A liability file can look settled until the sources are placed side by side. The police report attributes the impact to one maneuver, a witness describes another, and the client’s intake account uses language that creates a third version. If an AI-assisted demand simply blends those accounts into a smooth narrative, the draft may hide the exact conflict an adjuster will isolate.
The better workflow is not to ask software which source is “right.” It is to preserve each source, identify material disagreements, and route those disagreements to the attorney before the liability section is drafted.
Conflicting sources are a review problem, not a writing problem
Liability narratives often draw from records created for different purposes and under different conditions. A responding officer may summarize statements at a chaotic scene. A witness may have had a better angle but only a partial view. The client may remember more detail during a later interview, while also using imprecise terms such as “merged,” “turned,” or “came into my lane.” Photographs, diagrams, vehicle damage, video, and event-data records may support one account without answering every disputed point.
These are not interchangeable inputs. A police report may contain the officer’s observations, a party’s account, a witness quotation, and a conclusion in the same document. Under California Evidence Code section 1200, an out-of-court statement offered for its truth raises hearsay issues, while exceptions such as the business-records framework in section 1271 have their own foundations and limits. The demand-stage question is not merely whether a document exists. It is what the document actually establishes, who supplied each fact, and whether another source contradicts it.
An adjuster does not need a complete alternative theory to challenge a demand. One inconsistency about lane position, signal color, speed, right of way, or the sequence of impact can become the center of the response. A polished paragraph that silently resolves the inconsistency may therefore be less useful than a structured conflict note that forces early attorney review.
Build a source map before generating the liability narrative
An AI workflow should preserve source boundaries before it produces prose. That means treating every liability proposition as a claim tied to evidence, not as a free-floating fact. A practical source map can use four fields:
- Proposition: the factual point that matters, such as which vehicle entered the intersection first.
- Source: the report page, statement, photograph, video segment, diagram, or intake note supporting it.
- Source type: direct observation, quoted statement, later recollection, physical evidence, or attorney inference.
- Status: corroborated, disputed, incomplete, or unresolved.
This structure separates extraction from advocacy. The system can identify that the report narrative says the defendant changed lanes, while a witness says the defendant was already established in the lane. It can also flag that the client’s account does not address the timing point at all. What it should not do is convert silence into agreement or select the most favorable version without instruction.
Police reports require proposition-level parsing
“The police report says” is usually too broad. The officer may personally observe final resting positions and damage, summarize two competing accounts, reproduce a witness statement, and then list a primary collision factor. Each component has a different source and a different role in the attorney’s analysis.
For AI-assisted drafting, the report should be broken into propositions with page references. If a conclusion depends on a quoted party statement, that dependency should remain visible. If the diagram conflicts with the narrative, the workflow should flag the mismatch rather than choosing one. Supplemental reports should be linked to the original rather than treated as standalone replacements.
Witness statements need vantage-point context
A witness statement gains or loses significance through context: position, line of sight, distance, attention, lighting, traffic conditions, and when the statement was taken. A short statement that appears definitive may address only the moment after impact. Another witness may describe pre-impact movement but not see the signal phase.
The workflow should distinguish a direct quotation from a staff summary and preserve uncertainty language. “I think,” “approximately,” and “I only saw” are not clutter to be cleaned up. They are part of the source. AI can help normalize formatting, but it should not upgrade qualified observations into categorical facts.
Client accounts should remain versioned
The client’s intake account, later interview, recorded statement, deposition preparation notes, and signed declaration should not be collapsed into one composite memory. Changes may be innocent clarifications, differences in questioning, or meaningful inconsistencies. The attorney needs to know which is which.
A versioned account shows the date and context of each statement, the exact proposition that changed, and any new evidence that explains the change. This is narrower than the broader recorded-statement workflow discussed elsewhere on the site. For a related treatment of issue classification, see AI for Demand Letter Issue Spotting: Separating Liability, Causation, and Damages Problems.
Use a conflict protocol instead of silent reconciliation
Once the source map is complete, the system needs a repeatable escalation rule. Not every wording difference is material. The point is to surface conflicts that could change liability allocation, credibility, or the demand’s factual framing.
- Normalize terms without changing meaning. Map obvious synonyms, but retain the original language beside the normalized label.
- Compare propositions, not entire documents. A report and a statement can agree on impact location while disagreeing on pre-impact movement.
- Classify the conflict. Use categories such as direct contradiction, missing fact, timing ambiguity, source-within-source, or physical-evidence mismatch.
- Measure materiality. Escalate disagreements about right of way, lane position, signal phase, speed, visibility, distraction, admissions, and impact sequence.
- Request attorney disposition. The attorney can choose to investigate, obtain a supplemental statement, narrow the narrative, acknowledge the dispute, or reserve the point.
- Preserve the decision trail. Record which source the final narrative relies on and why, without overwriting contrary material.
The output should be a conflict table or review memo before it becomes demand prose. A useful entry might say: “Report narrative attributes the lane change to the defendant; independent witness states both vehicles maintained lane position; damage photographs do not independently resolve the movement.” That is operationally useful because it identifies the gap without pretending the software can decide credibility.
Draft only after the attorney sets the liability theory
After review, AI can help produce a liability section that is both persuasive and supportable. The draft should cite or describe the sources actually selected, avoid overstating ambiguous evidence, and keep unresolved issues out of categorical sentences. If the attorney decides the conflict should be acknowledged, the language can explain why the total evidence still supports the firm’s position without mischaracterizing the contrary account.
Before approval, the reviewer should ask:
- Does every material liability statement trace to an identified source?
- Did the draft preserve qualifications and uncertainty from the source?
- Were police-report conclusions separated from the underlying statements?
- Were witness vantage point and observation limits considered?
- Were changes across the client’s accounts surfaced rather than merged?
- Does the final narrative reflect an attorney decision, not an automated credibility choice?
This review also protects work-product discipline. AI-assisted drafts remain attorney work product when prepared in that context, but the attorney remains responsible for accuracy, strategy, and the decision to send. Firms should apply their confidentiality, access-control, and vendor-review policies to every source uploaded for processing.
How Legal Power AI fits
Legal Power AI’s demand-letter workflow is designed to organize source material, surface review points, and support attorney-directed drafting rather than replace liability judgment. Legal Power AI processes data with BAAs from OpenAI and Anthropic, and the attorney remains responsible for confirming the factual record and approving the final demand.
Conflicts should become visible before they become credibility problems
A clean demand is not the same as a trustworthy demand. When police reports, witness statements, and client accounts diverge, the workflow should make that divergence easier to examine—not easier to hide. Source-level extraction, proposition-based comparison, materiality rules, and a documented attorney disposition create a stronger drafting process. The result is not automatic truth-finding. It is a liability narrative built from evidence the attorney has consciously reviewed and chosen to rely on.
See how source-aware demand drafting works in practice.