Workflow Automation: Conflicts

The team's process map separates conflicts and new matter intake into 9 recurring steps, even though the details change across practices and record systems. Once the work has been separated, it divides into language, source-backed retrieval, legal judgment, and deterministic configuration, with different evidence, failure states, controls, and accountable performers attached to each one.

The language work appears at steps 1, 2, 5, and the drafting portion of step 8, where a model can extract party names from an email, propose spelling variants, condense an authorized matter record, or draft from an approved template. The reviewer still needs the source in view, authority to correct the output, a stopping point before the draft controls the matter, and a record of the decision that allows the work to move.

Configuration carries a different burden at step 9 because the firm has already chosen the billing rules, rates, required fields, and outside counsel requirements. Versioned rules, validation, and human verification can enforce those decisions without asking a model to improvise a rate or reconstruct a guideline from memory.

Corporate family expansion at step 3 determines the factual set that step 4 would search, allowing an unsupported addition to create noise, an omission to create a false negative, a source failure to reduce coverage, and an unresolved discrepancy to pass through a perfectly functioning engine until the reviewer receives a clean report the evidence never earned.

The record moves into professional judgment at steps 6 and 7 under the rules governing current clients, former clients, imputation, and prospective clients. Technology can organize relationships, display the evidence, preserve the audit trail, and route the question, while the attorney retains responsibility for adversity, waiver, conditions, clearance, or decline.¹

The process map shows a request moving from intake to search, review, and decision, although the case moves differently once a misspelled party name enters the form. The database returns nothing, the reviewer leaves the system for an old matter note, the note leads to an attorney who remembers why the relationship has received different treatment 4 years earlier, and the answer moves into an email thread because the form has never learned how to ask the question that resolves it.

When the request reappears in the intake queue, the official workflow still shows one clean path even though the matter has crossed a form, a database, a note, a person's memory, and an inbox. The exception queue preserves the truer route through names that arrive wrong, families that change, descriptions that bury adversity, and decisions that depend on history no system has captured.

While the request waits, the working process separates further from the written procedure. A missing field sends someone back to the requester, ambiguous authority leaves the matter waiting for a partner, disconnected records force a manual search, an unavailable decision maker stops the queue, and local knowledge appears only when the person who carries it happens to be present.

Following the retyped field through the incomplete request and the person who chases it exposes where errors create rework, where ambiguity requires judgment, which exceptions depend on one person's knowledge, and what the written procedure has omitted when it reduces the work to a sequence of boxes.

Each detour gives the build a requirement because the reviewer's next action reveals what information has to survive, where the exception needs to travel, when the system has to stop, and whose authority can resolve the question. The answer then needs a route back to the record so the next matter will not depend on the same accidental memory.

Once the working process becomes visible, most of the engineering weight moves toward retrieval. A model can extract, compare, summarize, and draft while remaining unable to make a stale record current, prove an unrecorded relationship, resolve a contradiction the sources leave open, or choose silently between sources that answer different questions.

A preregistered evaluation of legal research products from LexisNexis and Thomson Reuters found hallucinations in 17 percent to 33 percent of tested responses. The study examined legal research queries and particular product versions rather than conflicts systems or corporate entity data, so it supplies no conflicts-intake error rate and supports the narrower conclusion that retrieved sources can reduce model error while leaving material error in the answer.²

The case requires one query to cross public registries, commercial hierarchy data, confidential matter records, screened files, and stale caches before the reviewer sees one name. Access, authority, coverage, freshness, and the meaning of a silent source therefore belong to the retrieval design rather than its background documentation.

Before candidate assembly begins, the access layer must separate client and prospective-client information, attorney mental impressions, screened matters, declined-matter records, public registry data, and commercial hierarchy data, because placing them in one undifferentiated model-visible corpus turns retrieval into disclosure and moves purpose limitation, data minimization, segmented indexes, field-level permissions, and access decisions ahead of generation.

The sources themselves move at different cadences and support different claims: SEC EDGAR updates its APIs in real time as filings disseminate, with typical processing delays under 1 second for submissions and under 1 minute for XBRL data, GLEIF publishes Golden Copy files 3 times daily, Delaware presents entity information current as of the search while prohibiting data mining and automated tools through its public search, and OpenCorporates directs users to correct the official public record first and allow 30 days for its own record to update.^{3 4 5 6}

A source may be current and still answer only part of the question, which prevents freshness from creating a universal hierarchy among records whose coverage differs. A state registry may establish existence and filing facts without displaying a complete corporate family, SEC filings may evidence disclosed relationships for public filers without promising a complete current hierarchy, GLEIF may provide identity and reported parent records where coverage exists, including reporting exceptions, and a commercial source may reveal candidate relationships that still require validation.

As each proposed relationship enters the search set, its evidence trail has to identify the supporting source, the date on which that source last changed, the coverage the source does not claim, the relationship type drawn from the record, and any corroboration still required before the reviewer treats the name as part of the family.

When every source in the defined search plan answers, the system can report a covered zero. If a source fails, times out, returns partial coverage, or cannot be queried under its terms, the same blank screen represents a degraded zero. Neither condition proves the absence of a conflict unless the reviewer can see the names searched, variants used, relationships considered, source dates, response statuses, and unresolved gaps.

The proposed index reaches immediately beyond public company data into client names, prospective-client disclosures, matter descriptions, attorney impressions, waiver histories, screened relationships, and the fragments from which the firm will later reconstruct what it knew when the request arrived. Professional responsibility rules protect client and prospective-client information, while some records may also carry attorney-client privilege or work product protection, categories that overlap without becoming identical and leave confidentiality extending far beyond privileged communications.

ABA Formal Opinion 512 treats generative AI as a competence, confidentiality, communication, supervision, candor, and fee issue, and where a self-learning tool creates the disclosure risk the opinion describes, the lawyer must obtain informed consent before entering information relating to a representation because generic engagement-letter language does not provide it. The same opinion carries outsourcing diligence into vendor review, including the vendor's conflicts system, security, retention, contractual protections, and available remedies.⁷

A prospective client may disclose enough information through an intake chatbot to trigger protection before the firm has accepted the matter, which is why Oregon Formal Opinion 2026-208 connects marketing and intake systems to Oregon RPC 1.18 and directs Oregon lawyers who receive sufficient prospective-client information through a chatbot to enter the person into the conflicts system promptly. The jurisdictional instruction belongs to Oregon, while the operational problem travels wherever intake automation must capture protected information quickly enough for the conflicts process to use it.⁸

In the opinion's autonomous-agent example, information gathered during intake moves into engagement-letter drafting and then into the client management system under a goal to engage a new client. One automated chain can therefore cross disclosure, conflicts, acceptance, communication, and record creation before any lawyer has decided what the firm intends to do.

While a generated letter remains in a review queue, a lawyer can decide whether it should communicate acceptance, request a waiver, impose conditions, or decline the representation. Once the system sends it, the recipient may reasonably read the same language as the firm's decision, and the workflow has already exercised the authority it was supposed to preserve for counsel.

Screened matters must disappear before retrieval assembles the candidate set because hiding a summary after generation leaves the protected content inside the model's context, whereas a properly timed access decision excludes the record while preserving enough audit evidence to show that the control ran without revealing what it protected.

Formal Opinion 512 warns that a self-learning tool may surface information from one representation to people who should not receive it, and the Pennsylvania and Philadelphia bars connect that risk to conflicts and confidentiality by warning that inadequately safeguarded models may use information from one representation to inform another and by directing lawyers to keep client confidential information out of tools that lack adequate protection.⁹

Once the design allows matter records to feed a retrieval system, those duties follow the information into corpus selection, permissions, vendor contracts, logs, retention, training, and output review. They also reach the point at which a person has to stop the system before a fluent answer crosses a boundary the underlying record cannot.

A client may create the next exposure on a platform the firm never selected, as United States v. Heppner shows. The Southern District of New York held that a criminal defendant's independently generated exchanges with a public AI platform carried neither attorney-client privilege nor work product protection, relying on the platform relationship, the privacy policy before the court, the absence of attorney direction, and the failure of the documents to reflect counsel's strategy when created.¹⁰

One sentence in Heppner observed that attorney direction might arguably support an agency theory. The court preserved a possibility without creating a safe harbor, placing client instruction inside the firm's legal workflow because counsel can control what it tells the client long before it can control what a public platform stores, reuses, trains on, or discloses under its terms.

Warner v. Gilbarco reached a different result on the same day as Heppner's bench ruling, when a magistrate judge protected a pro se civil plaintiff's AI exchanges as work product after reasoning that ChatGPT served as a tool and that disclosure to it did not waive protection without a path to an adversary. Their different procedural settings and approaches to work product make the cases poor foundations for a universal rule and strong reasons to document platform terms, attorney direction, litigation purpose, and disclosure risk.¹¹

By the time a client's AI exchange reaches counsel, the platform may already hold the text and the confidentiality analysis may depend on terms the client never read. The firm has also lost the opportunity to shape how the material came into existence.

The supervision problem appears in the difference between work that carries an author's name, role, instructions, and drafting history and output that arrives without any trace of how the answer has come together. Generative AI can remove those signals and leave prose polished enough to enter the file before anyone reconstructs what has produced it.

When a row arrives without a source, retrieval status, drafting history, or accountable chain, its fluency conceals whether the name comes from a registry, a stale cache, a matter note, or a probabilistic completion. The reviewer then has no reliable place to direct attention.

The interface has to restore the signals that identical rows erase. Source class, source date, relationship type, retrieval status, unresolved discrepancy, and any model contribution need to reach the reviewer with the name rather than appear later in an audit log no one consults during clearance.

A matter summary needs the same route back through the authorized text the reviewer can inspect, the boundary between record content and inference, the person who accepts or corrects the draft, the decision log, and the matter file that will carry the conclusion into the next request.

ABA Formal Opinion 512 relies on Model Rules 5.1 and 5.3 when it directs firms to create policies and training and suggests marking AI-produced material in client or firm files so later readers understand its fallibility. Oregon Formal Opinion 2025-205 similarly requires training on the actual tool, including its capabilities, limits, data handling, privacy, and confidentiality, and requires lawyers to verify AI-assisted work.^{12 13}

The adoption test looks past the license and follows the work one year after rollout. The license can remain active while the intake team returns to email threads, inherited forms, memory, and the same side channels the product is supposed to replace, leaving the dashboard to record availability while the work records abandonment.

Because the request already arrives through email, a form, and an existing matter system, the design places the extracted parties, proposed variants, source-backed relationships, source dates, missing fields, and coverage failures where the reviewer already works. A separate navigation habit would add effort before the work could begin and give reviewers another reason to return to the path they already know.

Placing the inbound request beside the extracted parties, each relationship beside its provenance, each summary beside the authorized record it condenses, and each unresolved gap beside the person or source that can close it keeps an inference from acquiring the visual status of fact merely because the interface gives both the same row.

The team still needs training on the tool's error profile, data handling, stopping points, and escalation paths, although the product team carries the learning cost of reaching the work rather than transferring it to reviewers who already know where the intake request appears.

At the moment of review, the screen has to reveal what the system read, what it inferred, which sources it reached, where retrieval failed, and what moved to legal judgment. Without that path, the reviewer would supervise the answer while the work that produced it remained hidden.

Eight hundred uses can describe 80 reviewers returning across difficult matters or 3 people processing easy requests while everyone facing a hard call retreats to email and memory. Because aggregate activity cannot distinguish those conditions, the measurement plan follows repeat usage per eligible reviewer and the matters on which reviewers return.

Following each matter from request to decision brings the useful measures into view along the path, including the share of extracted parties that survive without correction, the distribution of clearance time, the volume and age of the exception queue, the share of reports with complete source coverage, and the rate at which reviewers override proposed relationships or summaries.

In its 2026 AI in Professional Services Report, Thomson Reuters Institute reported organization-wide AI use at 40 percent, based on more than 1,500 respondents across 27 countries, while only 18 percent said they knew their organization tracked return on investment from AI tools in some manner. Those figures cover professional services broadly, so an intake team must still establish whether its own workflow reaches the correct parties, sources, reviewer, and decision.¹⁴

Instrumentation that stops at the click can show that someone opened the tool while leaving the firm unable to tell whether the candidate set is complete, the exception reached the right lawyer, the reviewer trusted the provenance, or the final decision improved.

Matter descriptions that say "general corporate advice," affiliate records that omit relationships, waiver documents that remain untagged, and outside counsel requirements stored in disconnected systems bring the model to the edge of the available record almost immediately. Model quality cannot cure the resulting records problem, which returns the work to ownership, maintenance, source mapping, and decisions about what the system can responsibly claim.

When one person's availability controls the queue, faster extraction and cleaner summaries deliver more matters to the same unavailable decision maker while the product improves the waiting room and the decision remains exactly where it has been.

Where exceptions dominate, an exception-routing and knowledge-capture layer offers more value than another generation feature because it can record the gap, find the person who can resolve it, preserve the answer for the next request, and return that knowledge to the governed record.

Following the workflow while it runs keeps the case attached to the request through every handoff and detour. The work separates into steps small enough to judge, with an accountable person or system attached to each one and a distinction among language, source-backed retrieval, deterministic configuration, and legal judgment before anyone chooses where a model belongs.

At that level of separation, the architecture can attach visible source text and review to language tasks, coverage and provenance to retrieval, versioned rules and validation to configuration, and organized evidence to the attorney whose judgment controls the legal decision.

Corporate family expansion carries the sharpest retrieval risk in this case because fluent output can alter the factual set before the search begins. Other legal workflows conceal the same problem inside a step that resembles writing even though the answer depends on evidence the model cannot create.

A team that finds the evidentiary step, traces what enters it and what leaves, and builds the source, access, review, and escalation structure around the decision can intervene before a model's fluency makes an unsupported answer look complete.