AI Demand Letters for Personal Injury: EvenUp, Supio or ChatGPT?

Choose EvenUp or Supio when your demand volume justifies a quoted contract, because both are built for plaintiff firms and tie their figures to the records; use ChatGPT or Claude only as a drafter on redacted files. Decide by testing one closed case three ways, then check every bill, wage figure and date against a hand-built specials ledger.

The check matters more than the brand. A demand letter lives or dies on its numbers: specials (the medical bills and wage loss you can prove), the treatment dates and the narrative that links them. An AI draft can read beautifully and still carry a duplicated bill or a date outside the injury window, and the lawyer who signs it owns every figure.

Follow me on Instagram@sagnikteaches

What each route is, and what is verified about it

EvenUp

EvenUp's site presents a platform that drafts demands "in minutes" in your firm's tone from templates, and offers a choice between instant AI demands and legal-reviewed demands depending on case type. It also describes turning medical records into interactive chronologies and verifying facts and citations with line-level citations. It describes itself as SOC 2 audited with a health-privacy attestation; ask for the reports before any records go in. There is no public price list (its /pricing address simply returns you to the home page), so the price comes from a sales conversation.

Connect on LinkedInSagnik Bhattacharya

Supio

Supio presents itself as an agentic legal AI platform for plaintiff and mass tort work, with separate products for chronologies, demand letters and a case ledger. Its demand-letter page says the agent learns your templates from your prior demands, that every claim in a demand carries a linked source, and that on high-value cases human experts verify medical bill totals and damage calculations. The case ledger, it says, reads medical bill PDFs into billed amounts, adjustments, insurance payments and balances, removes duplicate line items and links each line to its source document. It states SOC 2 Type II certification and compliance with health-privacy and data-protection rules. There is no public price either: its /pricing address leads to a demo request. Get a quote in writing, and ask for the demonstration on a case file of your own, with consent, rather than the vendor's sample case.

Subscribe on YouTube@sagnikteaches

ChatGPT or another general assistant

A general assistant has no connection to your case management system and no concept of your template unless you provide one. It is the cheap baseline. ChatGPT Plus is listed at $20 a month; business plans have a two-seat minimum and, by default, do not train on business content. Claude Team lists $20 per user a month on annual billing ($25 monthly), also with a two-seat minimum. Individual plans handle your data under different default settings, so check them before any client information goes in. For medical records the safer assumption is that nothing identifiable goes into any tool until the firm has approved it.

If you are still deciding what case software to buy at all, the personal injury case software checklist covers the wider purchase.

Set up the three-way test on one closed case

Use a case that has settled, so that you know the outcome and the full records are in the file. Remove or replace names and identifiers for any tool the firm has not yet approved, and get the client's consent or use a closed matter permitted by your firm's policy. Redacting client records before a general chatbot sees them shows what to strip.

1. Route A: vendor tool. Ask EvenUp or Supio for a trial or supervised demo on the closed case, and ask what they do with the file afterwards.
2. Route B: general assistant. Upload the redacted records to a business-plan chat or project and use the prompt below.
3. Route C: your own template. A paralegal drafts the letter the usual way and records the time it takes. This is your control.

The control matters because the honest question is not "is the AI draft good?" but "is it better than what we would have written, and by how much time?" Without the control you cannot answer.

Build the specials ledger before you look at any draft

The ledger is the answer key. Create it by hand, from the source bills, before you read any AI output, so that you are not anchored by its numbers. Each line needs a provider, date of service, billed amount, amount paid or adjusted where the records show it, and a page reference.

Here is an illustrative ledger for a made-up rear-end collision claim with five months of treatment. All figures are invented.

Check the sum yourself: $1,250 + $4,110 + $1,900 + $2,970 + $750 + $1,260 = $12,240. A ledger with a sum you checked is the thing every AI total will be compared with. The paid-or-adjusted column is left out here to keep the table narrow; in a real ledger it sits beside the billed figure.

The prompt for the general-assistant route

For Route B, hand the assistant the records and the ledger structure, and tell it to show its working. Asking for a table with page references is the single most useful instruction, because it makes errors visible.

Below are redacted medical records and billing statements for a closed
personal injury matter. Do not use any information from outside the documents.

Task 1: Produce a table of every charge: provider, date of service,
amount billed, and the page where it appears. If an amount appears on both an
itemised bill and a summary statement, list it once and note both pages.
Task 2: Total the table and show the arithmetic.
Task 3: List any item where the date of service falls outside the period from
the incident date to the last treatment date, or where the amount differs between
two documents.
Task 4: Only after Tasks 1 to 3, draft a three-paragraph demand narrative
using ONLY items from the table. Mark each factual sentence with its page number.
If you are unsure of a figure, write [CHECK] instead of guessing.

The order of tasks is deliberate. By making the model build and total a table first, you get a checkable intermediate product before any persuasive writing appears. Run the same four tasks, by hand, against the vendor tools' output by asking them for their itemised source list.

Compare the totals line by line

Suppose the general assistant returns a total of $12,990. The ledger says $12,240, so the gap is $750. Find it rather than averaging it away. In this illustration the cause is a $750 pharmacy charge that appears on the itemised pharmacy statement and again inside the clinic's summary invoice. The model listed both. That is a classic double count, and it would have inflated the demand by 6%, a number the other side's adjuster will find in minutes. Supio says its case ledger removes duplicate line items automatically and lets you review what it excluded from totals; in the vendor run, open that exclusions list and check it against your ledger too, because a line removed in error costs the client money.

Three other discrepancy types turn up in real files and are worth testing deliberately:

• Billed versus paid. Records may show a billed amount and a lower amount after a write-down or insurer adjustment. Which figure belongs in the demand depends on the rules that apply to your matter, so the ledger should carry both columns and a lawyer decides.
• Dates outside the window. A therapy visit three weeks before the incident, from a prior complaint, should not appear in specials. Ask the tool explicitly to flag them, as Task 3 does.
• Units that multiply wrongly. "18 visits at $165" is $2,970. A model that reads "20 visits" from a scheduling note will say $3,300. Always match the count to the dated entries, not to a plan of treatment.

Record each discrepancy as a row in a scoring sheet: tool, item, what it said, what the record says, severity. After the test you will have a list that shows what kind of mistakes each tool makes, which is more useful than a general impression.

Scoring the narrative

Numbers first, then prose. Read each narrative with the records beside you and mark every sentence as supported, unsupported or overstated. A sentence like "she has been unable to lift her child since the incident" needs a source: a treatment note, a pain diary, a client statement. If nothing in the file says it, it is an invention, and a plausible one is more dangerous than an obvious one.

Illustrative example: an assistant writes "the client's injuries have permanently affected her ability to work". The employer letter in the ledger says she missed six days and returned full time. The sentence is unsupported and, if sent, damaging to credibility. Delete it and replace it with what the records show. Where the draft overstates, a useful follow-up instruction is "rewrite using only wording that appears in the records, and mark anything you inferred".

The chronology that sits behind the narrative has its own pitfalls, covered in whether AI can build a medical chronology for an injury case.

What a vendor demo must show you

When EvenUp or Supio runs your closed case, treat the demo as a test, not a presentation. Ask for these in writing or on screen:

1. The source citation for three figures you choose at random, opened to the exact page.
2. What the tool does when two documents disagree: flag it, pick one, or ignore it.
3. How a lawyer edits the draft, and whether edits are tracked.
4. Where the data lives, who at the vendor can see it, and the deletion process at contract end.
5. For vendors offering legal-reviewed demands, who reviews, with what qualification, and what their turnaround is.
6. The unit of pricing and every extra charge: per demand, per case, per seat, onboarding, integration with your case management system.

A vendor that cannot open the citation to the page, or that avoids the question of disagreeing documents, has told you something. The two-week trial scorecard for lawyers offers a ready structure for recording these answers alongside cost.

Working out whether it pays

With a quote in hand, the sum is simple, and every input is yours. Illustrative figures: your control demand takes a paralegal six hours at a loaded cost of $45 an hour, so $270. If the vendor route cuts drafting to two hours of checking, that is $90, a $180 saving per case. If the quote works out at $250 per demand, the tool costs more than it saves on a straightforward case. On a complex case, where the control takes 20 hours ($900) and the tool cuts it to 6 ($270), the saving is $630 against the same $250. The break-even depends on your mix of cases, so count how many of each you ran last quarter before you decide.

That arithmetic ignores quality effects, which can matter more: a demand that settles for more, or one that goes out a week earlier. Do not count either unless you have the firm's own figures; a vendor's case studies are not your data. And remember that time saved only counts if the checking step is real. If the saving comes from skipping the ledger comparison, you have traded time for risk.

A filled-in checklist for the finished letter

Before a draft from any route leaves the building, run it against a short checklist. Below is the checklist filled in for the illustrative rear-end collision claim, so you can see what a completed one looks like.

Notice that the checklist produced five changes in a letter that read well on first pass. That is typical of the exercise and the reason to keep it as a permanent step, however good the tool becomes. If you keep the completed checklist on the matter file, you also have a record that the figures were verified by a person.

Asking the tool to correct itself, and what comes back

When a comparison shows a mistake, your next prompt matters. A vague complaint ("the total looks wrong") usually gets a confident re-statement. A specific one gets a fix you can verify. Compare these two follow-ups for the double-counted pharmacy charge.

Weak: "Check your total again."

Better: "Your total is $750 higher than the sum of unique charges. List every
charge that appears more than once across the documents, with both page
numbers, then give a corrected table and total."

Illustrative output from the better prompt: "Duplicate found: pharmacy charge of $750 appears on the pharmacy statement (p. 22) and within the clinic summary invoice (p. 30). Counted once, the corrected total is $12,240." That is what you want: the duplicate identified, both pages given, and a total that matches your ledger. If a tool cannot explain the difference, do not accept a changed total, because a number that moves without a reason is as untrustworthy as the one before it.

Then ask for the opposite check, which catches missing items: "Which charges in the documents did you leave out of the table, and why?" A model that dropped the MRI because the invoice was a scan with poor text will say so if asked, and you can supply the figure by hand.

A mistake that looked fine until the ledger caught it

Here is a realistic failure to plan for, shown as an illustration. Suppose a draft states that the client attended physiotherapy "twice a week for ten weeks". It sounds right, and it matches the treatment plan in the first physiotherapy note, which proposed 20 sessions. The dated visit entries show 18 visits spread across 12 weeks, with a three-week gap in the middle when the clinic was closed. The letter has turned a plan into a fact. An adjuster who counts the dated entries would find the gap and the overstatement together, and the credibility cost would fall on the rest of the letter. The ledger comparison catches it because the ledger counts entries, not intentions. Make "plan versus actual" an explicit line in your test, for visits, for medication, and for work restrictions.

Where the human check cannot be skipped

Keep the following with a qualified person whichever route you choose: the choice of billed or paid figures; the decision to include or exclude pre-existing conditions; the valuation of general damages; anything that depends on the law in the place the claim is brought; and the final signature. An AI draft can help with structure and phrasing. Work that needs legal judgement stays with the lawyer, and the file should show who did it.

It also helps to write the check down as a step. Setting up human review for AI work describes a light version that does not slow a team down, such as a one-page sign-off listing the figures verified and the pages used.

A short rollout once the test is done

Choose one route for one case type, such as soft-tissue road-traffic claims, and run it on the next five live matters with the ledger check on every one. Count errors caught by the ledger per case. If the number falls to zero over five cases, extend to a second case type. If it does not, the problem is the tool or the prompt, and you want to know before volume grows. For other jobs worth handing over, seven jobs for personal injury lawyers to hand to AI first gives a sensible order.

Questions firms ask about AI demand letters

Do EvenUp and Supio publish prices?

No. Neither lists a price on its site: EvenUp's pricing address returns you to its home page and Supio's leads to a demo request. Both are sold through a conversation with the vendor, so budget by asking for a written quote that names the unit (per case, per seat or per demand), what is included, the contract term and the renewal price. Do not rely on a figure quoted in a forum or review.

Can I use ChatGPT for a demand letter?

You can draft with it, but client records are sensitive health and financial information. Use a business plan with training switched off by default, redact identifiers, and treat every figure as unverified until a person has matched it to the records. Check your firm's rules and the client's consent before uploading anything.

How do I check the figures an AI used?

Build a specials ledger by hand from the bills and records, with a page reference for each line, then compare the AI's totals line by line. Differences usually come from duplicate bills, billed versus paid amounts, and dates of service outside the injury period.

Who is responsible for an error in an AI-drafted demand?

The lawyer who signs it. A vendor's review service or citation feature reduces the chance of an error but does not move professional responsibility. Keep a record on the matter file showing who checked the figures, against which documents and on what date, so the check can be shown if the demand is ever questioned.

Further reads

• Personal Injury Lawyer Chatbot: Intake Questions, Limits and Cost — Look at the intake side of a personal injury practice.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Decide between a general chatbot and a legal tool.
• What to Check in an AI Tool's Privacy Policy and Terms — Read a vendor's privacy terms before uploading records.
• Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Check confidentiality duties before using any chatbot.
• Can Lawyers Use AI Note Takers in Client Meetings? — Decide whether AI may record client meetings.
• Per-Seat or Pay-As-You-Go? Legal AI Pricing for Small Firms — Weigh per-seat pricing against pay-as-you-go.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: EvenUp home page; Supio home, demand-letter and case-ledger pages; both checked 1 November 2026; Anthropic and OpenAI plan pages.