Why we don’t put an AI in the prescribing seat
Whatever software you use, the prescription is yours. So the useful question is a simple one: when the list on your screen is wrong, can you see why? That is where a repertory search and an AI chatbot are very different.
We make Homoeoclinics, and we compete with the products discussed here. Keep that in mind. Every fact below has a public source, and the two studies we quote are peer-reviewed. They set a bar for us too.
Let’s start with the biggest point, because it is true. The prescription is yours. Kent used a repertory. So did Boenninghausen, Boger and Lippe. None of them let it prescribe for them. Whatever software shows you — ours included — is a shortlist that you, a registered practitioner, either use or throw away. Adding AI doesn’t change that.
So the question isn’t “does the AI choose the remedy?” The question is what happens when the list is wrong, surprising, or doesn’t match your own reading of the patient. At that moment, you want to question the list. That is where the two kinds of software part ways, and it has now been measured.
Ask it twice, get two answers
In July 2026 the journal Healthcare published a study by a team that included practitioners from the Academy of Homeopathy Education. They took 100 acute cases from a teaching clinic, gave each one to four free AI chatbots, and compared the answers with the remedy the practitioner had actually given.
On average, the chatbot’s top suggestion matched the practitioner’s remedy 20.8% of the time, and the remedy appeared anywhere in its suggestions 36.5% of the time. In 41% of cases, none of the four chatbots suggested it at all.
But the finding that matters most for software you open every day is one the researchers weren’t even looking for:
“answers to multiple queries on the same AI platform, even when performed within minutes, could result in different answers”
They tried this twelve times. Half the time, the top remedy changed — same chatbot, same case, same words, minutes apart.
Think about what that does to your second prescription. Six weeks later, you want to know what the reference said last time, not what it happens to say today. A reference that changes its mind isn’t a reference.
A tool that agrees with you can’t check you
The same study found something else. When the researchers pushed back on a chatbot’s answer, one chatbot kept changing its mind — switching to the practitioner’s own remedy 44 times out of 75.
That isn’t agreement. It is a tool built to please you, which is the opposite of what a repertory is for. A rubric is useful because an author wrote it down, and it doesn’t care what you think this morning. When your reading and the book disagree, that disagreement tells you something. A tool that gives way the moment you push it has lost the one thing you were asking it for.
This isn’t new. An earlier study of an automated remedy finder found the practitioner’s remedy anywhere in the software’s list 59% of the time, in the top three on 37 occasions, and not at all in 41% of cases. It also noted the tool “is not designed to mimic the rapport building or empathetic communication that is a critical element of chronic case taking” (PMC12345833). We are not claiming to beat these numbers. We are claiming you can check our work.
“Which book says that?”
Ask our app why a rubric is on your list, and it shows you: the rubric, the books it appears in, the grade each author gave each remedy, and the page. Ask an AI chatbot the same question and you will get an answer too — a confident one — that sometimes points to a page that doesn’t exist. Made-up references are a known habit of these tools. And they are hardest to spot exactly when you are least sure, which is exactly when you were relying on the tool.
What can go wrong, side by side
Reliability isn’t a slogan. It is a list of everything that can break, and what happens to your clinic day when it does. Our suggestions are worked out by our own software, on our own servers, from the books. No AI model, no outside company, no credits to run out. That removes whole rows of problems from the table below, simply because we never added them. (We once built an optional outside AI step. It was never switched on for real case notes, and in August 2026 we deleted it completely.)
| If this happens | Suggestions from an AI model | Homoeoclinics |
|---|---|---|
| You enter the same case twice | The top remedy may change — it did in half the repeat tests Healthcare 2026;14(7):909 | Same case, same list. It is worked out from the grades printed in the books, not guessed |
| You ask which book says it | A reference you have to check yourself, which may not exist | Book, author, grade and page, right on the row |
| The AI gets updated | Last month’s answer can’t be repeated | The books don’t change between two readings of the same note |
| The AI company has an outage | No suggestions — and it isn’t your software company’s to fix | No outside service is involved, so there is nothing to go down |
| Your AI credits run out | Many plans in this market limit AI use by tier Published pricing pages of two Indian competitors, read 31 Aug 2026 | Nothing on the suggestion path is metered. One price covers it all |
| The internet drops mid-clinic | An online AI call can’t finish | Repertories and materia medica download for offline use; notes written offline sync when the network is back |
| You ask where the case text went | To an outside AI company, under its own storage rules | Nowhere. Case notes are read by our own software on our own servers |
| Someone asks how you chose | You can describe what the AI did, but not why | The rubrics you kept, each with its book — the same chart repertorisation has always made |
If you read only one row, read the last one. A repertorial chart has always been how our profession shows its working, and the CCRH case record template asks for one by name. Software that gives you a chart you can explain is doing the traditional job faster. Software that gives you a number you can’t explain has quietly changed the job.
Why you edit the list before anything is ranked
On the case-note screen, the rubrics sit above the remedies, and you can work on them: search and add one we missed, or mark one “not this”. Then press Continue. Nothing gets ranked until you are happy with the rubrics. Choosing the totality is your job, so that comes first. But there is also good research behind it.
Researchers at Wharton found that people stop trusting a computer faster than they stop trusting a person, after seeing both make the same mistake (Journal of Experimental Psychology: General, 2015 — “algorithm aversion”). Their next study found the fix: people will keep using an imperfect tool if they are allowed to adjust its answer, even a little (Management Science, 2018 — “Overcoming Algorithm Aversion”).
Put that next to a black box. A tool you can’t adjust gets one chance with an experienced doctor. It gets a hard case wrong, as every tool eventually does, and is never opened again. A tool you can correct survives being wrong, because being wrong becomes a conversation instead of a verdict. We would rather be corrected than blindly trusted.
Why we show no confidence percentage
Some products show a confidence score next to each suggested remedy. We choose not to. And it isn’t because we couldn’t work one out.
A percentage next to a remedy looks like it means “this patient has an 87% chance of getting better”. It doesn’t — nothing has measured that. It only describes how a set of rubrics scored inside the software, dressed up to look like a clinical fact. We once tried a number on the rubric list in our app and took it out. When the rubrics were clearly right, a low number made the software look like it was doubting itself. When the number was high, it invited exactly the blind trust the warning below it is there to prevent. Either way, people argued with the number instead of reading the list.
So the list shows what it can stand behind — which books, which grades, which pages — and asks you to remove anything that doesn’t fit your patient. It is also why you won’t find an accuracy percentage for our app anywhere on this site. One number is misleading both ways at once.
Where AI really does help — and what we use
We are not saying AI has no place in a clinic. It is good at jobs that have nothing to do with choosing a remedy: typing up a consultation, translating a patient’s words, summarising a long history, tidying dictation. Those are language jobs, and AI is good with language.
We do use one small smart helper ourselves, for one narrow job: finding rubrics when the patient’s words share nothing with the book’s (“bossy” and DICTATORIAL have no letters in common). It runs on our own servers and sends nothing anywhere. It can’t make up a rubric that isn’t in the books, and it only steps in when your own words found nothing. We tell you this openly, because a page asking for honesty shouldn’t hide its own tools. Our line is simple: nothing that guesses gets to write the reference.
How to decide in ten minutes
Take your hardest recent case — one you have already prescribed for and followed up, so you know the answer. Then try it in every product you are considering, ours included:
- Enter it twice, ten minutes apart. Did the top of the list change? If so, you have learnt the most important thing about that tool.
- Pick the strangest suggestion and ask which book and page it came from. Then open the book. Every classical repertory is free online.
- Remove a rubric you disagree with. Does the software let you? Does the ranking really change?
- Ask the company, in writing, where your case text goes and how long they keep it. Ours stays on our own servers; get the other answer in writing too.
- Ask what happens the day your AI credits run out in the middle of OPD.
If a product passes all five, buy it. We would rather lose to one that passes than win against one that doesn’t. We think you will find very few pass, for a simple reason: you can’t make a guessing tool give the same answer twice just by adding a warning under it.
In one sentence
Both kinds of software give you a list, and either way the prescription is yours — but only one of them can still show you, six months later, exactly why that list looked the way it did. For a method built on individual cases and the second prescription, that isn’t a nice extra. It is the requirement.
Sources
- Doherty R, Pracjek P, Luketic CD, Straiges D, Gray AC. Comparing AI Chatbots to Live Practitioners of Homeopathy: A Comparative Retrospective Study. Healthcare (Basel) 2026;14(7):909. https://pmc.ncbi.nlm.nih.gov/articles/PMC13073836/
- The Application of Artificial Intelligence in Acute Prescribing in Homeopathy: A Comparative Retrospective Study. https://pmc.ncbi.nlm.nih.gov/articles/PMC12345833/
- Dietvorst BJ, Simmons JP, Massey C. Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err. J Exp Psychol Gen 2015;144(1):114–26. https://pubmed.ncbi.nlm.nih.gov/25401381/
- Dietvorst BJ, Simmons JP, Massey C. Overcoming Algorithm Aversion: People Will Use Imperfect Algorithms If They Can (Even Slightly) Modify Them. Management Science 2018;64(3):1155–70. https://pubsonline.informs.org/doi/10.1287/mnsc.2016.2643