Is an AI Twin Any Good? My Experience Using and Building Them

Yes, an AI twin can be genuinely useful, but it depends on how well it represents the person behind it. I want more than familiar phrases and a convincing tone. I want advice grounded in what that expert actually teaches, applied to my situation, with an honest answer when something is outside its knowledge. If the concept is new to you, start with our guide to what an AI twin is.

I’ve used Ask Iman, Delphi-based twins, and custom GPTs. I’m also building Padro, so I have a commercial interest in this space. But my opinion comes from experiencing both sides: conversations that made the idea click for me, and others that showed how quickly a twin can lose your trust.

For coaches, consultants, and creators, I think there’s a real opportunity here to serve an audience better and earn from expertise in a new way. The question is whether the twin is good enough for someone to keep using after the novelty wears off.

The interaction that convinced me an AI twin could be useful

I asked Ask Iman to help me validate business ideas. It laid things out in detail, using terms and a tone I recognized from Iman Gadzhi’s content. It felt much closer to receiving advice through his perspective than getting a generic answer about starting a business.

That familiarity mattered. I wasn’t looking for every possible opinion on the internet. I wanted help thinking through my ideas using the approach of someone whose content I already knew.

I can’t verify that Iman himself would have given the same response, and I wouldn’t confuse that conversation with actual market validation. But it helped me think through the ideas, which was what I needed from it.

For an expert’s audience, that is a meaningful use case. Someone may have watched your lessons and understood the principles, yet still struggle to apply them to their own situation. A good twin should help close that gap when you’re unavailable.

The interaction that showed me the problem

The same product disappointed me when I asked about something Iman had posted recently. It seemed not to know the material, shifted onto something else, and gave me what I judged to be a hallucinated answer.

I don’t know whether the cause was missing content, a failure to find it, or something else behind the scenes. What I do know is that the answer didn’t acknowledge the gap.

That’s a serious problem for anything presented as someone’s twin. When the system sounds like a person you trust, it’s easy to give its answer more weight than it deserves.

My standard is that it should never invent a belief, experience, or recommendation and attribute it to the creator. It can explain an existing framework or apply it to a new situation, but it should make that distinction clear. Applying someone’s principles isn’t the same as knowing what they would personally say.

When the relevant knowledge isn’t there, I would much rather hear that than receive a confident answer to a different question.

Being correct is only part of being a good twin

A response can be factually reasonable and still fail to represent the expert.

Imagine a coach who always diagnoses the client’s situation before recommending a strategy. If their twin immediately produces a ten-step plan, it might sound helpful while skipping the most important part of their method.

This is why I care about tone, language, teaching style, and the reasoning behind the advice. Does it ask the questions the expert would ask? Does it recognize the exceptions they care about? Does it explain why a recommendation fits, using their actual material?

It also shouldn’t agree with everything the user says. In research published by Anthropic in 2023, the assistants tested sometimes favored agreement with a user’s views over truthful responses. That study isn’t a verdict on today’s AI twins, but it identifies a behavior worth testing, especially in a tool used to validate ideas.

If an expert is known for challenging weak assumptions, a twin that enthusiastically endorses every plan isn’t representing them well.

Some of the most valuable knowledge was never written down

Building Padro has made me think more carefully about what actually goes into a knowledge base. Course videos and published articles are an obvious starting point, but so are handwritten notes, journals, and explanations that have never become polished content.

Even those materials won’t contain everything an expert knows.

A consultant might recognize a warning sign almost immediately, yet never have explained which details tipped them off. A coach might make an exception to their usual framework because of something learned through years of experience. That unwritten judgment is often called tacit knowledge.

At Padro, we’re using an interview mode to ask questions based on the creator’s knowledge and experiences, helping bring more of that thinking into the material their twin can use. The useful questions go beyond asking someone to describe their approach:

• When would you advise someone not to follow your usual framework?
• What would you need to know before making this recommendation?
• What happened with a past client that changed how you teach this?

The aim is to capture concrete explanations and examples the creator can review, rather than ask the AI to guess what’s in their head.

This isn’t an idea unique to Padro. Delphi also offers an interview mode designed to capture knowledge and ways of thinking through questions. For me, the meaningful comparison is how much useful judgment an interview uncovers and how faithfully the twin applies it afterward.

There is relevant research, too. A study involving 1,052 people found that agents given interviews or survey responses predicted participants’ answers better than agents given demographic information alone. It studied simulations of attitudes and behavior, not commercial coaching twins, so it doesn’t prove that an interview can recreate an expert. It does give us a reason to take richer personal input seriously.

Is it better than a custom GPT?

The custom GPTs I’ve used were mostly disappointing until I supplied a lot of information, which took effort and time. They often felt narrower than the experience I wanted from a twin.

That’s my experience, not proof that custom GPTs are inherently poor or incapable of representing a broader body of work. A dedicated AI-twin product still has to earn the comparison.

What I want is a coherent view of the expert’s approved knowledge: their courses, explanations, examples, opinions, and exceptions. Then I want a manageable way to keep that knowledge current, review weak answers, and make the experience available to an audience.

Simply calling the product a twin doesn’t establish any of that. My concern with this category is how easily a generic chatbot can acquire a person’s name and photograph without becoming meaningfully better at representing them.

I would compare the answers to the same real questions before deciding which approach is better.

Can an AI twin help a coach earn more?

This is the business opportunity I’m most interested in. Saving time on repetitive questions is useful, but a twin can also become another way to sell access to your expertise.

You could offer it as a standalone subscription, include it in a premium coaching package, or make it part of a course or membership so people can get guidance between sessions. We’ve written more about the ways you can monetize your expertise with an AI twin.

There are public examples of these models. In a Delphi-published case study about George B. Thomas, the company describes a $25-per-month HubSpot Helper subscription and reports more than $20,000 earned from new revenue streams. The article doesn’t give a clear period for that revenue figure, and it’s a vendor case study, not independently audited evidence or a typical earnings forecast.

That’s encouraging, but it doesn’t mean an audience will pay just because a product uses AI. I think the strongest starting point is an expert people already seek out for advice, with useful material and recurring questions the twin can help answer.

I wouldn’t put a strict audience-size requirement on experimenting. I would separate trying it from assuming it will sell.

The numbers that matter after launch are paying users, repeat usage, renewals, and what remains after platform costs, support, and upkeep. If it’s bundled into a package, you also need to understand whether it genuinely improves that offer rather than simply adding another feature.

What the research does and doesn’t tell us

Research on AI coaching is promising, but we should be careful about what we claim from it.

A 2026 research preprint involving 517 participants found that an AI career coach improved reported goal progress compared with receiving no support over a two-week follow-up. However, it did not significantly outperform a structured written-reflection exercise on overall goal progress.

That is evidence for a particular form of short-term support, not proof that a digital version of any coach is as effective as working with them personally.

The same distinction applies to popularity. Delphi reports more than 2.5 million questions answered by Matthew Hussey’s twin. That shows substantial usage according to the provider, but a question count doesn’t tell us whether each answer was accurate or whether users got better outcomes.

I find these examples worth paying attention to. I wouldn’t use them to promise a perfect replica, guaranteed earnings, or a replacement for a human relationship.

How I would test an AI twin before sharing it

I would start with questions real clients ask, then deliberately include some awkward ones. A polished demo usually shows what a system knows. I also want to see what happens when it doesn’t.

Sources deserve particular attention. Research on citation evaluation distinguishes between sources that fully support a statement, partially support it, and don’t support it. A reference appearing below an answer isn’t enough on its own.

I would also test the experience with a small group of intended users. Can they get useful answers without elaborate prompting? Is the wait reasonable? Do they come back? What questions expose missing knowledge?

Those gaps should feed into ongoing improvement. One of the things we’re focusing on at Padro is helping creators identify where users run into problems so they can supply missing information and review the result.

That work doesn’t end at launch. Neither does the balance between response speed and cost, which becomes especially important when people use a twin for very different kinds of conversations.

Finally, users should know they’re speaking to AI. Creators should be deliberate about what private notes or client material they share, and define when a conversation needs their personal attention.

So, is an AI twin worth it?

My answer is yes, when it gives people useful access to an expert’s knowledge and approach while that person is unavailable. I’ve had that experience myself.

But sounding like someone is not enough. A twin needs to represent their teachings faithfully, explain its recommendations, and recognize the limits of what it knows. Otherwise, the familiar tone can make a bad answer more convincing.

For coaches and creators, I see it as a potential new revenue stream built around expertise people already value. Whether it deserves to become a paid product comes down to the quality of the experience and whether the audience finds it worth returning to.

That’s the standard I’d apply to any AI twin, including the ones we’re building at Padro.