Grounded AI persona: keeping answers to your material

This page is about an AI that speaks for one real person from their own material; it isn't about the fictional customer profiles marketing teams call personas, or a chatbot playing a character.

What is a grounded AI persona?

A grounded AI persona is an AI that speaks for one person and takes its answers from that person's own material. On AIMayDo we call it a twin.

Grounding matters because your clients will read its answers as yours. A general chatbot will sound confident about almost anything. A twin that carries your name should sound like you on the things you actually cover, and say plainly when a question goes past them.

How does an AI persona trained on your own content work?

For each question, the twin picks the parts of your material that bear on it and answers from those, together with the rules you approved. The underlying model isn't retrained on your files. When you add material, its next answers can draw on it.

You can add:

• Documents (.docx), plain text files (.txt) and spreadsheet exports (.csv)
• Audio and video files, which are transcribed for you
• A video link, which is downloaded as audio and transcribed
• Text you paste straight in

If your material is in a PDF, paste the text in.

The persona comes from the same material. When you build or rebuild your twin, we draft a short profile of how you talk and reason, plus a set of decision rules. A decision rule is a line you'd hold on a question you haven't answered before, such as when you'd tell someone to see a professional in person. You read the draft and publish it, or approve rules one at a time. Rules you haven't approved don't shape any answer.

Text chat, your cloned voice and the live talking avatar all speak from that one set of material and rules. The voice and face come from one photo and a short voice clip. They only change how an answer sounds and looks.

Can an AI answer only from your material?

Yes, as a standing instruction: every AIMayDo twin is told to answer only from your material and your approved rules. If a question falls outside that material, it's told to say it's outside what you focus on instead of making something up. You can narrow it further with the topics you list in your dashboard. Ask about something off that list and it politely declines.

The twin replies in the visitor's language, even when your material is in another one, and keeps the same scope.

Some rules apply to all twins, on top of yours:

• It gives information, not regulated advice. For medical, legal or financial questions about someone's own situation, it points them to you or another qualified professional.
• If a message suggests someone is in crisis, it steps out of character, doesn't try to counsel them, and points them to emergency services and a helpline.
• If someone asks whether they're talking to the real you, it says plainly that it's an AI twin built from your material.

A crisis notice with emergency and helpline numbers also sits on each twin page. It shows the lines for the visitor's own country where we have them.

How do you test a grounded AI persona?

You answer your own test questions and a separate AI judge scores how closely the twin's answers match yours, which gives you a fidelity score.

1. You write test questions, the kind your clients ask, and answer each one yourself.
2. The twin answers the same questions with the same instructions, rules and material it uses live. It never sees your answers.
3. A separate AI judge compares each pair and scores it out of 100 for how closely the twin's answer matches your stance, reasoning and voice. It marks down contradictions, claims that go against your view, the wrong framing and a missed main point. It doesn't mark an answer down for adding detail that fits your view.
4. Your twin's score is the average across your questions.

It's a measured accuracy score, where accuracy means how near the twin gets to the answers you wrote. Once you've measured, the score appears on your twin's public page with a plain label saying what it measures. It doesn't tell a visitor whether an answer is right. A twin can match you closely and still repeat something you'd now correct.

In your dashboard each test question shows its own score and a short note from the judge, so you can see where the twin drifted. Measuring is manual. After you add material or change your rules, run it again. More questions give you a steadier number. The test uses only the profile and rules you've approved, so approve your draft first. You can measure before your twin goes live, so the first person to test it can be you.

How do you reduce AI hallucinations in a persona?

Fill the gaps in your material, list your topics and test the awkward questions, then read some answers yourself. A hallucination is an answer the AI makes up and states as fact, and no setup stops them entirely. These steps cut them down on your twin:

1. Cover the questions you actually get. If clients keep asking something you've never written about, record a short answer and upload it.
2. List your topics. A short list of what you cover gives the twin a clear line to decline at.
3. Read the draft rules before you publish them. This is where you catch a position you've since changed, or a rule you'd never put that way.
4. Write test questions on the edges. Easy questions flatter the score. Ask the awkward ones, where a general chatbot would give the textbook answer and you wouldn't.
5. Measure again after a big change. If the score drops after new material goes in, open the question that moved and read the judge's note.
6. Read some answers yourself. The judge doesn't mark down extra detail that fits your view, so a made-up example can still score well. Your own read catches it.

How do visitors know it's an AI?

Each twin page carries a banner the visitor can't dismiss. It says they're talking to an AI twin built from your published material, that it isn't you in person, that it can make mistakes, and that it shares general information, not professional advice. When your twin runs on your own website, the widget carries its own AI label.

What if you work in a regulated profession?

You can build a twin, and you decide for yourself whether one fits your work. The platform is for everyone: coaches, consultants, teachers, doctors, psychologists and psychotherapists among them. The twin shares what you've published as general information. It doesn't deliver therapy, diagnosis or treatment.

Some clinical words can't be set as twin topics, such as therapy, psychotherapy, psychiatry, counselling, mental health or diagnosis. You can list the subjects you write about instead, like stress, sleep or relationships. If you leave the list empty, the twin keeps to your area. Check your professional body's position before you publish a twin under your name.

If you work in a clinical field, a psychiatrist's live twin shows the AI banner and the crisis notice the way your visitors would see them. Coaches will find the same steps for coaching material in how to build an AI coach.

Can you switch it off or delete it?

Yes, with four controls in your dashboard that take your twin offline or remove it:

• Suspend takes your twin offline and its public page stops working. You can bring it back later.
• Withdraw consent and take down takes it offline and deletes your cloned face and voice at our providers. Your material is kept, but the twin can't go live until you consent again.
• Delete this twin deletes your cloned face and voice at our providers, then removes the twin with its material, rules, test questions and consent records.
• Delete my account removes all the twins you own, your profile and your login.

Withdrawing consent and both delete actions ask you to confirm first. You'll find them in the Data & privacy tab of your dashboard. A twin also can't go live in the first place until your consent is recorded and stored with it, and you've finished Stripe's payout setup, which asks for your identity details.

What should you look for in grounded AI persona tools?

Look at where the answers come from, what happens past your material, and how much control you keep. These questions sort tools quickly, so ask them of us too.

• Where do the answers come from? Your material only, or the model's general knowledge as well?
• What happens when a question goes past your material? Does it say so, or improvise?
• Can you read and approve the rules it follows before they shape any answer?
• Is there a measured score, and can you see how it's produced?
• Is the AI label visible to every visitor, on the tool's own pages and on your site?
• Can you take it down, and does that delete your cloned face and voice?

Where can a grounded AI persona still go wrong?

Mostly at the edges of your material, where a twin can be confidently wrong. No single check stops every mistake, and each one covers a different gap:

A twin can only draw on what you've written or recorded, and it may repeat a view you've moved on from if that view is still in your uploads. It doesn't know anyone's history with you in person, and it can't take responsibility for anyone the way a professional does. The model writes the sentences, so it can phrase a point in a way you wouldn't. That's why the AI label never comes off, and why your score shows only once you've measured it.

How do you build a grounded twin and test it first?

Create your twin for free, add your material, rebuild and approve your rules, then write a few test questions and measure it before anyone else talks to it. There's no subscription, and how it works for experts has the steps in order. Visitors pay by the minute once the free part you set runs out, and only while they're talking. You set the price and keep 80% of whatever you charge above the platform base. How a pay-per-minute AI twin works explains the meter, and how to monetize your AI avatar walks through the money.