The AI Suggests, You Decide: Designing Assistive AI

In its explainer on how the generator works, Hypnothera draws a line I like a lot: the AI "doesn't diagnose or prescribe," and "the output is a script, not a treatment plan." The software writes something personal, and a person decides what to do with it.

I work on the same line in a very different field. I build DokuTrak, which helps accountants, lawyers and similar firms collect the documents they need from their clients. When a client uploads a file, an AI reads it first and may flag it: wrong year, wrong kind of document, unreadable. Then a person at the firm either accepts it or asks the client for a new one. The AI never makes that call.

"The AI assists, the human decides" is easy to put on a website. Making it true took six specific design choices, and I think each one applies anywhere an AI drafts something and a person signs off, including a session script.

In short: know who lives with the outcome; let the AI say when it isn't sure; show reasons, not just verdicts; keep the labels honest; make the final yes something only a person can give; and let people switch the AI off, with everyone affected told it's there.

1. Start with who lives with the outcome

The first question isn't what the AI can do. It's who carries the consequences when the result is wrong.

For our users, the answer is clear. An accountant who files a return with the wrong year's tax form, or a lawyer who relies on an expired ID, is personally accountable for it. The software isn't. So the software can't be the one that accepts the document. Everything else in this article follows from that.

The same question has an answer in hypnosis. The listener lives with a session. A practitioner lives with the script they hand to a client, which is why Hypnothera's own piece on safe AI sessions describes therapists as "gatekeepers" who review AI-created scripts. Whoever lives with the outcome should hold the final say, and the design should make that easy for them.

2. When the AI isn't sure, it should say so out loud

An assistant can be wrong in two ways, and they aren't equal. Ours can flag a perfectly good document, which costs the professional a quick check. Or it can call a bad document fine, and the professional accepts it in good faith. The second mistake reaches their client.

So we chose to be noisy on purpose. When we set the bar for launch, we allowed up to 30% of clean documents to be flagged for a second look, and we were strict about the other direction: our launch target required catching at least 97% of wrong documents. The instructions to the model say that if there is any meaningful doubt, it should answer "uncertain" rather than "valid."

We don't take the model's word for it either. A "looks valid" result only stands if the model reports high confidence (0.9 or more on its own scale) and quotes something it actually read on the page. Otherwise it's turned into "needs human review." That safety rule only ever moves toward caution. Nothing gets upgraded.

The cost is real. People see more flags than the AI's accuracy alone would justify, and we tell them to expect it. But a person can overrule a cautious assistant in seconds. They can't overrule a mistake they never saw.

For a tool that drafts sessions, the same trade-off would look like this: when the model isn't sure a theme or a suggestion fits the person, it says so and asks, rather than writing it in quietly. A question costs the listener or the practitioner a moment. A wrong suggestion delivered with confidence costs more.

3. Show the reason, not just the verdict

A verdict on its own leaves you with two options: trust it or don't. A reason gives you a third: check it.

Our check has seven possible outcomes, and each has a plain label: "Wrong document type," "Wrong period," "Possibly expired," "Missing required info," "Unreadable file," "Needs human review," or "Looks valid." When a document is the wrong type, the professional sees what was expected and what arrived. Behind each result is one plain sentence explaining why, and short quotes from the document that support it.

That's what makes a person faster instead of just busier. A reason like "this tax form is for 2024, and the request asked for 2025" takes a glance to confirm. A red dot with no explanation sends you back to reading the whole file.

I think the parallel for a generated session is visibility and editing. Hypnothera's explainer is honest about where the model makes "interpretive choices" (the metaphor, the imagery), and its practitioner workflow has a review-and-customize step where any section can be edited. A choice you can see is a choice you can keep or change.

4. Watch what your words promise

Labels quietly tell people who is in charge. We say "Looks valid," never "Approved." The model is told, in the first lines of its instructions, that its review "is a suggestion only — a human reviewer makes the final decision." In our own vocabulary, the AI "flags" and the professional "approves." We keep those words apart on purpose, in the product and in how we write about it.

Getting this wrong is easy and costly. If the interface says "Verified," people stop looking, and the review you designed turns into a rubber stamp. "A script, not a treatment plan" does the same job for Hypnothera: it tells the reader exactly how much weight the output can carry.

5. Make the final yes something the AI can't give

Saying "a human decides" is one thing. Making it impossible for anything else to decide is another.

In DokuTrak, accepting a document or asking for a replacement only works for a person signed in to the app. Anything else, whether an automated program or an AI assistant acting on the professional's behalf, is refused, and the attempt is written to the log. An assistant can help. It can't do the accepting.

Two smaller details matter too. An approval applies to the exact version the person looked at: if the client uploads a new file, it starts unapproved. And when a professional corrects something the AI got wrong, we keep the correction instead of letting it vanish.

In a practice, I'd expect the equivalent to be simple: a script reaches a client only after the practitioner has read it, and if the script is regenerated, it gets read again.

6. Let people switch it off, and tell whoever is affected

An assistant people can't decline isn't really an assistant. Before a firm sends its first request, an admin chooses whether the AI check is used at all or whether every document goes straight to a person. If that setting is missing or unclear, the answer is "human only." If the system can't read the setting at the moment a file arrives, the document simply waits for a person.

The people being assisted aren't the only ones affected, either. Our upload page tells the firm's clients that, if the firm turned it on, their documents may be read by AI for an assistive check, and that the firm remains responsible for the final decision.

Questions to ask any AI tool that says it "assists"

If you're choosing an AI tool for your practice, or for yourself, these are the questions I'd ask. They sit well next to Hypnothera's checklist for choosing an AI hypnotherapy app, which covers privacy and expert oversight:

1. When it's unsure, does it tell me? Look for an "I'm not sure" state, not only right and wrong.
2. Can I see why it made a choice? A reason or a quote is something you can check; a score isn't.
3. Do its labels promise more than it does? "Suggested" and "draft" are honest words. "Verified" and "approved" usually aren't, coming from software.
4. Is the final step mine, every time? Nothing important should happen because the AI decided it was ready.
5. Can I turn it off, and do the people affected know it's there? If a client hears a session an AI helped draft, is that disclosed, and did a person look at it first?

None of these makes an AI less useful. They make it the kind of help you can keep using, because you stay the one who decides.

Arthur Teboul is the founder of DokuTrak, where an AI flags the documents clients send and the professional makes the call on every file.