How to Write a Demand Letter with Claude AI (2026)
A five-stage attorney protocol for drafting settlement demand letters with Claude AI: sanitized inputs, valuation framing, 17 sections, and verification.
Key Takeaways
- The core idea: Settlement demand letters succeed or fail on structure, not prose. Missing citations, unexplained treatment gaps, and unsupported valuations signal weakness before negotiation begins.
- Why it matters: Attorneys must sanitize all client data before AI interaction. Substitute roles for names, intervals for dates, and maintain a secure mapping document outside the session.
- How it works: The model drafts and organizes, the attorney supplies every fact, figure, and citation. This division of labor creates defensible output while reclaiming hours spent on manual assembly.
- What to do: Verification gates prevent fabricated citations and unsupported claims. Every authority must trace to a primary source you confirmed before the letter leaves your control.
Source: The Leveraged Years Briefing. Permalink
The short answer: Claude can draft a complete settlement demand letter in an afternoon, and the draft is worth exactly as much as the protocol around it. The working method has five stages: sanitize the file, build a structured case file, frame the valuation with your own numbers, draft the letter section by section in your firm's voice, and run a verification gate before anything is sent. The attorney supplies every fact, every figure, and every citation. The model structures, drafts, and stress-tests. That division of labor is what makes the result defensible.
This article walks through each stage the way we teach it to practicing attorneys. No tool can replace the judgment that decides what a claim is worth or when to hold a number. What a disciplined workflow replaces is the four hours a paralegal spends assembling a chronology by hand, and the generic letter an adjuster skims and discounts.
Why the demand letter rewards structure more than any document you send
A demand letter is the anchor for every number that follows it. Adjusters and defense counsel read hundreds of them, and they sort fast: complete, specific, and supported gets a real evaluation; generic gets a formula response. The three things that most often cost claimants money are missing or unverifiable citations, treatment gaps left unexplained for the other side to weaponize, and a demand figure with no visible method behind it.
All three are structural problems. Structure is what a well-run AI workflow is good at, provided you keep the judgment where it belongs: with you.
Stage one: sanitize before anything enters the session
This stage is not optional, and it comes first because everything else depends on it. Client files, and plaintiff-side files especially, contain the most sensitive material a practice handles: medical records, treatment notes, identifiers everywhere. Real records identifying a real client never go into a general AI session, under any circumstance.
The workable discipline is substitution. The client becomes [Claimant]. Providers become roles: [ER Physician], [Treating Orthopedist]. Calendar dates become intervals: day 1, day 24, week 6. You keep a key document inside the secure client file, outside the AI session, that maps each placeholder back to its identity, so the output reconciles with the source file later. Five extra minutes, and you have a session you could describe to your malpractice carrier without flinching.
Stage two: build the case file before you draft anything
Most people open an AI tool and ask for the letter. That is backwards. The letter is the last artifact, not the first. Start by having Claude convert your sanitized narrative into a structured intake sheet: incident, parties by role, liability factors, injuries, damages signals, coverage, jurisdiction. Then have it list what is missing as direct questions. That missing list is a file-completeness audit most firms never run this early.
Next, the treatment chronology. Give it the sanitized treatment events and ask for a dated table with every gap in care flagged, every provider transition noted, and every inconsistency between the records and the narrative surfaced. A six-week gap the defense discovers is a discount. The same gap found now, with the innocent explanation written down, is a managed risk you address on your own terms in the letter.
Stage three: frame the valuation with your numbers, not the model's
Here is the line that matters most. Claude has no database of verified settlements, and a number it invents is worth nothing. Never negotiate from one. What it does well is the frame: laying out your specials, applying the multiplier arithmetic you choose, holding the comparable outcomes you researched and verified yourself, and showing how the number moves when an assumption moves.
The output to aim for is a one-page valuation frame: specials by category, a general damages range with your stated rationale, your comparables alongside, and a target, floor, and ceiling that you decide and can explain to the client without mentioning the tool. If you cannot explain the number without the tool, it is not your number yet.
Stage four: draft the seventeen sections in passes, in your voice
Complete demand letters answer every question the carrier's evaluation asks, in roughly the order it asks them. We teach a seventeen-section map: parties and coverage, liability narrative, injury chronology, treatment summary, medical specials, wage loss, future care, pain and suffering, a pre-rebuttal section that answers the weaknesses you mapped before the adjuster can use them, comparables support, the demand with its deadline, an exhibits index, and the structural sections around them. Adjust the map to your practice; the discipline is that your whole firm drafts to one map, so no letter goes out missing its strongest section because someone was in a hurry.
Draft in passes, not in one shot: facts and liability first, damages second, pre-rebuttal third, the demand last, correcting each pass before the next so an error in section two never echoes through section fourteen. Give the model two of your prior letters, fully sanitized, and it will pick up your sentence rhythm and formality level instead of sounding like a template.
And one absolute rule: the model never adds authority. You supply the citations you verified; it weaves them in. A public database has tracked more than 1,300 court filings containing fabricated AI citations. In a demand letter, a citation the adjuster looks up and cannot find does not just weaken the argument. It ends your credibility for the rest of the negotiation.
Stage five: stress-test it, then verify everything before it leaves
Before the letter goes out, make Claude switch sides. Have it review the draft as a veteran claims adjuster: where the letter is strong, what gets discounted and why, and what would move the evaluation. Every weakness the simulated reviewer finds is a free revision. The same session can rehearse the negotiation itself: the likely lowball, your counter anchored back to the demand and the record, the concessions you are willing to make in decreasing size, and the floor where you stop.
Then the gate. Every citation checked against the primary source. Every figure checked against the file. Every record reference confirmed against a document you can produce. This is not bureaucracy; it is the supervision the rules of professional conduct already require of you, documented. The letter that clears the gate is yours in every sense that matters, whoever typed the first draft.
When buying software is the better answer
There is a growing market of purpose-built demand platforms, and an honest protocol includes knowing when to use one. Tools in this category offer verified citation databases, valuation models trained on settlement outcomes, and managed negotiation tracking, at per-demand fees that run from roughly 125 to 800 dollars across the market. If your firm sends dozens of demands a month, evaluating one is rational. If you send a handful, the workflow above produces comparable structure at the cost of your time, and it is portable: the method works on employment settlements, commercial disputes, and any matter that ends in a negotiated number.
Either way, learn the method first. Attorneys who understand the workflow underneath these platforms negotiate better with the vendors, supervise the output properly, and never mistake a tool's number for a valuation.
Where to go deeper
We teach this exact workflow, from the sanitization protocol through the adjuster simulation and the verification gate, as a four-lesson capstone case study inside The Leveraged Attorney, our course for practicing attorneys who want a repeatable, defensible way to work with Claude. The course covers the full system: the clean room, contract review, research organization, drafting support, and the supervision protocols that make all of it safe to use on real matters. You can also see documented examples of professionals running on AI across other fields.
Common questions
Can Claude AI write a legal demand letter?
Yes, it can draft a complete letter from attorney-supplied facts. The draft is only as defensible as the protocol around it: sanitized inputs, attorney-supplied figures and citations, and full verification before sending. Claude structures and drafts; the attorney decides.
Is it safe to put medical records into an AI tool?
Real records identifying a real client should never go into a general AI session. Work from a sanitized fact summary, with roles instead of names and intervals instead of dates, and keep the mapping key in the secure client file.
Will the AI invent case law in my demand letter?
It can, which is why the rule exists: the model never adds authority. You supply verified citations, and you check every one against the primary source before the letter goes out. More than 1,300 court filings with fabricated AI citations have been tracked publicly. Do not join them.
Should my firm buy demand letter software instead?
At high volume, possibly: verified citation databases and outcome-trained valuation are real advantages, priced per demand. At lower volume, the disciplined workflow gets you comparable structure for the cost of your time. Learning the method first makes you a sharper buyer either way.
This article is general information about workflow, not legal advice and not CLE. Attorneys remain responsible for confidentiality, supervision, and every document they sign.
Frequently Asked Questions
Can I trust Claude to draft demand letters without inventing case law?
Only if your protocol forbids the model from adding authority. You supply verified citations, the model incorporates them, and you check every reference against the primary source before sending. The tool structures; you decide and verify.
How do I protect client confidentiality when using AI for demand letters?
Sanitize every input before it enters the session. Replace names with roles, dates with intervals, and keep a secure key document that maps placeholders to real identities outside the AI environment.
Should I invest in specialized demand letter software or use a general AI workflow?
If you send dozens monthly, purpose-built platforms with verified citation databases may justify their per-letter cost. For occasional use, a disciplined Claude protocol delivers comparable structure at the cost of your time, and learning the method sharpens your evaluation of paid tools.