft work, a human approves before action is taken. The pattern of choice for anything irreversible, externally visible, or financially sensitive.
Learn the agentic glossary →Where this workflow tends to break in production — and what to put in place before you ship it.
A related entity or former name is never searched
Mitigation: Extraction prompt requires variants and relationships stated in materials; reviewer adds known affiliates; unclear references become follow-up questions to the prospect.
Model invents a relationship and staff search a wrong name, or trust it
Mitigation: Every name carries its source quote; prompt forbids unstated relationships; reviewer deletes anything without a source.
Summary is read as clearance
Mitigation: Summary template never states the matter is clear; attorney records the decision in a separate field on the matter.
Confidential intake data sent to an unapproved or training-enabled tool
Mitigation: Single approved tool under reviewed terms; consumer apps barred in the signed AI-use note; terms re-checked when the provider changes them.
New parties added mid-matter are never checked
Mitigation: Trigger on party additions to open matters re-runs the workflow for the new names.
Do not use this as a conflict decision tool — it builds the search list and organizes the results; the decision, any waiver analysis, and any client communication about it belong to the responsible lawyer under your state's version of the rules. Skip it if your firm cannot confirm, in a contract it has read, that the model provider does not train on your inputs; in that case run the extraction checklist by hand, because Formal Opinion 512 makes informed client consent the price of using a self-learning tool with information relating to the representation. Skip it if your conflicts database is incomplete — a perfect search list run against a system missing the firm's pre-migration clients finds nothing, so fix the data first. And check your state bar's own guidance on generative AI, which may be stricter than the ABA opinion.
A phased approach to get this workflow running and delivering ROI.
Days 1–30
Foundation
Days 31–60
Optimization
Days 61–90
Scale
Smith.ai will sell you the same intake call two ways — $2.00 handled by AI or about $10 handled by a human. Knowing which calls belong in which bucket is the whole decision.
Goodcall bills by unique caller, Rosie by minute and Smith.ai by call. We read each pricing page and worked out what a 300-call-a-month office pays on each, including Smith.ai's human receptionists.
None of the three big legal AI platforms publishes a per-seat price. Here is what each one does publish (research content, security commitments, plan terms) and how a 25-lawyer firm should structure the quote.
One practical AI workflow per week. No fluff.
Get the full guide with step-by-step setup, workflow templates, and copy-paste assets.