Lesson objective

Sequence an intake workflow so that the conflicts check runs on names alone before anyone at the firm, human or software, receives the prospective client's account of the facts.

Plain-language explanation

Why this matters now

Automation changes the order in which a firm learns things. A person on the telephone might instinctively stop a caller and ask for names first. A form or a chat agent does exactly what it was designed to do, and most intake designs ask for the story early because the story is what the firm needs to decide whether the matter is worth taking.

That order matters. Information a prospective client gives the firm about their situation is confidential, and depending on your jurisdiction's rules on prospective clients and on conflicts of interest, receiving it can create obligations that later limit what the firm may do for existing clients. This lesson does not state those rules, because they differ by jurisdiction and you must read your own. It teaches a design that avoids receiving the story until the firm knows whether it can receive it.

Professional rules: ABA Model Rule 1.6 concerns the confidentiality of information relating to the representation of a client. It is a model rule, adopted with different numbering and wording in each jurisdiction, and your jurisdiction's treatment of information received from prospective clients may be set out in other provisions too. Read them before designing any intake flow.

The automation question is sharper than the manual one. Once a narrative is inside an intake system, it is not in one person's head. It sits in a database, may be copied into a relationship management record, may be read by an agent that summarises it into a shared queue, and may be retained in logs. Unreading it is not possible.

A question to sit with

Suppose your conflicts check finds that the prospective client's opponent is a current client of the firm. How many people, systems and logs already hold the prospective client's account of the dispute — and could you list them?

What goes wrong

1. The story box comes first. The web form opens with "Tell us what happened". The conflicts check runs afterwards, on names the firm extracted from the story.

2. The agent reads the narrative to find the names. A design that has an agent pull party names from a free-text account has already had software read and store the confidential account before the check.

3. Only the client's name is checked. The adverse party, related entities, other parties involved and former names are not collected, so the check runs on the one name least likely to reveal a conflict.

4. Matching is too literal. "Harrow Logistics Ltd", "Harrow Logistics Limited", "Harrow Group" and the director who owns it are treated as unrelated. A search that misses variants produces a clean result that is simply wrong.

5. The agent clears the conflict. Hits below a similarity score are dismissed automatically and the inquiry proceeds. Deciding that a match is not a conflict is a judgment, and it has been made by a threshold.

6. The check fails open. The conflicts system is unavailable, the orchestration times out, and the workflow proceeds to narrative collection because nothing told it to stop.

7. The narrative arrives anyway. A prospective client emails a detailed account before any form is filled in, and the intake agent summarises it into the shared queue the whole intake team can see.

The practical answer

Split intake into two stages and gate the second on clearance.

Stage What is collected Who or what may see it Exit condition
A. Names Prospective client; related persons and entities; adverse parties; other involved parties; former and trading names; matter category from a fixed list Intake agent; conflicts search; conflicts analyst Conflicts review recorded by a person
B. Narrative The account of the facts, documents, dates Only after clearance, only the people assigned to assess the inquiry Engagement package ready for attorney decision

The matter category in stage A is a selection from a fixed list, such as "employment dispute" or "commercial lease", not a description. A category supports routing and the conflicts search; a description is the beginning of the narrative.

Tell the prospective client why. An approved sentence explains that the firm needs to check it can act before hearing the details, and asks them not to share the facts yet. This is courteous, it reduces unsolicited narratives, and it makes the design visible.

Give the agent the mechanical work and a person the judgment.

Step Agent may do alone A person does
Normalise names Standardise spelling, suffixes and punctuation
Generate variants Produce trading names, abbreviations, former names given, known related entities from firm records Add variants from knowledge of the client base
Run the search Query the conflicts system on every name and variant
Assemble results Present every hit with the matter, role and status it came from
Decide what a hit means Conflicts analyst reviews every hit, including weak ones
Resolve a potential conflict A licensed attorney decides, and records the reasoning
Record clearance Named reviewer records the result before stage B opens

No threshold dismisses a hit. The agent may rank hits by strength so that the analyst sees the strongest first, but every hit is shown and every one is dispositioned by a person.

Fail closed. If the conflicts system is unavailable, a search errors, or any name could not be searched, stage B stays locked and a person is notified. The design states this explicitly, because the default behaviour of most workflow tools is to continue.

Quarantine unsolicited narratives. When a prospective client sends facts before clearance, the message goes to restricted storage that the general intake team and the intake agent cannot read. A person with access extracts only the names for stage A. The message is released to the assessing team only after clearance, and handled under the firm's policy if clearance fails.

Keep the intake agent's memory per inquiry. An agent that retains context across conversations can carry one prospective client's details into another's. Security standard: the OWASP Top 10 for Agentic Applications for 2026, published on 9 December 2025, lists memory poisoning among its entries, and the underlying lesson for intake is that anything an agent remembers can resurface where it should not. The entry list was verified from a secondary summary and should be confirmed against the OWASP PDF; the list is a voluntary standard, not law. Design intake so that each inquiry starts with no memory of the last.

Remember this

  • Collect names first and the narrative only after a person has recorded conflicts clearance; the story cannot be unread once it is in a system.
  • Stage A collects every relevant name, including adverse and related parties and former names, plus a matter category from a fixed list.
  • The agent normalises, generates variants, searches and assembles; a person dispositions every hit and a licensed attorney resolves any potential conflict.
  • Fail closed: if any search could not run, stage B stays locked and a person is told.
  • Quarantine unsolicited narratives, and give the intake agent no memory across inquiries.

Legal example

The fictional firm Marrowgate Keane LLP inherits an intake form from a marketing agency. Its first field is a large text box: "Tell us about your situation". Behind the form, an agent reads the text, extracts names, and runs a search in the firm's conflicts database. The conflicts analyst, Rory Callender, is asked to approve a plan to extend the same design to the commercial disputes group. He runs the trace-one-story test with a fictional inquiry mentioning a distinctive detail, a warehouse fire at a fictional depot. The detail appears in seven places: the form submission table, the relationship management record, the agent's summary in the shared intake queue visible to fourteen people, the agent's run log, a notification email to the intake inbox, the analytics export, and a nightly backup. He also finds that the agent searched only the prospective client's name, because the adverse party was mentioned in the story as "our landlord" without a name. The redesign has two stages. Stage A asks for the prospective client's name, any trading or former names, related companies, the names of the other side and anyone else involved, and a category from a list of twelve. An approved sentence asks the person not to describe the facts yet. The agent normalises names, generates variants, searches every one, and presents every hit to Rory with the matter and role it came from; nothing is dismissed by score. If any search errors, the inquiry is locked and Rory is notified. Potential conflicts go to the supervising partner, Imogen Faraday-Holt, whose decision and reasoning are recorded. Only then does a stage B link go to the prospective client. In the first month, three prospective clients email detailed accounts before completing stage A. Those messages route to a restricted folder that neither the intake team nor the agent can read, and Rory extracts only the names. One of those inquiries turns out to be adverse to an existing client. Its account never reached the shared queue.

Responsible-use warning

Never let an agent read a prospective client's narrative in order to find the names for a conflicts check. By the time the names are extracted, the confidential account is already in the agent's context, its logs and whatever queue it writes to. Collect names separately and first. Equally, never let a similarity score, a rule or an agent dismiss a conflicts hit, and never let the workflow continue to narrative collection when a search failed. A clean result from a search that did not run properly is worse than no result.

Human-review requirement

A named conflicts analyst dispositions every hit, and a licensed attorney resolves every potential conflict and records the reasoning before stage B opens. The attorney who owns intake signs off the two-stage design. The reviewer confirms that stage A collects adverse, related and former names as well as the prospective client's; that no hit is dismissed automatically; that the workflow fails closed; that unsolicited narratives are quarantined; that the intake agent keeps no memory across inquiries; and that the design has been checked against the adopted rules of the firm's own jurisdiction.

Practical exercise

Redesign a fictional intake flow so that conflicts run on names before narrative.

  1. Invent a firm and describe its current intake flow in five steps, including where the prospective client first describes the facts.
  2. Run the trace-one-story test on paper: list every system, queue, record, notification and log where the account would appear under the current design.
  3. Write the stage A field list. Include the prospective client, related persons and entities, adverse parties, other involved parties, former and trading names, and a category list of at least eight entries.
  4. Write the approved sentence that asks the prospective client not to share the facts yet and explains why.
  5. Complete a table of the conflicts steps showing which the agent does alone and which a person does.
  6. Write the fail-closed rule: the events that lock stage B and who is notified.
  7. Design the quarantine for unsolicited narratives: where they go, who may read them, and when they are released.
  8. State how the intake agent's memory is limited to a single inquiry.
  9. Rerun your trace on the new design and list where the account now appears before clearance. The target is nowhere.
  10. Name the attorney who signs off the redesign by role.

Knowledge check

  1. Why must the conflicts check run on names before the firm receives the narrative?

    Answer guidance: Because the prospective client's account is confidential and, depending on the jurisdiction's rules, receiving it can create obligations that limit the firm. Once it is in a database, queue, agent context or log it cannot be unread, so the firm should know whether it can act before it receives the story.

  2. What does stage A collect?

    Answer guidance: The prospective client's name; related persons and entities; adverse parties; other involved parties; former and trading names; and a matter category chosen from a fixed list rather than a description.

  3. What may an agent do in the conflicts step, and what must a person do?

    Answer guidance: The agent may normalise names, generate variants, run the search and assemble every hit with its source. A person must disposition every hit, a licensed attorney must resolve any potential conflict, and a named reviewer records clearance. No score dismisses a hit.

  4. What does failing closed mean in intake?

    Answer guidance: If the conflicts system is unavailable, a search errors or any name could not be searched, the workflow does not continue to narrative collection. Stage B stays locked and a person is notified.

Sources and further reading