Where AI fails in a life care plan
Answer in brief: AI fails a life care plan in 3 places: it misreads the record (possible items, superseded orders, frequencies, duplicates), it makes up what the record can't supply (prices, life expectancy, literature), and it leaves you unable to explain your method. Each failure has a cheap check; the table below lists 12.
Limits: Tool facts come from each vendor's own website, read on 2026-11-12; we did not test any product hands-on. The cost example is synthetic, and the rates in it are placeholders, not recommended growth or discount rates.
The 12 failure modes and the check for each
| Failure | What it looks like | The check |
|---|---|---|
| 1. A possibility becomes a costed row | "May need revision surgery" shows up as "revision surgery recommended" | Keep the provider's exact words in every row; sort by "may", "consider", "if" |
| 2. A superseded order stays in | A March fusion recommendation survives a June note that withdrew it | Ask for later notes that change each recommendation, with dates |
| 3. Frequency misread | "Every 6 months" becomes 6 visits a year | Open the page for every row that carries more than a small share of the total |
| 4. Duration filled in by default | Every row runs to the end of life expectancy because no stop was stated | Mark durations "stated" or "planner opinion"; never leave the tool's default |
| 5. Same recommendation counted twice | 1 note appears in the hospital, clinic and attorney copies and becomes 3 rows | Remove duplicate pages first; merge rows with the same provider, date and item |
| 6. Side, dose or level wrong | Left knee becomes right; 300 mg becomes 30 mg | Check side and dose on every row that has them |
| 7. Missing records unnoticed | A referral to neurology with no neurology note, and the plan has no neurology row | Ask for referrals and orders with no matching note; request the records before costing |
| 8. A price with no source | A chatbot answers "a below-knee prosthesis costs about $X" | No cost goes in without a dataset, edition and area, or a dated quote |
| 9. Wrong cost basis | Medicare fee schedule amounts in a plan meant to show usual and customary charges | State the basis at the top of the cost section and check every source matches it |
| 10. Life expectancy from a model | The tool suggests a number of years for the horizon | Horizon comes from a physician's opinion or a cited table, never a model |
| 11. Invented literature | A chatbot cites a study for a replacement interval that doesn't exist or doesn't say that | Open every cited paper; cite only what you read |
| 12. A method you can't explain | At deposition you can't say what the tool did or how you checked it | Keep a short note per case: tool, version or date, what it did, what you verified |
Misreading the record: failures 1 to 7
These are the failures record tools are supposed to prevent and sometimes cause. The common root is paraphrase. A tool that rewrites "anticipate" and "may" into the same neutral sentence takes away the words you need for the probable-versus-possible call. Pick tools that quote, and use prompts that ask for quotes.
Duplicates matter more in life care planning than in most record work. A catastrophic file often holds the same discharge summary 3 or 4 times: from the hospital, from the treating physician's chart, from a prior production. A tool that doesn't remove duplicate pages will hand you the same recommendation several times, and a busy planner can cost it twice.
Missing records are the quiet failure. A plan built on an incomplete file looks complete. Ask the tool what was referred or ordered but never documented, and decide whether to request those records before you cost anything that depends on them.
Making things up: failures 8 to 11
A general model will answer any question about cost, life expectancy or the literature in a confident sentence. None of those answers belongs in a plan as is.
- Costs need a dataset or a quote, an edition or date, and a geography. FAIR Health, a fee reference, a fee schedule, a vendor quote. The cost data list has the common ones.
- The cost basis needs to match the plan's purpose. A Medicare set-aside uses Medicare pricing; a life care plan in litigation usually doesn't. A tool that "estimates costs" without naming its basis has made that choice for you.
- Life expectancy is a medical opinion or a cited table. It sets the horizon for every row, so a wrong number scales the whole plan.
- Literature cited by a chatbot must be opened and read. A replacement interval supported by a paper that doesn't say so is worse than an interval stated as your opinion.
Explaining your method: failure 12
Under Rule 702 and its state versions, the question is whether your opinion rests on sufficient facts and a reliable method applied reliably. AI in the workflow doesn't change that test. It adds questions: which tool, what it did, and how you checked it. "It found the recommendations, I opened every cited page for the costed rows, and I added items from my interview" is a method. "The software generated the table" is not.
Keep 1 short note per case: the tool, the date, what you asked it, and what you verified. It takes 5 minutes and answers the whole line of questioning.
When not to use AI at all
- When the only option is a consumer chatbot account. Records are PHI; without a signed BAA covering the plan you use, don't upload them.
- On a short file you can read in an hour. The setup and checking time can exceed the reading time.
- For the interview, the home visit and calls with treating providers. These are where plans find needs the record never shows.
Questions about AI limits in life care planning
What is the most common AI mistake in a life care plan?
Turning a possibility into a costed row. Notes say "may" and "consider" often, and tools that paraphrase lose that word. Ask for exact quotes and sort by them.
Should I disclose that I used AI?
Be ready to describe it if asked, and follow any rule from your jurisdiction, the retaining attorney or your certifying body. A short per-case note of tool, date and checks makes that easy.
Can AI find needs the record doesn't mention?
No. Home modifications, family caregiving and transport needs often never reach the chart. They come from your interview and site visit.