Can an AI conflict check find every possible match?

No system can guarantee that. Results depend on the completeness and quality of intake data, historical records, naming conventions, matter descriptions, access permissions, and search rules. The workflow must expose uncertainty and support human review rather than claiming perfect coverage.

What is a sensible first conflict-check pilot?

Use a representative set of previously reviewed matters with known outcomes. Run the assistant in shadow mode, compare its candidate matches and source evidence with the firm's existing process, and measure missed matches, false positives, review time, intake completeness, source traceability, and escalation quality.

AI Conflict Check Assistant for South African Law Firms

A potential client phones with an urgent commercial matter. The person gives a trading name, mentions two directors, refers to “the group”, and wants advice before close of business. The firm searches one spelling in its matter system, checks an old spreadsheet, asks a partner in a group chat, and waits.

The delay frustrates the prospect. A rushed clearance can create a much larger problem.

An AI conflict check assistant law firms South Africa can use responsibly should not decide whether a firm is free to act. It should make the search more complete and reviewable: collect the right identities, resolve naming variations, search approved records, show possible relationships with sources, and route the evidence to the lawyer or risk owner who has authority to decide.

What an AI conflict check assistant actually does

A managed conflict check assistant can support the administrative and evidence-preparation stages of new-client, new-matter, lateral-hire, vendor, or other firm-approved checks. Depending on the firm’s policy and systems, it can:

open a structured request with a unique reference
collect legal names, trading names, former names, registration numbers, and identifiers
capture related entities, directors, shareholders, counterparties, witnesses, experts, funders, and other relevant people
ask approved follow-up questions when information is incomplete
normalise punctuation, spacing, initials, titles, and common company suffixes
generate approved name variations and aliases for search
preserve the name exactly as supplied
search permitted client, contact, matter, document, billing, and relationship records
identify exact, close, phonetic, transliterated, and historical-name candidates
show why each record may match
link every candidate to its source record
separate identity confidence from possible conflict relevance
group duplicate results
prepare a concise review pack
route high-risk or ambiguous results to the correct reviewer
record reviewer questions and decisions
prevent matter opening until required approval is recorded
monitor for approved changes in parties after opening
preserve approved search rules and lessons in the Company Brain

It should not conclude that two people are the same without sufficient evidence, decide that a conflict exists or does not exist, reveal confidential client information to an unauthorised requester, assess professional duties, approve a waiver, create an information barrier, accept a mandate, or open a matter without the firm’s required human decision.

The role is search preparation, evidence, routing, and process control. Professional judgement stays with authorised humans.

Why conflict checks become slow or unreliable

The search box is rarely the whole problem. The quality of a conflict check depends on the information collected, the records searched, the relationships understood, and the decision trail preserved.

Common breakdowns include:

an intake request containing only one person’s common name
trading names used instead of registered entities
old company names missing
initials recorded in one system and full names in another
spelling and transliteration variations ignored
groups, subsidiaries, trusts, partnerships, and related entities omitted
counterparties added after the initial request
directors, witnesses, experts, insurers, or funders not considered when relevant
historical matters described too vaguely to understand the relationship
closed files held in a separate archive
contact and matter records containing duplicates
records inaccessible to the person running the search
private knowledge trapped in a partner’s memory or inbox
staff searching different systems with different rules
false positives returned without useful context
possible matches dismissed because the names are not exact
confidential matter details exposed unnecessarily during review
urgent requests bypassing the normal approval path
verbal clearance not recorded
a cleared matter changing parties without a new check
no evidence of which names, systems, dates, and rules were searched

AI cannot compensate for an incomplete client and matter record. It can expose gaps, prepare better queries, and make review more consistent, but the firm’s data and governance still determine the ceiling.

A structured AI Intake Assistant for Law Firms can improve the information entering the process. The conflict workflow then needs its own permissions, review standards, and stop conditions.

Measure the annual conflict-check bleed

A value case should include capacity and risk controls without pretending that every risk can be converted into a precise rand figure.

Measure:

conflict requests per month by practice area and office
requests returned because information was incomplete
average intake and search-preparation time
systems and archives searched manually
lawyer and risk-team review time
urgent checks outside normal working hours
duplicate candidate records reviewed
false positives by cause
late-discovered party or relationship changes
matter-opening delays
prospective clients lost during avoidable waiting
internal interruptions and partner-wide messages
searches repeated because evidence was not preserved
corrections to names, entities, and relationships
checks that required remediation after matter opening
records that could not be searched reliably
time spent creating review and audit evidence
incidents or near misses caused by incomplete information or process bypass

Do not price the project from the number of prompts or hours needed to configure software. Start with the annual operational bleed and the value of a more consistent, evidence-backed gate.

The paid AI Opportunity Audit maps the real process, access boundaries, data quality, approval authority, failure modes, and first safe pilot before any AI employee is trusted with live clearance work.

Map the current conflict-check journey

Take a representative sample: a simple individual client, a corporate group, litigation with many parties, a property transaction, an urgent instruction, a declined matter, and a check that produced many possible matches.

Map each step:

Who requests the check?
What information is mandatory before search begins?
How are people, entities, groups, trusts, and relationships represented?
Which identifiers are collected?
Which name variations must be searched?
Which databases, matter systems, archives, documents, and institutional knowledge sources are approved?
Who may access each source?
What information may be shown to the requester?
How are candidate matches ranked?
Who investigates identity?
Who assesses conflict relevance and professional duties?
What creates an automatic stop or escalation?
Who may clear the request?
How is the decision recorded?
When is a waiver or information-barrier process considered?
Who controls those steps?
What prevents premature matter opening?
What triggers a recheck after opening?
How are new parties added?
How is sensitive evidence retained or restricted?

The process map should distinguish search completeness, identity matching, relationship interpretation, professional analysis, and final approval. Combining them into a single “clear” button hides where errors happen.

Build the Company Brain behind consistent checking

A generic model does not know the firm’s approved search universe, entity rules, confidentiality boundaries, reviewer roles, escalation criteria, or decision policy.

A Company Brain for conflict checks can hold:

request types and mandatory fields
person and entity naming standards
approved identifier types
relationship taxonomy
practice-specific party checklists
approved search systems and archives
query-generation rules
common name-variation patterns
exact, close, phonetic, and historical-name matching thresholds
rules for groups, subsidiaries, trusts, and partnerships
source hierarchy
access and confidentiality controls
automatic stop conditions
reviewer and approval matrix
evidence-pack format
recheck triggers
matter-opening controls
quality metrics
known failure cases
approved examples and reviewer corrections

The Brain does not contain a universal answer to legal conflicts. It contains the firm’s governed operating process for preparing and reviewing them.

The firm should own that process knowledge and improvement history in a readable, exportable form. Vendor models are replaceable. The firm’s approved rules, records, and decisions are the durable asset.

Improve intake before improving search

A sophisticated search cannot find an entity the requester never names.

The intake should collect what is appropriate for the matter type, which may include:

prospective client legal name
identity or registration number where lawful, necessary, and authorised
trading, former, maiden, abbreviated, or alias names
related entities and group structure
directors, trustees, partners, members, or controlling people
opposing and interested parties
co-parties
witnesses and experts where relevant
insurers, lenders, funders, or other participants where relevant
existing advisers
matter type and concise description
jurisdictions and locations
dates or historical context
urgency and reason
requesting lawyer and responsible partner

The form should be dynamic. A corporate transaction, family matter, estate, conveyancing instruction, employment dispute, and litigation file do not need the same relationship map.

The AI can ask: “You listed Ubuntu Trading as the counterparty. Is that a registered company, a trading name, or both? Please provide the registered entity and registration number if available.” It should not guess the answer from a website and silently make it part of the official request.

Normalise names without erasing source truth

Name normalisation improves recall, but the original supplied value must remain visible.

The assistant may prepare variations such as:

full name and initials
surname spacing and punctuation variants
common title removal
legal entity suffix variants
old and new company names
trading and registered names
hyphenated and unhyphenated forms
common transliteration variants
reordered names where convention allows
approved phonetic candidates

Each generated variation should show why it exists. An aggressive fuzzy match can produce an unmanageable queue, especially for common names. A narrow exact match can miss relevant history.

The system should rank candidates rather than hide them. Useful signals can include registration or identity number, full legal name, address, contact details, associated entities, people, matter context, dates, and source quality.

Matching confidence is not conflict relevance. Two records may clearly refer to the same company while the professional question remains unresolved. Conversely, a weak identity match may still deserve review because the potential consequence is serious.

Search approved sources and preserve evidence

A useful result must show what was searched and where a candidate came from.

The evidence record can include:

request reference
search date and time
names and identifiers supplied
generated search variations
approved systems and collections searched
unavailable or failed sources
query version
candidate record identifiers
source links or references
matched fields
conflicting fields
relationship summary drawn from approved records
access restrictions
reviewer notes
decision and authorised approver
recheck requirements

The assistant should never fill an inaccessible source with a claim that “no match was found”. “Archive unavailable during search” is materially different and should block or escalate according to firm policy.

A defensible candidate result might say:

Possible match: Ubuntu Holdings (Pty) Ltd, registration number ending 482, appears as a former client in matter record COM-2023-117. The supplied counterparty is “Ubuntu Holdings”, with no registration number. Exact entity identity is unconfirmed. Restricted relationship details available to the authorised reviewer.

This preserves uncertainty, source, and confidentiality boundaries.

Control false negatives and false positives

A weak conflict workflow optimises only for speed. A safe workflow tests both kinds of failure.

False negatives can arise from:

missing parties
incomplete aliases
inaccessible archives
poor historical records
spelling variation
entity restructuring
relationships stored only in documents or people’s memory
search rules that are too strict

False positives can arise from:

common names
overly broad fuzzy matching
duplicated contacts
unrelated entities sharing words
stale or incorrect records
weak context in old matter descriptions

The answer is not simply to return more names. It is to improve intake, use layered matching, show evidence, group duplicates, and route uncertain candidates to a qualified reviewer.

The pilot should use known historical cases to test whether the assistant surfaces the candidates humans considered important. It should also measure the review burden created by noise.

Keep confidentiality inside the review path

Conflict checking can itself expose sensitive information: the existence of a client relationship, matter description, adverse party, internal concern, or restricted representation.

The workflow should apply:

role-based access
minimum necessary result disclosure
restricted matter flags
separate requester and reviewer views
secure storage and transfer
encryption
purpose-limited processing
provider and operator controls
activity logs
retention rules
tested matter and client separation
incident response
review before external communication

The requester may need to know that a result requires risk review, not the confidential details behind it. The assistant must respect the same need-to-know boundaries as the firm’s human process.

POPIA, professional duties, confidentiality, privilege, and firm risk policy are related but not interchangeable. The implementation should be reviewed for the firm’s actual obligations by appropriate professionals.

Keep the final decision human and attributable

The review pack should help the authorised professional decide. It should not make the decision look automatic.

A useful pack can contain:

scope and completeness of intake
names and relationships searched
sources searched and unavailable sources
candidate matches grouped by confidence
source-linked relationship evidence
identity questions still open
new-party or recheck requirements
restricted information controls
required reviewer
a clear place for questions, decision, conditions, and approval

The decision record should name the authorised approver, time, outcome, conditions, and evidence version. If the matter changes, the original decision remains part of the history and a new check can be opened.

No AI-generated explanation should be treated as professional clearance merely because it sounds confident.

Launch in shadow mode against known outcomes

Do not begin by allowing an AI assistant to clear live matters.

A controlled pilot can:

select one practice area or request type
define mandatory intake fields and approved sources
use previously reviewed requests with known results
test name variations and relationship extraction
compare candidates with the historical human review
investigate false negatives and false positives
test restricted-access behaviour
run the assistant in parallel on selected new checks
require full human search and approval during the pilot
record corrections and update only approved rules

Measure:

intake completeness
missing-party prompts accepted
known candidate recall
false positives per request
identity-resolution accuracy
source-link accuracy
failed or unavailable source detection
preparation and review time
confidentiality-control failures
escalation precision
user adoption
decision evidence completeness

A 30-day working interview should prove that the assistant improves preparation and visibility before any expansion in scope or permissions.

Monitor changes after matter opening

A conflict check is not always a once-off event. Parties, directors, advisers, witnesses, funders, transaction structures, and claims can change.

The workflow can create controlled recheck triggers when:

a new party is added
an entity name changes
a counterclaim or third-party notice appears
a new witness or expert becomes material
a transaction structure changes
a related entity enters the matter
the client scope expands
the responsible lawyer requests a recheck
firm policy sets a review milestone

The assistant may detect and prepare the change. An authorised human decides the appropriate response.

What a useful monthly report shows

Management does not need a vanity count of AI searches. It needs evidence that the gate is becoming more reliable.

A useful report can show:

checks opened and completed
turnaround by request type
incomplete requests
urgent requests
candidate matches requiring review
source failures
common false-positive causes
corrections to names and relationships
rechecks triggered
process bypass attempts
decision records missing evidence
staff time saved conservatively
Company Brain rules improved
unresolved data-quality work

That report turns operational experience into an owned learning loop rather than allowing the same search problems to repeat.

Start with controlled preparation, not automatic clearance

The goal is not a faster green tick. The goal is a more complete, consistent, source-backed conflict-check process that protects clients, the firm, and the professionals making the decision.

BizSage installs AI employees for law firms around defined administrative work, strict permissions, professional approval, monitoring, and a firm-owned Company Brain.

If conflict checks depend on inconsistent forms, private memory, repeated searches, and unclear evidence, start with the AI Opportunity Audit. We will map the actual process, quantify the operational bleed, test the data and access boundaries, and define a narrow pilot without handing professional judgement to a machine.

Frequently asked questions

What does an AI conflict check assistant do?

It structures the request, prepares name variations, searches approved records, surfaces possible matches with sources, and routes the evidence to the responsible reviewer. It makes preparation more consistent; it does not provide clearance.

Can AI clear a legal conflict automatically?

It should not. Identity, relevance, duties, confidentiality, waivers, information barriers, and acceptance require authorised professional judgement under the firm’s policies and obligations.

Can it find every possible match?

No. Search quality depends on complete intake, reliable historical records, available sources, permissions, and tested matching rules. A responsible system reports unavailable sources and uncertainty rather than promising perfect coverage.

What should a law firm test first?

Start with historical checks that have known outcomes, then run selected live requests in shadow mode. Measure candidate recall, false positives, source traceability, review time, confidentiality controls, and escalation quality before changing any approval process.