Automation & AI Agents
Automated client intake at law and accounting firms: question by question
The first call decides whether an inquiry warrants a paid consultation. How an AI agent collects facts without giving advice, checks for conflicts of interest before the firm is exposed, and rates urgency by verifiable anchors.
In short
An AI intake agent at a law or accounting firm collects the facts of an inquiry without giving advice, checks for conflicts of interest before the firm hears the story, and rates urgency by a deadline written in a document. Anything beyond that, such as a match on the conflict list or a professional question, goes to a person with an organized record.
Contents
On the first call at a law or accounting firm, an AI agent does four things: it collects facts without giving advice, checks for conflicts of interest before the firm hears the story, rates urgency by anchors that can be verified, and hands everything beyond that to a person. It answers the phone line during the hours when nobody else does, and continues on WhatsApp to collect the documents. Here you'll find the sequence of questions, where the conflict list comes from and how often it's refreshed, where the caller's words are sent, and how to test it all before you connect it to the line. This is one of the places where AI agents for business save the firm's most expensive time: a partner's time spent answering a first call.
The first call at a firm is work nobody puts a price on
At a professional firm, there's no real "lead." There's an inquiry, and it becomes a case only after someone has listened long enough to understand what happened, whether the firm is allowed and able to represent the caller, and whether the caller is willing to pay for a consultation. Nobody bills this call by the hour, so it falls to whoever is free, and that's where nobody asks about the opposing party until the caller has already told the whole story.
Screening a firm's inquiries is triage, not sales, and that's the part you can standardize: the same questions, in the same order, at any hour. That's why we build AI agents that answer client inquiries at this layer and not at the layer where the professional decision is made, usually for firms whose inquiry volume went up thanks to SEO for law firms or SEO for accountants, while their phone still went to voicemail at 6 p.m.
Who answers at 8 p.m., and on which channel?
The agent answers on two channels, and both are defined in writing in advance. On the phone, it steps in by an explicit routing rule: after five rings, on every call outside the 9 a.m. to 6 p.m. window, and on a second call that comes in while the secretary is on the line. On WhatsApp it answers every message, and that's where the inquiry continues after the call, because that's the channel the documents arrive on.
The handoff to a person is the part everyone announces and nobody builds: an alert that nobody is waiting for isn't an escalation ladder, just a light that blinks. That's why the on-call rotation is assigned by name. An urgent inquiry triggers an alert on the on-call person's phone, and if it isn't acknowledged within 15 minutes, the system calls them, and after 45 minutes it moves on to the next name. For Friday and Saturday there's a stated response window, and the agent says when a person will get back to the caller instead of promising a callback that won't happen.
How do you collect the facts without giving legal or accounting advice?
The agent asks; it doesn't conclude. The separation is enforced in the script: each type of matter has a closed list of factual questions, and any wording that touches on chances, eligibility or an amount is flagged as out of scope and passed to a person. "When did you receive the letter?" is a question. "That sounds like wrongful termination" is an answer, and it belongs to whoever is sitting in front of the documents.
And when callers ask "So what are my chances?", a flat refusal kills the conversation: the agent offers a process instead of an assessment (which documents to bring and when), and the question is logged as an open question in the file.
The sequence of questions in one inquiry, and the record that comes out of it
This is what it looks like in an employment termination matter. The fourth question is the boundary: after it the story begins, and before it the conflict check runs.
- What the matter is about, in two sentences, in your own words
- Full name and ID number
- The employer's name and company number as they appear on the pay slip
- Who else is involved whose name might appear in our records: the business owner, a direct manager, a lawyer who's already in the picture
- When the employment started and when it actually ended
- Whether a letter was received, and what date is on it
- Whether a pre-dismissal hearing was held, and on what date
- Which documents are in hand right now: a contract, pay slips, a letter, correspondence
What comes out isn't a free-form summary but a form: the type of matter, the identities of the parties, noting what was verified and what was given verbally, the time and result of the conflict check and against which version of the list, a short factual summary, every deadline with its source, the documents that exist and the ones that are missing, and an urgency rating with its reasoning. And one rule worth more than any field: whatever the agent isn't sure about is marked as such.
How do you check for conflicts of interest before the firm is exposed?
You check at the beginning, not at the end. The agent asks for names and ID numbers before the caller starts telling the story, runs them against the list of clients and opposing parties, and if a match comes up, it stops and passes the call to a person. The real risk isn't that the firm will take a case with a conflict of interest, but that it will receive information it should never have heard. The moment the caller has laid out the dispute, the firm is exposed, even if it turned the inquiry down a minute later.
Matching on ID numbers is stronger than matching on names, but it isn't certain: an Israeli ID number and a company number both have nine digits, callers give one instead of the other, and swapped digits can land on another valid number. A check digit proves that the number is built correctly, not that it belongs to the person who said it. In Hebrew it's worse, because spellings with and without vowel letters, and different transliterations, create misses and false matches. That's why a partial match goes to a person, and the threshold is calibrated against the last 100 inquiries.
Where the list comes from, and how often it's refreshed
A week-old list produces the one failure that must never happen. There are three options: a direct API, if the case management system exposes one, and in many Israeli systems that isn't the case; a scheduled export of an encrypted file to a locked folder every night at 2 a.m.; or a shared spreadsheet in a small office. At accounting firms, the list usually sits in the bookkeeping system: Hashavshevet, Priority or something similar. Two things are needed around the export: a timestamp, so that if it's older than 26 hours or the export failed, the agent takes only a name and contact details; and an intraday update, because a party added in the morning must be on the list before the opposing party calls in the evening.
Conflicts of interest the caller can't name
Real conflicts sit at the group level: a parent company, a subsidiary, a shareholder who's a client in another matter, or a positional conflict. The caller doesn't know who owns the company that's suing them, so you take the company number, use it to pull the shareholders and directors from the Registrar of Companies, and compare at the group level. Unknown ownership is logged as an open, blocking item. A positional conflict can't be found by comparing lists, so the record carries a line that states the legal question itself.
Where the details the caller shared on the call are sent
Part of the processing happens at an outside vendor, and that's a fact to be stated at the partners' meeting, not discovered later. The AI engine that transcribes and drafts the record doesn't run inside the firm but on a vendor's servers, usually outside Israel, so things protected by attorney-client privilege pass through a third party before any engagement has been signed.
- A locked processing region, no use for training and zero retention at the vendor, as written in the contract and not as a promise over the phone
- A data processing agreement and a list of subprocessors: who transcribes, who hosts, who sends the WhatsApp messages
- Masking before sending: ID and account numbers are replaced with a token, and the full value stays on the firm's side
- The conflict check doesn't leave the firm at all. Comparing lists doesn't need a language model, and there's no reason for client names to leave the office
- Deletion on a schedule: how long the recording is kept and how long the transcript
And on the firm's side: what's said at the opening (that this is an initial intake that isn't advice and doesn't create representation), and who sees what. A family law or criminal inquiry isn't open to everyone with access to the system.
How do you tell an urgent case from someone shopping around on price?
Not by tone, but by signs that can be verified. The agent rates by four anchors, not by how stressed the caller sounds.
- A deadline stated in a document: a hearing date, a deadline for an objection or an appeal, a deadline for filing a report
- A document already in hand: a tax assessment, a statement of claim, an order, a demand for payment, a dismissal notice
- An action already taken: an agreement that was signed, an application that was filed, a payment that was transferred
- Exposure that grows over time: interest and index linkage, late penalties, a running statute of limitations
Almost every caller says it's urgent, so instead of arguing, the agent asks for an anchor: what the document says and what date is on it. The opposite direction is no less dangerous, because some people mention in passing "I got some letter from the tax authority" without realizing they're already on a clock. And someone who opens with "How much does it cost?" with no document and no date doesn't justify a partner's time at 2 p.m.
How Israelis contact a firm, and which rules apply here
The inquiry almost always starts on the phone and continues on WhatsApp. Within an hour, photos of the tax assessment or the statement of claim arrive, usually shot at a crooked angle, and sometimes a voice message at night. An intake system built around a website form doesn't match this reality.
Local details trip systems up. An ID number dictated aloud loses its leading zero, so you validate the check digit and not just the length. The difference between a limited company and a Licensed Dealer (a VAT-registered sole proprietor) changes the type of matter. Dates are given as "after the holidays" or appear as a Hebrew calendar date, and they need to be converted to one calendar and read back. Callers mix Hebrew with Arabic, Russian or English. And the phone rings at 8 p.m. and on Friday, when the authorities are closed.
And on top of all that, there are rules that need to be translated into system settings, not into a paragraph in a policy document:
- The Israel Bar Association (Advertising) Rules, 2001 prohibit solicitation: the agent steps in only on an inquiry the caller initiated, and it doesn't present itself as a lawyer
- Privilege under Section 90 of the Israel Bar Association Law applies even without a signed engagement, so the masking and the deletion aren't window dressing
- For accountants, there are also independence rules under the Certified Public Accountants Regulations, and for certain services an identification duty under the Prohibition on Money Laundering Order, which a person completes
- After Amendment 13 to the Privacy Protection Law, 1981, in force since August 2025, an intake database almost always holds information of special sensitivity. Questions for your adviser: whether a privacy protection officer is required, and what the notice wording should be
- Recording: the Wiretap Law permits recording by a party to the conversation, and who counts as that party when the answering is automated deserves a written decision. In the meantime, a notice at the opening
- Section 30A of the Communications (Telecommunications and Broadcasting) Law: answering an inquiry isn't an advertisement, but a later marketing message is, and opt-out requests must be handled
Where it breaks, and how to test before you connect the main line
It breaks in four known places, and each one has a behavior defined in advance; otherwise the failure reaches the desk as if it were a fact.
- Misclassifying the type of matter. The classification is a suggestion, not a verdict: the agent reads back "I've noted that this is about the end of your employment," and a correction is saved as a corrected field
- A date read from a crooked photo isn't entered as a fact: it's read back, and without confirmation it's marked as unverified
- A transcription that breaks when the caller switches between Hebrew and Arabic or Russian. Below a confidence threshold, the agent doesn't guess: it asks again with a closed question, and after two failures it passes the call to a human queue with the audio
- A caller who hangs up on realizing they're talking to a system: an opening that says who's speaking, a transfer to a person the moment they ask, and measuring hang-ups in the first 20 seconds
Then you test in shadow mode. On WhatsApp it's easy: the agent drafts, a person approves every message before it's sent, and the gap between them is the data. On the phone you start with a narrow window, only after 6 p.m. and only one type of matter, for two weeks or 100 inquiries.
Acceptance criteria are set before the run, and two of them are blocking: zero cases in which a factual summary was collected before the conflict check was completed, and zero statements that assess chances, eligibility or an amount; a single failure stops the go-live. The rest of the threshold is quantitative, and you set it: no more than two records out of 20 with a material field wrong. The handoff rate is measured in two buckets, handoffs that were meant to happen and handoffs that are a failure of the tool, and only the second kind is a bug. And a rollback switch, a single action that returns the line to its previous routing, in the hands of the person named for it.
What you don't entrust to it is just as clear: the decision whether to take a case, fees, contact with the opposing party, and any document that goes out in the firm's name. An AI agent for lawyers and an AI agent for accountants work best when they do one thing in depth: bring an organized inquiry to the desk, clear of conflicts of interest and correctly rated. The rest we build together, and it starts with a short conversation about how your firm works.
Frequently asked questions
How is client intake at an accounting firm different from a law firm?
At an accounting firm, the inquiry starts from a business status rather than a dispute, so the first questions deal with the type of entity, the reporting frequency and who handles the books today. Urgency is measured against filing deadlines, and the conflict check is aimed at the business's partners and related companies.
What happens when both sides of the same dispute contact the same firm?
It's common, especially in disputes between neighbors, partners and family members. The list must include previous inquiries too, not only clients who signed, or the other side gets taken on as if nothing had happened. When a match comes up, the decision goes to a person.
How does the agent handle a caller who raises several matters in one call?
A business owner who calls about a tax assessment may also mention a dispute with a former employee along the way. The agent splits them into separate matters, each with its own type, urgency rating and conflict check, so the urgent matter doesn't get swallowed up in the long story.
Should you keep the intake call recording or only the transcript?
The recording is worth the most in the first weeks, when you're calibrating the questions and checking where the transcription fails. After that, the value drops and the exposure remains, so the simple logic is a recording for a short window and a structured record for the long term.
Is automated intake right for a firm of two or three people?
In a small firm in particular, the load of the first call falls on the people whose time is the most expensive. The difference is in scope: two or three types of matter instead of ten, and a conflict list that's a spreadsheet rather than an export from a system.
The bottom line
The first call at a firm is the moment when conflicts of interest, priorities and the quality of the information that will reach the meeting are decided, and today it's handled in the least organized way of the entire process. An agent that handles it properly doesn't advise and doesn't decide: it standardizes the questions, checks against an up-to-date list before the firm is exposed, and hands off to a person by the clock. What's left to decide is where the information is processed, what's kept, and what threshold must be met before you connect it to the line.

