AI Intake for Law Firms: Run the Conflict Check Before the Callback

Fast is the wrong first goal

Most advice about law firm intake is about speed, and the numbers behind it are real. In Clio's 2024 Legal Trends Report, secret shoppers contacted 500 firms: only 40% picked up the phone and just 33% answered an email. Smith.ai's own analysis of 3.2 million law firm calls, published in October 2026, puts roughly one call in five after hours. So firms put an AI chatbot on the website or an AI receptionist on the phone, and inquiries stop going unanswered.

Speed is the second thing intake has to get right. The first is order. Before anyone at the firm hears what happened, the names of everyone involved should already have been checked against every client, former client and adverse party the firm has. An AI intake that is fast and gets the order wrong collects, at 11 p.m. and at scale, exactly the information that can disqualify the firm from work for an existing client.

This is the builder's guide to doing it in that order: what the rules ask of intake, what an AI assistant may and may not say, the architecture I build, and when to buy instead. I am an engineer, not a lawyer. The ethics opinions cited here are advisory, they vary by state, and the calls belong to your firm's counsel.

What Model Rule 1.18 asks of intake

ABA Model Rule 1.18 says a person who consults a lawyer about possibly forming a client-lawyer relationship is a prospective client. Three things follow, even if the firm never takes the matter:

• Confidentiality. What the lawyer learned may not be used or revealed, except as Rule 1.9 would allow for a former client.
• Disqualification. If the lawyer received information that could be significantly harmful to that person, the lawyer may not act against them in the same or a substantially related matter, and neither may anyone else in the firm.
• A narrow way out. The firm can still act with both sides' informed consent in writing, or if the lawyer took reasonable measures to avoid hearing more than was reasonably necessary to decide whether to take the matter, is timely screened, takes no part of the fee, and the prospect gets prompt written notice.

Two comments turn that into intake design. Comment 4 says to limit the initial consultation to what that decision needs. Comment 2 says a consultation is likely when a lawyer, through advertising in any medium, invites information about a potential representation without clear warnings that limit the lawyer's obligations. A website chat that asks "tell us what happened" is that invitation.

ABA Formal Opinion 510, issued in March 2024, filled in the practical part. As the ABA Journal summarized it, the lawyer should warn the prospective client that the firm has not agreed to take the matter and that information should be limited to what is needed to decide whether to move forward. Without reasonable measures, the lawyer's conflict is imputed to the whole firm.

For a system, that means three requirements. Collect names before facts. Say, before the first question, that no lawyer has agreed to act. And keep a record of what was asked and why, so the firm can show its reasonable measures if it ever needs the screen.

What state bars say about AI chat intake

Florida Bar Ethics Opinion 24-1 (January 2024) was the first to address intake chatbots directly. A lawyer must inform prospective clients that they are communicating with an AI program, not a lawyer or an employee of the firm. It warns against an overly welcoming chatbot that may give legal advice, notes that a prospective-client or even a lawyer-client relationship can be created without the lawyer's knowledge, and applies the Rule 1.18 duties when AI interviews prospective clients.

Oregon Formal Opinion 2026-208 (February 2026) is the most specific so far, and it separates rule-based chatbots from AI agents that act on their own. Its answer is "yes, qualified", and the qualifications read like a specification:

• Clear disclaimers on the website and in the chat, because a chatbot that sounds like a lawyer could make someone believe a relationship has formed.
• A responsible lawyer notified of each interaction and able to review it promptly, to decide whether the firm will proceed or tell the person that no relationship will follow.
• Prospects entered promptly into the conflict management system once the chatbot has collected enough to identify them.
• Lawyer approval for anything an agent would otherwise do by itself. The opinion's own example is an agent that sends an engagement letter because it decided that furthers the goal of signing a client.
• Due diligence on the vendor and every product underneath it, including whether client information is kept out of training.

The Oregon opinion also points to the New York City Bar's Formal Opinion 2024-5, which treats a chatbot that does not identify itself as a possibly misleading communication, and to the Pennsylvania and Philadelphia bars' joint Opinion 2024-200 on AI that gives legal advice. Behind all of them sits ABA Formal Opinion 512 (July 2024): lawyers need a reasonable understanding of the tools they use, must get informed consent before client information goes into a tool that learns from its inputs, and cannot rely on boilerplate in an engagement letter to get it.

Put together, the assistant's script has hard edges:

The order I build: names, conflicts, then conversation

The model handles language. The conflict search and the deadline arithmetic are code. Every decision that commits the firm is a person's. In order:

1. One queue for every channel. Website chat, web form, email, and after-hours calls as transcripts from an answering service or a voice agent, each with its source and time.
2. Extract names and dates only. A model fills a fixed schema: every party and their role, companies and the people behind them, the type of matter, the state, key dates such as the date of service. It marks inferences as inferences and never invents a party.
3. Run the conflict search in code. Every party, plus known parents, subsidiaries and former names, against clients, adverse parties, related parties, closed matters and earlier prospects in Clio Manage, MyCase or Filevine. Matching is fuzzy on purpose, with thresholds the firm's risk partner sets.
4. Flag deadlines from the lawyers' own table. Limitation periods vary by state and by type of claim, and once someone has been served there is a response deadline: 21 days to answer a federal complaint under Rule 12(a) of the Federal Rules of Civil Procedure, other periods in state courts. A hit marks the inquiry urgent and alerts a lawyer. The prospect never sees a date from the system.
5. Hold or clear. A possible match goes to the conflicts person with the matched records side by side, and nobody has a substantive conversation until it is cleared. A clean search goes to the intake lawyer's list.
6. A lawyer accepts or declines. Accepting books the consultation from that lawyer's calendar and drafts the engagement letter for the lawyer to send. Declining drafts a non-engagement letter for the lawyer to approve.
7. Everyone goes into the conflict system. Accepted or not, the prospect and every party they named are recorded, because the Rule 1.18 duties survive a decline and the next inquiry has to find them.

Each step is logged: what the assistant asked, what the person sent, the exact search inputs and results, who cleared each hit and why. That log is the firm's evidence of reasonable measures. The full blueprint, with the rollout stages and a worked US example, is in law firm client intake.

Why an exact-match search is not enough

Prospects give the name they use, not the name in your system. "Acme Restaurant Supply" is "ACME Restaurant Supply, L.L.C." in the matter list, a distributor is listed under the trade name it had before a rebrand, a person appears under a maiden name or with a transposed middle initial, and the company suing your prospect is a subsidiary of a client you represent elsewhere. An exact search returns nothing for all of them, and "nothing found" is the result that lets the callback go ahead.

So the search strips legal suffixes and punctuation, normalizes case and spacing, compares token by token, and includes known related entities. It produces more false positives than an exact search, on purpose, and a person clears each one. If you want to see how many near-matches your current search would miss, I built a free conflict check tool that runs fuzzy matching on your own list of names, in your browser.

Where calendaring fits

"AI calendaring" means two different things in a law firm. Booking the consultation is the easy one: it comes after the conflict check and after a lawyer accepts, from that lawyer's calendar. The other is court deadline calendaring once a matter is open, and that deserves its own system with a lawyer confirming every date. Malpractice insurer ALPS, reading the ABA's Profile of Legal Malpractice Claims 2020-2023, puts administrative errors such as failing to calendar properly at nearly 25% of all claims. That side is covered in court deadline docketing.

Buy, build or both

Start with what you already pay for, and test its conflict search before anything else. Here is what the vendors document as of October 2026:

Sources: the Clio help center, LawSites on Grow AI, the Lawmatics help center, the MyCase support center, and Smith.ai's documentation. Smith.ai's own blog adds that its check only matches names and that it takes no responsibility for misspellings, which is honest and is exactly the gap.

My advice is the same as on the AI receptionist build-or-buy question: buy when it fits. A firm with one practice area, a clean contact list and Clio should try Grow AI first, and an answering service with a name check is a real improvement over voicemail. Ask each vendor three questions, and ask to see the answer on your own data:

1. Does the search find a party under a misspelling, a former name and a parent company, or only the name the caller gave?
2. Can the system stop a callback or a booking until a person clears a possible match?
3. Does it record who cleared each hit, what they saw and why?

A build earns its cost where those answers are no: several practice groups with their own triage rules and deadline tables, a conflict search that has to cover corporate families, or a firm that wants the intake logic, the logs and the model provider under its own control. Often it is both: Clio Grow or Lawmatics stays the home of intake, and the build sits in front of it, running the wider search and the deadline flags before a lead is created.

What it costs

A build like this usually lands in the multi-step workflow tier of my AI workflow automation work, $15K - $30K over 3-5 weeks. The effort sits in the conflict search, the practice management integration and the deadline tables, not in the model. A firm practicing in several states, several practice groups or a search across corporate families pushes it toward the top. Running costs are small: model calls cost cents per inquiry, because each one is extracted once and the drafts are short, plus hosting and any company data lookups.

What to measure from the first week

• Possible conflicts held before contact. Every possible match should reach the conflicts person before anyone discusses the facts. The target is all of them.
• Misses in shadow mode. Possible conflicts your current search found and the new one did not. It must be zero before the old search is retired.
• Transcripts with advice in them. Every chat is reviewed at first, then a sample. The target is zero.
• Time to first human contact. Split by office hours and after hours, and by urgent and normal inquiries.
• Declined prospects in the conflict system. Every one of them, with the parties they named.

Bottom line

AI intake solves the speed problem easily, and that is the trap. The ethics rules care about order: names before facts, a warning before the first question, a person clearing every possible conflict, and a lawyer, not a chatbot, deciding whether the firm acts. Build that order into the system and the after-hours chat becomes safe to switch on. Leave it out and the firm is collecting disqualifying information faster than ever.

Frequently asked questions