AI Document Review for British Columbia Lawyers

AI can help British Columbia lawyers sort records, build chronologies, compare accounts, surface possible contradictions, and prepare questions for human review. It should not be treated as the record, a privilege decision-maker, or a substitute for counsel's analysis. A defensible process defines the legal and factual review question, controls the document set, preserves a reference to every source page, separates extraction from inference, and sends material findings back to the original document before they are used.

Key facts and limits

Define the review before selecting a tool

“Review these documents” is not a review protocol. Start by identifying the proceeding, stage, issues, document population, custodians or sources, relevant dates, desired output, and decisions that remain with counsel. A chronology review needs reliable dates and event links. A production review needs scope, responsiveness, privilege, and quality controls. A witness-preparation review needs exact page or line references and balanced treatment of confirming and conflicting passages.

Define exclusions as carefully as inclusions. State whether the collection omits attachments, images, handwritten notes, password-protected files, audio, corrupted files, or records outside the agreed date range. Record OCR limitations and language requirements. An output cannot be evaluated honestly if nobody knows which material the system could not read.

Use a review charter

• Question: the factual or workflow question the review must support.
• Population: the authoritative source set and how completeness is reconciled.
• Categories: agreed labels with examples and escalation rules.
• References: file name plus page, paragraph, line, exhibit, or other stable locator.
• Human decisions: privilege, responsiveness, legal significance, strategy, and final reliance.
• Quality sample: how misses and false positives will be measured and corrected.

Keep the technology workflow aligned with Rule 7-1

For actions governed by the Supreme Court Civil Rules, Rule 7-1supplies the current discovery-and-inspection framework. As consolidated to July 28, 2026, subrule (1) generally requires each party of record, unless all parties consent or the court otherwise orders, to prepare and serve a Form 22 list within 35 days after the end of the pleading period. The list includes documents that are or have been in the party's possession or control and that could, if available, be used by any party of record at trial to prove or disprove a material fact, plus other documents the party intends to refer to at trial.

The rule also matters to review design. Subrules (6) and (7) address privilege claims and require a description that permits the claim to be assessed without revealing privileged information. Subrule (9) requires prompt amendment when a list is inaccurate or incomplete or a qualifying document later comes into possession or control. Subrules (10) through (14) address demands, responses, and court powers concerning omitted or additional documents. Subrules (15) and (16) address inspection and copies, while subrule (21) states a consequence for failing to discover or produce a document as required by the rule, unless the court orders otherwise.

A seven-step AI document review workflow

1. Freeze and inventory the working set. Preserve the authoritative collection, assign stable identifiers, detect duplicates without silently deleting variants, and log unreadable or unsupported files.
2. Separate materials by sensitivity and purpose. Keep privileged, potentially privileged, confidential, personal, settlement, and public materials within the access model approved for the matter.
3. Run a narrow extraction pass. Ask for observable items—dates, people, communications, document types, quoted passages, and source locators—before requesting narrative conclusions.
4. Build structured review views. Organize candidates into a chronology, issue table, witness map, or contradiction queue. Preserve the exact source link beside each entry.
5. Escalate legal and ambiguous decisions. Route possible privilege, scope uncertainty, conflicting dates, unclear authorship, and material gaps to a lawyer rather than forcing a confident label.
6. Sample and correct. Review both included and excluded material, test high-risk categories separately, record error patterns, and revise the protocol when the sample reveals systematic misses.
7. Reconcile and sign off. Compare processed counts with the collection, confirm exceptions were resolved, and have the responsible lawyer approve the work product actually used.

Adapt the review to the document type

A single instruction should not govern every record. Each document category carries a different source structure and a different risk of losing context. The review charter can share core controls—stable identifiers, source locators, exception logging, and human escalation—while changing the fields and quality sample for each group.

Attachments and document families deserve particular attention. A short email may be intelligible only with the spreadsheet it forwards; a report may rely on an appendix; a transcript exhibit may sit in a separate file. Preserve those relationships in the review view and sampling plan. If the system cannot maintain them, record the limitation and use a different process for the affected group.

Language and format also affect the protocol. Scanned pages, handwriting, diagrams, stamps, tracked changes, comments, embedded files, and password protection can create silent gaps. Identify those formats during inventory, test representative files before the full run, and route unsupported material for manual or specialized review. Translation is another layer: preserve the original wording, identify the translation method, and avoid treating an automated translation as the authoritative quotation.

Worked example: building a source-linked chronology

Assume a BC civil-litigation team has pleadings, correspondence, invoices, interview notes, and examination transcripts concerning a disputed project delay. The immediate job is to build a chronology for counsel's review, not to decide liability or prepare a final list of documents.

1. Define the event schema

Use fields for date, time if material, actor, event, document identifier, page or line, quoted support, confidence, and reviewer note. Permit “date unclear” instead of encouraging the system to infer a precise date from surrounding text. Keep document date, event date, and received date separate.

2. Extract candidates in batches

Process controlled batches and require a locator for every row. Capture extraction failures. Near-duplicate emails may contain different attachments or recipient lists, so grouping should not erase the variant needed to understand the record.

3. Compare, do not adjudicate

Ask the system to place passages concerning the same event side by side. A later recollection that differs from a contemporaneous email is a review candidate, not automatically a contradiction or a credibility finding. Counsel reads the full passages and decides their significance.

4. Verify material entries

The final chronology should link every material entry back to the record and distinguish direct quotation, paraphrase, inference, and unresolved conflict. That structure lets counsel test the account instead of inheriting an opaque narrative.

Privilege, confidentiality, and minimum necessary access

The Law Society's AI materials expressly identify confidentiality and information security as professional-responsibility issues. Before matter documents enter a system, review the provider, terms, retention, access, administrative controls, data use, incident process, and the firm's policy. Use the minimum necessary data for the task and restrict the people and tools that can access the workspace.

Privilege review needs a separate decision path. AI may help route documents containing counsel names, legal-advice language, or litigation-related patterns, but those signals do not establish privilege and can miss privileged material. Conversely, the presence of a lawyer's name does not settle the issue. Have a lawyer review candidates, record the basis for the decision, and prepare any description required by Rule 7-1 without disclosing the protected information.

Quality control that tests the review

A polished summary is not a quality metric. Quality control should test traceability, coverage, and decision consistency. Sample documents the system included and excluded. Oversample high-risk groups such as possible privilege, handwritten pages, poor scans, tables, attachments, and long transcripts. Track false positives, false negatives, missing locators, OCR errors, and unsupported inferences.

When a sample exposes a repeatable error, correct the workflow and rerun affected material; do not simply fix the sampled row. Preserve the model or tool version, instructions, collection identifier, run date, exceptions, and reviewer so the result can be explained and updated.

Design the handoff for the next lawyer

Review output should be understandable without reconstructing the entire run. State the question, collection boundary, categories used, exceptions, sampling method, known limits, and date of review. Give each material finding a stable source locator and identify whether it was extracted, inferred, or confirmed by a reviewer. Separate unresolved items from completed findings so uncertainty is not flattened during drafting. If the collection changes, record what was added and which portions of the analysis were rerun. This handoff discipline matters when responsibility moves between associates, counsel, experts, or trial teams: a later reviewer should be able to open the source, understand the decision path, and see what remains unchecked. It also makes proportional updating possible because the team can target the affected source set instead of treating every prior output as current.

A verified Curia workflow

Curia's current product implementation supports matter document uploads and a matter Analysis stage that reviews current matter documents. The analysis can surface a timeline and cross-document findings such as corroborations and contradictions, retains document context for follow-up questions, and provides a separate citation-check action when Canadian case citations are found in matter documents. These are review aids; they do not decide privilege, production scope, credibility, or the legal effect of a record.

Common errors and edge cases

Letting the summary replace the collection

A summary compresses context. Preserve the source set and require stable references so reviewers can return to the complete item.

Asking for legal labels too early

Extract observable features first. Responsiveness, privilege, admissibility, credibility, and legal significance require their own analysis.

Ignoring failed files and attachments

A processed count can look complete while attachments, scans, or encrypted files were skipped. Reconcile at both file and family level.

Accepting invented precision

If a date, speaker, or page cannot be established, the output should preserve uncertainty. Plausible detail is not evidence.

Testing only the results the system selected

Review excluded material too. Otherwise the team measures obvious false positives but never discovers important misses.

Practical checklist

Before review

• Confirm the proceeding, governing requirements, review question, source population, and output.
• Approve the tool and access model for the sensitivity of the material.
• Preserve originals, stable identifiers, attachments, and an exception log.
• Define categories, examples, escalation rules, and the human decision owner.

During review

• Require a source locator for every material extraction.
• Keep observed text separate from inference and lawyer analysis.
• Escalate privilege, ambiguity, material conflicts, and missing context.
• Sample included and excluded documents, with extra attention to high-risk formats.

Before relying or producing

• Return every material finding to the original document and read it in context.
• Reconcile file counts, families, failures, and later-added material.
• Apply current Rule 7-1, orders, agreements, and matter-specific legal judgment.
• Have the responsible lawyer approve the final work product and record unresolved limits.

Key takeaways

• Start with a review charter and authoritative collection, not a broad prompt.
• Use AI for classification, extraction, comparison, and organization—not final legal decisions.
• Preserve source locators and send every material finding back to the original record.
• Keep privilege and confidentiality within a controlled, lawyer-supervised process.
• Test exclusions and failures, not only the results the system selected.
• Align production work with the current rule, court orders, agreements, and the actual file.

Primary and authoritative sources

• BC Laws: Supreme Court Civil Rules, Rule 7-1 — Discovery and Inspection of Documents (consolidation current to July 28, 2026 when checked).
• Law Society of British Columbia: Guidance on Professional Responsibility and Generative AI.
• Law Society of British Columbia: Artificial intelligence and the legal profession.
• Law Society of British Columbia: Code of Professional Conduct for British Columbia.

Related Curia resources: AI legal research in British Columbia, safe AI uploads for lawyers, and AI document review for Ontario litigation.