Added on: 07 August, 2026
AI demand letter generators have earned their place in the personal injury workflow — and the platforms building them are refreshingly candid about what they actually deliver. Ask most AI demand generation tools how finished their output really is, and the honest answer is somewhere in the range of 85–90% ready. Not 100%. Not "send it as-is." Roughly nine-tenths of the way there, with a specific, predictable stretch of work still required before it's ready for an adjuster's desk.
That last 10–15% is not a rounding error. It's where accuracy, tone, and case-specific judgment live — and it's exactly the gap this article is about.
What "85–90% Ready" Actually Means
The 85–90% figure shows up consistently across the AI demand letter drafting category, described in slightly different language by different platforms. One industry analysis puts it plainly: the attorney receives a document that's roughly 90% complete and ready for review rather than a blank page once the AI has processed the case records and generated a structured first draft. Another puts the split even more directly: AI handles the first 80% of the draft, and the attorney's judgment adds the remaining 20% — different phrasing, same underlying reality.
The point isn't that AI demand letter tools are unreliable. It's that even the platforms themselves frame their output as a strong first draft, not a finished work product. Extraction accuracy claims illustrate why: one industry comparison notes that a platform's extraction accuracy below 95% still means significant time correcting errors, and even well-regarded platforms report meaningful accuracy gaps — one AI demand platform reports its own service-date mapping model at 90% accuracy, notably ahead of general-purpose models like GPT-4, which the same source measured at 68% on the same task.
Ninety percent accuracy sounds impressive until it's translated into dollars. As one industry source frames it, missing just 20% of billing records across 100 cases, at an average impact of $5,000 each, can total over $100,000 in lost claim value per attorney — a concrete illustration of what the unreviewed 10–15% can actually cost.
What the Remaining 10–15% Actually Involves
Treating "85–90% ready" as a finish line rather than a checkpoint is where AI-generated demand letters most often go wrong. In practice, the review step that closes that gap covers three specific categories of work:
Reviewing for Factual Accuracy
Every damages figure, treatment date, provider reference, and cited authority in an AI-generated draft needs to trace back to the actual source record — not be accepted because it reads convincingly. This is where hallucinated citations, unrelated visits folded into the treatment timeline, and unflagged prior injuries most often hide, since none of these errors look obviously wrong on a surface read.
Adjusting Tone
AI-generated language can read as generic, formulaic, or inconsistent with a firm's usual voice and negotiating posture. One industry source is direct about the risk of skipping this step: adjusters notice when firms send AI-generated demands without review, and formulaic language or generic narratives weaken the firm's negotiating position. Tone isn't cosmetic in a demand letter — it signals how seriously a firm has engaged with the case.
Filling in Case-Specific Details
AI drafting tools work from the records and data they're given. Nuances that require independent judgment — how hard to push on a disputed liability point, which precedent or argument carries the most weight for a specific adjuster or venue, what context belongs in the narrative that isn't captured in the medical records alone — still need a human layer to complete the picture.
Why This Gap Matters More As Firms Scale AI Adoption
The risk in that final 10–15% doesn't stay constant as a firm's AI-generated caseload grows — it compounds. A firm sending a handful of AI-assisted demand letters a month can plausibly catch errors through informal attorney review. A firm running dozens or hundreds of cases a month through an AI drafting tool is relying on that same informal review process to scale at the exact moment it's least likely to hold up.
This is the structural reason "review before sending" tends to erode under volume: it's a manual step competing against the very efficiency gains that made AI drafting attractive in the first place. Firms that don't formalize the review step don't eliminate the risk in that final 10–15% — they just make it less visible until an adjuster, opposing counsel, or a malpractice review surfaces it.
Closing the Gap Without Losing the Speed
The instinct to skip or shortcut review is understandable — it's the fastest way to keep the efficiency gains AI drafting promised in the first place. But the data above suggests that instinct is exactly backwards: the firms getting real ROI from AI demand letter tools are the ones treating that final 10–15% as a formal, systematic step, not an optional read-through.
That's the specific gap FactAssess' AI Output Audit service is built to close — independent, source-traceable verification of exactly the categories covered above: factual accuracy, source-traceability, and completeness, applied consistently across every AI-generated demand letter or medical chronology, regardless of caseload volume or which AI drafting platform produced the draft.
FAQs
Common questions from this post.
The last 10–15% is where accuracy lives. Make sure it's covered.
Our Services
About Us
Get in Touch with Us!
© 2026 FactAssess. All rights reserved. Privacy Policy