AI at Work: From Resistance to Results with Dr. Gleb Tsipursky
21 min · August 31, 2026
Dr. Gleb Tsipursky on AI agents for financial advisors, resistance to AI in law, hallucinations, choosing and switching between LLMs, and learning AI by asking it to teach you.
Key Takeaways
- Dr. Gleb Tsipursky, author of The Psychology of AI Adoption at Work: From Resistance to Results, describes an AI agent built for a financial investment firm that finds local announcements of people about to receive a large sum, such as a business sale or divorce settlement, and writes and sends a personalized outreach email.
- The agent runs on a scheduled task every hour and costs nothing beyond an existing Claude subscription, he says, and the firm now spends its time on inbound leads and client service instead of outbound research. He says his expertise is teaching people to fish, not just building tools.
- A second agent analyzes a client intake form or call and recommends a portfolio, so the account manager arrives at the meeting with a prepared package, while inbound email replies are drafted but never sent without a person checking the tone.
- On law, Richard Gearhart says attitudes toward AI have shifted as it has become more reliable, but courts have sanctioned over 700 attorneys over hallucinated cases, and AI-drafted patent applications can be redundant, such as 14 applications when one or two would do.
- Gleb says general tools like Copilot can hallucinate cases, but legal tools grounded in Westlaw or LexisNexis do not, though they can be out of date, and that humans must review every filing.
- Gleb says Claude is the smartest tool, followed by ChatGPT, and recommends using only Claude, ChatGPT, Gemini, and Copilot, with the last two only if company policy requires them. To move from ChatGPT to Claude, ask Claude for a prompt that extracts your context from ChatGPT, then paste the answer into Claude.
- For high-stakes work, such as a critical legal filing, strategic plan, or major sales presentation, he suggests the most powerful model, Opus for most work and Sonnet for everyday tasks. To document a process, he suggests recording it with Claude Cowork, and says the best way to learn an AI tool is to ask it to teach you.
Main Topics
See how a scheduled AI agent finds leads, writes outreach, and preps account managers for meetings.
Hear why attorneys are cautious, how hallucinations lead to sanctions, and why humans must review AI drafts.
Learn which LLMs to use, how to transfer context from ChatGPT to Claude, and when to pay for a top model.
Featured Guests
Meet the Hosts

Elizabeth Gearhart, Ph.D. is the host of Real AI Use Cases: Business Owners Roundtable, where she explores how businesses are actually putting AI to work. She is also CMO of Gearhart Law and co-host of the nationally syndicated Passage to Profit radio show and podcast.

Richard Gearhart, Esq. is co-host of Real AI Use Cases: Business Owners Roundtable and founder of Gearhart Law, a full-service intellectual property law firm. A longtime podcaster, he hosts Intellectual Property News and is the host of the nationally syndicated Passage to Profit radio show and podcast.
The roundtable’s through-line: build agents and pick models deliberately, keep humans in the loop, and ask AI to teach you how to use it.
Episode Summary
Elizabeth and Richard Gearhart talk with Dr. Gleb Tsipursky, author of The Psychology of AI Adoption at Work: From Resistance to Results, about how businesses use AI in practice. He describes an hourly AI agent that finds people coming into money and sends personalized outreach for a financial firm, and a second agent that prepares portfolio recommendations before client meetings.
The conversation covers resistance to AI in the legal profession, sanctions over hallucinated cases, and why patent drafts from AI can be redundant. Gleb recommends Claude and ChatGPT over other tools, explains how to move your context from one LLM to another, advises when a top model is worth paying for, and says the best way to learn AI is to ask it to teach you.
Frequently Asked Questions
How can an AI agent find and contact leads?
How much does an AI agent cost to run?
Why are attorneys wary of AI?
Do legal AI tools hallucinate?
How do you move from ChatGPT to Claude?
When is it worth paying for the most powerful model?
Episode Transcript
View Full Transcript
This is Real AI Use Cases, Business Owners Roundtable, with hosts Elizabeth Gearhart, podcast consultant, marketing expert and PhD researcher using AI every day, and Richard Gearhart, entrepreneur, seasoned business owner and intellectual property attorney specializing in innovation. Here’s how real companies are using AI right now.
Welcome to Real AI Use Cases, Business Owners Roundtable. I’m Elizabeth Gearhart here with my co-host Richard Gearhart, and we’re here with Gleb Tsipursky today. Tell us how you’re using AI. How are businesses using AI in your experience?
So, and this is all based on my book, The Psychology of AI Adoption at Work: From Resistance to Results, peer reviewed, out with Georgetown University Press. Businesses are using AI in a variety of ways. I was just giving a training this morning for a company that is using AI for a financial investment firm, so financial advisor, wealth advisor. One of the tools that we worked on building for the company is a tool that looks in the local area where the company provides client service for any announcements that somebody is about to come into quite a bit of money who might not have that much experience managing it. So a business owner, let’s say they’re selling their business and $7 million is coming in, or there’s a significant divorce settlement, and one of the spouses doesn’t know how to manage the money. The other spouse managed the money all this time, so a several-million-dollar settlement. Through the tool, the AI tool finds this information, then customizes a pitch to this person. It finds their email. And then it creates a pitch specific to that individual saying, hey, we specialize in working with people who are recently divorced, and we know it’s a challenging situation. What we want to do is just help you take one of the number of loads that you’re dealing with right now, by helping you figure out how to invest this money safely and securely for your future. And then it sends the email, so it automatically sends the email.
Wait, do you review the email before it goes out?
Not at this point. They worked with it until they were happy with it. So right now, instead of having to do outbound research and trying to reach out to people, they’re just waiting for inbound emails. It’s saving their time, spending their time much more efficiently on providing client service and ensuring client retention. And they’re getting inbound emails that are a result of their outreach.
Do you have any idea what their return on investment is? How many leads they get compared to how much money they spent?
They’re spending very little. They just have a subscription to a Claude account, and this Claude account is $20 per month per person for up to five. That’s what’s doing it. So as long as you already have a Claude account or a ChatGPT account, you use that, and that’s not costing them any additional money. You just run a scheduled task that runs every hour, researches various announcements, reaches out, sends an email, and doesn’t cost any extra money.
Are they using agents to build a repeatable algorithm that consistently finds the information in the right way, identifies the emails in the right way, creates the content in the right way?
So it’s an agent. Agent is a general concept for an AI task that does things in the real world in a repeatable way. So they schedule it to run once an hour and then they are automatically emailing. And then they’re getting inbound leads, and they can spend their time on dealing with inbound leads while also, of course, serving their existing clients.
So did you write the agent for them or did they do it themselves?
I taught them how to do it. My expertise is not just developing tools. My expertise is in teaching people to fish.
So how do you actually build such agents?
Another agent that they developed, we want to talk about this one, is a meeting prep agent. The client puts their information into an intake form, or it’s the result of a phone call with a client, either way, whichever the client prefers. Then it analyzes that information and it recommends a portfolio for the client based on that information. The account manager goes over the portfolio, and then they present to the client to talk about what to do and next steps.
So it really is automating a lot of the process, but you still need people in the mix, right? You still need people at the end of the task, maybe.
Well, yes. For inbound emails there is an email draft tool that automatically drafts a response, but it doesn’t send the response. A person needs to evaluate that and decide, is this the right tone for this specific context? And tweak the email and then send it with whatever the next steps happen to be. Now, the AI tool can automatically follow up, and there are automatic follow-ups based on that. You still need the person who manages that account. Then you need to figure out if they want an intake call, and of course you don’t want to have an AI agent manage that. That’s a high-value client, and the account manager has to meet with the client. But they now have that pack prepared for them with recommendations for what to do next, which is much easier than they had otherwise.
Great. You wrote a book, The Psychology of AI Adoption at Work: From Resistance to Results. So Richard tried to bring AI in years ago. How’d that go?
Initially it met with some resistance. I would say, though, a couple of years ago AI was a lot different than it is now. It’s improved as a tool, and I think attitudes have shifted in part because it’s probably more reliable and the output is generally better. There’s still a lot of suspicion about AI in the legal profession. Courts have sanctioned over 700 attorneys in pleadings where the AI hallucinated cases and the court checked up on it and said, they didn’t find anything like this. As of probably three or four months ago, it was still hallucinating cases, but I do find it a good sounding board for testing legal theories, trying to understand the psychology of different people involved in a case. And I think the results there are usually worth considering. But I’m not at the point where I would adopt it on a wholesale basis. I think a part of the skepticism in the past is not technophobia, it’s wanting to do a good job for the client. As it gets more reliable, I think more attorneys are going to be using it. There also may be some fears that because it can do things so quickly, it’s going to undermine what we do and the service we provide.
For patents?
For patents, it’s usually pretty bad. I don’t think it’s gotten better in some ways. In some ways it seems like it’s gotten worse. I think drafting that kind of legal document is something that AI is well adapted for. What sometimes happens is we’ll have clients who will put their inventions into AI and it’ll literally draft 14 patent applications for them, which is overkill for most and expensive. If each application costs $20,000 from start to finish, and you’re an entrepreneurial company starting out, it’s ridiculous. Traditionally the invention can be protected in one, maybe two applications. AI still has a lot of redundancy. And the clients come to us and they think that because it’s AI, because it looks nice and uses some legal language, that what they have is correct. They get upset if you suggest, well, maybe we don’t need 14 applications, maybe we just need one. They think AI is smarter than we are on that type of topic.
But you still use AI every day?
I still use it every day, I use it for all sorts of stuff. But I’m curious about whether you would have an approach to handling that kind of situation.
Sure. I work with a number of law firms, as you can imagine. Professional service firms are my expertise, so I’ll talk about law firms that are general practice as well as ones that include patent. You’ll have hallucinations when you use tools like Copilot or something like that. If you use Westlaw or LexisNexis or specialized AI tools that are grounded in legal cases, you’ll not have hallucinations. You’ll not have cases that are made up. The worst that happens is that they are not as up to date as they should be. This has to do with Westlaw or LexisNexis not having the latest court case inside and not citing the latest court case. I’ve heard complaints about that. That is not hallucination. That is just, oh, this case is slightly outdated, so you should have cited something else to make a more convincing argument. Second, in terms of legal AI tools and what you can do, if you use sophisticated tools, you can have very strong pleadings, very strong drafts, very strong outcomes, but you need to have humans review them. Absolutely no question about it. I’ll tell you about an example I had myself with a personal situation where I’m involved in a legal dispute. I was talking to my lawyer and he sent me his draft of what we wanted to file. I looked at it, I ran it through the AI tool and said, well, this is quite good, but here are some weaknesses in the lawyer’s language, and here are ways it could be strengthened in accordance with Ohio law, since it was filed in Columbus, Ohio, to make the argument more robust against counterarguments. I worked with an AI tool to have some ideas for strengthening this pleading. I told him, if you think it’s helpful, please integrate it. If you think it’s not helpful, you’re the expert, you make the judgment. And he said, oh yeah, this would actually make the pleading stronger, I’ll integrate this language and then it’s ready to file.
I wanted to ask you about this. I read about a research paper when I was scrolling, and it was very recent. And I started asking questions about this topic to Perplexity, which supposedly uses RAG, retrieval-augmented generation, so that it supposedly goes out and finds it. It totally did not find the study.
No, Perplexity is outdated.
So which tool do you think does find the most recent? You think Claude is the best one?
It’s no question. It’s the smartest tool, followed by ChatGPT. It’ll still do a solid job. Perplexity is behind Copilot, so it’s going to be worse than Copilot. You should not be using Perplexity.
Have you used it lately? Because I hadn’t used it for a long time and it seemed like it was a lot better all of a sudden.
I’m just talking about where the major labs are. So you have Claude from Anthropic, you have OpenAI, you have Google Gemini, you have Microsoft Copilot. Those are the four that anyone should be using. And the latter two you should only use if your company policy requires it, because they’re less smart than Claude from Anthropic or ChatGPT from OpenAI.
So I started using ChatGPT and I created a lot of different project folders. And it seems like ChatGPT knows me pretty well now. It brings in things from earlier discussions and chats into what I’m doing and it does it in a relevant way. I would like to maybe do more work in Claude. I think about how I rebuild the experience that I have with ChatGPT and transfer that to Claude, because I’ve been using ChatGPT for probably a year, doing seven or eight queries a day, something like that. And I’ve put all sorts of different kinds of information in there. So now to duplicate that on Claude seems like it would be a big time investment. Is there a way to transfer information between the LLMs?
Absolutely. Ask Claude, tell me the prompt that I would have to give to ChatGPT in order to get the context from all of my conversations and then transfer it to Claude. Claude will create a prompt for you, which you then copy and paste into ChatGPT, and ChatGPT will respond with the context from your previous conversations, and then you would input that into Claude and Claude would have that context. Now it wouldn’t be perfect, it wouldn’t have all your projects, so it wouldn’t have the structure, but it would have the context in the memory.
Interesting.
But that would save at least some time.
Do you think there’s any advantage to rerunning the same prompts in different LLMs just for the purpose of having a backup, because the technology is evolving differently at different organizations?
It’s a lot of extra time that you don’t need. What I told you will work for every LLM. If for some reason you want to transfer from Claude to ChatGPT, you would do the exact same thing in reverse.
I just got hit with something really bad. I had my speaker website and my blog on Manus. Are you familiar with Manus?
Yes. You know what happened, right?
Yeah. And then I was like, oh no, no, Meta, you don’t get to own Manus, Manus has to be a standalone company. So I went in there today to post a blog and I got this message like, we’re going to wipe out all your data, but you can back it up here. So I did back it up and put it on my Google Drive, but I don’t know if I did it right. I asked Claude if I did it right, and it said yeah, it looks like you did it right. So Claude is smarter. But I know from my blog posts, I saved them all in Word because I wrote them myself and then I had ChatGPT optimize them for LLM search. So I do have a separate copy of them on my hard drive, on my computer, in Word. But you know that happens, right? It’s scary.
The last valuation of Claude was over a trillion, the last valuation of OpenAI was over 800 billion, and there’s pretty much no company that can afford to buy either. You don’t need to worry about any of those. The same thing goes for Copilot and Gemini, for someone who’s using those, because obviously those are two of the largest companies in the world, Microsoft and Google.
I have another question for you too.
Sure.
I did buy some credits for Claude Fable 5. And I was using it, but I don’t think I need something that powerful for the queries that I’m doing. What are people normally using Fable 5 for? Is that for programming more than anything else?
If you want to review all the information of a legal case and make a really strong legal brief for the court, that’s going to be a critical filing. That’s something where you want to use Fable, something that’s really important to get right and pretty complex.
Could I use it to, or should I use it to try to do an agent? I talked to a friend of mine. She said, well, you can, Fable 5 will just make an agent for you.
You can use Sonnet to make an agent. You can use Opus to make an agent. You don’t need Fable 5 to create an agent for you. Fable can be used to create code, like you said, but it can also be used to create complex legal products, complex legal outcomes. Let’s say you’re doing your strategic planning and you want to pull together all the information for your strategic plan from everything that you have, and then draft the initial strategic plan based on your goals and guidelines. Fable can do that for you, something pretty sophisticated that you want to be really high level. Or let’s say you’re doing a presentation to make a really important sale to a client, high stakes. Fable 5 is good for that. Those are the use cases, something that’s really sophisticated, something that’s really high stakes, where it’s worth paying for. Otherwise, Opus is fine, and for everyday tasks, Sonnet is fine.
Okay. Yeah, it keeps saying, you’re using Fable 5, do you really need to? It’s giving me this one because I did pay for credits.
You don’t need to be using it for everyday tasks. It needs to be worth the effort and the pain.
I guess I have one more question. We have processes that people in the law firm do, and I want to have them screen record the processes and then have an AI delineate the process and really define it, so that maybe we could turn it into an agent at some point. What would you use for that? Claude?
Claude Cowork has that functionality. Download the Claude desktop app, turn on the Cowork functionality, and then ask it, what are the steps needed to record this using Claude Cowork? It’ll explain it to you very clearly. Once you’ve recorded it with Claude Cowork, it’ll break it down into the series of steps that it sees, and then you can create an agent from that.
That’s right. The best way of figuring out how to do something with an AI tool that’s high quality is ask it how to do it.
Yeah. Have it teach you. I find that’s how I’m using them more than ever, is having them teach me how to use them.
Wonderful. Exactly. Is there anything else we need to say before we go? Your book is The Psychology of AI Adoption at Work: From Resistance to Results, by Gleb Tsipursky, T-S-I-P-U-R-S-K-Y. And where can we find your book?
First of all, people can connect with me on LinkedIn. Tell me you heard me on the Real AI Use Cases Business Owners Roundtable podcast, because I won’t accept a random LinkedIn connection. There’s too much spam these days from AI bots. And then you can find my book on Amazon, Barnes and Noble, any bookstores. It’s out with Georgetown University Press, a traditionally published, peer-reviewed book, all that good stuff. If you want a free copy of the introduction and a chapter of the book, you can go to my website, DisasterAvoidanceExperts.com/AIBook. That’s for a free book chapter, the introduction to the book. If you’ve already bought the book, whether pre-ordered or ordered, just put the receipt number in there and you will get a free assessment on how to adopt AI effectively and a manual on the seven critical mistakes that leaders make in AI adoption.
That is excellent and so very timely, and it’s awesome that people like you are helping people figure this out, because I think it can help businesses a lot, but you have to know how to use it, and there’s so much, like you said, spam out there. Thank you.
Well, thank you.
This has been Real AI Use Cases, Business Owners Roundtable.
You have been listening to Real AI Use Cases, Business Owners Roundtable. We hope you found this valuable. Join us again for more stories, because the future of business is driven by AI. This podcast was recorded at the iHeart Studios in Manhattan as part of the Passage to Profit Radio Show.
Keep exploring the show
Hear how more business owners are putting AI to work in their day-to-day.
