How AI Gets Reliable Financial Data: Inside the AI Supply Chain
19 min · October 1, 2026
A financial data provider explains how it feeds answer engines like Perplexity, why quality checks come first and how AI built a deck in 10 minutes.
Key Takeaways
- Financial data providers do not just use AI, they also supply the data that powers AI answer engines and brokerage platforms.
- New startups are building their systems around AI from the start, which is easier than integrating AI into a legacy platform.
- The most important use of AI is whatever process is most tedious or taxing for your business.
- Quality assurance is half the job: automated outlier checks, human review and AI hallucination checks before publishing.
- Enterprise agreements with AI companies keep company data from being uploaded to open consumer tools.
- AI answer engines pull from specific data vendors for financial questions because accuracy has a right answer and the rest are wrong.
- Anyone selling data to LLMs needs consistent, structured, daily output from a reputable source.
Main Topics
Why accuracy and quality checks matter when AI answers financial questions.
How data vendors license their feeds to AI companies and brokerage platforms.
How enterprise agreements and role-based use shape AI adoption across a company.
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.
Elizabeth and Richard talk through automating their podcast workflow, comparing Perplexity and Claude, and why accuracy matters.
Episode Summary
Andrew Lebbos, who runs the licensing department at Benzinga, explains how his company both uses AI every day and supplies financial data to AI companies and brokerage platforms. When someone asks Perplexity about an earnings result, the answer can come from data Benzinga produced in-house.
The conversation covers quality assurance, including OCR for handwritten government trade filings and human review before AI-generated content is published, enterprise AI agreements and how different roles use AI. Andrew also shares how a tool called Genspark built a partnership deck in 10 minutes, and the hosts compare Perplexity, Claude and other tools.
Frequently Asked Questions
How can AI companies get reliable financial data?
How do you check the accuracy of data used by AI?
What is the most important use of AI in a business?
How can a company set guardrails for employee use of AI?
What AI tools does Andrew use?
What does it take to sell data to AI companies?
Episode Transcript
View Full Transcript
Welcome to Real AI Use Cases: Business Owners Roundtable. We have Andrew Lebbos here with us. He is in the financial industry, and he is doing something really interesting with AI.
Absolutely. We use AI every single day to develop all of our processes for communication. We are also a provider to some of the AI companies. We provide our data to Perplexity to help with their answer engine. If you go on Perplexity and ask about the earnings results for SpaceX yesterday, it will return some Benzinga data that we produced in-house. So not only do we use AI, we also contribute to AI. A lot of our partners are building AI systems in their brokerages and platforms, and we are powering those systems with our data.
What is your role in the company? Are you the founder, or do you have a team helping you with this?
I have been at Benzinga for seven years. I am not a founder, but I am one of the longest-tenured people there. I run the licensing department. Most people know Benzinga as a financial media company, because they see our news on their trading platforms, but we also have a data division that we license to brokerages, and that is the department I run.
I listen to the AI Daily Brief podcast when I can. Do you ever listen to that one?
No, but I think I am subscribed to their newsletter.
One thing the host said a few weeks ago was that when people build their systems now for their companies, instead of adding AI to what they already have, they are building them around AI. Are you finding that with your clients?
With the newer startups we are onboarding, yes, for sure. There is definitely a benefit to building today, because integrating into a legacy platform is a lot harder than building from scratch.
What do you think is the most important use of AI in a business today?
I think it depends on the business. Whatever process is most tedious or taxing for the business, if it can be automated with AI, that is the most important use of it. What do you think is the most important use of AI in a business?
I have to agree with you. Right now I am trying to automate a process for this podcast itself. We have an assistant who generates the YouTube video, and she also gives me the audio file and the transcript. Then I put those on our podcast distribution platform, generate the cover and put everything on a website. I am trying to figure out how to automate that whole process. Perplexity has been pretty amazing lately, and a friend said the newest Anthropic model is really good too. I have to figure out which one, and maybe I will dabble with both and see which gives me better results.
That is ultra complex, so it is interesting in terms of the accuracy of the data you provide. It is obviously important that if you are providing financial data to an LLM, the data is correct, and one of the challenges with AI is that it can sometimes make mistakes. What steps do you take to verify the accuracy of the data?
A big investment of time and resources on our end is QA. QA is half of our job. Aggregating and sourcing the data from primary sources is half, and bringing it in, organizing it and ensuring it is accurate is the other half. Depending on the data set, we have automated systems that run to make sure there are no outliers. We have some human QA, with humans actually looking at the output before we send it. We have some products built on AI, and before we publish the Bull Says Bear Says specifically, we have a human review it to make sure there is no hallucination. QA is a huge part of our business, and one of our differentiators with our clients is that we take it extremely seriously.
What are some of the steps involved?
One example is government trades. When someone in the Senate or the House makes a trade, they have to file it, and the way they file it is not standardized. Some of these folks are handwriting and spilling their coffee on the sheets. We built an OCR system in-house and then built a QA layer on top of it. Years ago, before the technology was there, we would go one by one through the filings and try to do it with our eyes. Human error is dangerous in financial data, and our data is live on brokerage platforms. You see our news on Fidelity and on JP Morgan Chase. Those platforms have no space for error, and if we miss, we hear about it right away. So we have doubled down on our QA, and with the advent of AI it has helped us win a lot of those partnerships.
How many large language models do you personally use? Do you have a favorite?
I primarily use Claude, including Claude’s computer capabilities. It is extremely helpful the majority of the time, and I also still use the web app, old school. I use Perplexity, including Perplexity Computer. I use Genspark for decks. I am not sure if you have tried it. It is similar to Perplexity, but built specifically for making decks. Sometimes I just turn on my recording, speak out loud what I want the deck to look like, attach a couple of things, and in 10 minutes there is a deck. I think that is the coolest use case. The first time I used it, I did a deck for a pretty big partnership, and there were a couple of points of contention where I could not figure out which direction to go. Not only did it build a deck that would have taken me 40 hours, it built it beautifully and it answered the questions that were lacking. It filled in the gaps I would otherwise have had to figure out. It is pretty incredible.
In your company, what are the policies around individual employees using AI for their work? Are there guardrails or guidelines management suggests, or is it left to the discretion of the individual employee?
That is a great question. I am not quoting our handbook, but in short, we all work out of company instances, so we have enterprise agreements with all these AI companies. We are not just going to an open consumer ChatGPT and uploading all of our personal or company data.
Nobody should do that. Use an enterprise version, especially when it comes to your business data and information.
Exactly. We have an enterprise relationship with a lot of these different LLMs and other models. Day to day, it is up to the team member to decide. I might ask the enterprise version of Claude about something, see what it says, and if I then present that information or make decisions based on it, I am held accountable for how it is used and implemented.
Is there a wide spectrum of how team members use it? Do you ever talk about how you use it?
The team uses it extensively. Our engineers are best at it, because they are tech-forward. We have amplified the number of tickets we can put out, and the time to MVP is exponentially faster than before AI. We are pumping out products, fixes are much faster and tickets are getting knocked out. Our analytics engineering and data engineering teams do an incredible job with AI. It is a wide range, and there are controls. We have a 50-person news desk that writes everything by hand, and we will continue to, because the art of editorial is the connections these writers have. But we use AI to help find sources for the news desk. The engineering team may be producing actual code with AI, while the news desk uses it as a resource to point it in the right direction. Every role at the company uses it in a different way.
I had another question. You said you were letting Perplexity train on your data, is that right?
No. Perplexity is a client of ours. Imagine you went to Perplexity and asked what the results of Tesla’s earnings were this quarter. Perplexity pulls from specific data vendors instead of just scraping the web, to avoid hallucination and give a more targeted response, and we are one of those data vendors. Earnings is just one example. They take a few of our calendars and our news feed, and they are really good about attribution. They are very nice to work with, a very upfront, honest AI company. They use us as a reputable source to answer questions in their engine.
And they are paying you for that?
Yes. That is our whole business. We license our data to a partner. If you see our news or information on your 401k provider’s platform, there is probably Benzinga data on it, and that platform pays us a monthly fee. Instead of just displaying it on their platform, these AI companies are using it to answer questions, but it is the same premise. We produce the data, you use it, and you pay us for the production and delivery of the data.
But not every bit of data is paid for by Perplexity. Presumably your data is behind some sort of paywall they could not access without your license, but there is also other data available free of charge from any number of sources. What would be the business model for somebody who has a valuable data set and wants to license it to an LLM? What steps would they take?
Every year we host an event, the FinTech Awards and Deal Day, and every year people come out of the woodwork with interesting data they are looking to sell. Ten years ago they wanted to sell to the hedge funds, five years ago to the brokerages, and now they want to sell to the LLMs. The biggest barrier to entry for anyone with data to sell is the consistency of the data and the delivery. They need something with daily output, something structured. That is how the data has to look for LLMs, or anyone, to buy it. That is the first and hardest part. Another barrier I have seen is that some people have data but are not a legitimate, reputable company. The foundation of the business and the data set needs to be at a certain level for people to even consider purchasing it, and I assume that is true for any service.
So you have to have the authority and an actual research system set up if you want to sell data. I think Benzinga gets to sell theirs because you have the authority. You have been around a long time, you have proven yourselves and your data has proven out because people are using it. I do have a question about accuracy. I was talking to Gemini one night, asking how we can show up and be the answer, and it said Google invested in Reddit and it is training on Reddit. There is no verification of the answers there. Anybody can say anything. Have you looked into that?
I do not know that I have the credentials to answer on the specific application of Reddit within these platforms. I know it is used in a question-and-answer fashion, and that Reddit is very human, with a lot of human error and bias. But I also know it is used to understand how humans talk from a training perspective, which I think is pretty cool. From an accuracy perspective, I am not really sure. It is something I should look into. Most of the LLMs that use our data are under NDA, so I cannot say which ones. But the reason they approach us is that financial information is cut and dry. There is a right answer and the rest are wrong. There is not going to be ambiguity around analyst ratings for a stock. The analysts gave these numbers, you add them together and divide by the count, and that is a consensus price target. In providing results for financial questions, you need really sound information. It cannot be ambiguous, deduced from 10 Reddit threads. If you are asking for a pizza recipe, you could pull the thousand-reply pizza thread and see what is most often said. For finance, I think they realize how important accuracy is, and input equals output.
Do you think they are trying to get more accurate in general? I know for financial they have been.
I think that is the goal, of course, and I think they are getting there. If I look back every quarter, the models are so much better. It is scary how fast they are getting so good.
Have you used the newest Anthropic model?
Yes, there was free access for part of June. I was blown away. I used it to organize things and create lists, and for a competitive comparison. You just give it a prompt, walk away, come back and have what is literally a Word document you could present to a client.
Do you think it is better than Perplexity? Have you tried Perplexity for the same task?
I cannot give a one-or-the-other answer, but I know it burns a ton of credits.
I asked each of them the same question. I had to buy credits for that one, where my Perplexity subscription worked, and I actually thought I got the better answer from Perplexity, to be honest. But people swear by the other, so I am going to keep experimenting.
I think it is good to test all of them and keep pushing. The more we use it, the better it will get, especially for your use case, your instance.
I also use Gemini because we do a lot with video, and I use ChatGPT because I like the images it generates, and I am using Manus. It takes a lot of time to compare all of them with the same prompt, but in some cases it is worth it for the double and triple checking. This has been a really interesting conversation. It is comforting to know there is actual real data being used to train these models. It is not all just the wild west.
For sure. Our goal at Benzinga is to democratize finance and make it easy for everyone to invest. We think AI is a great tool that enables personalized experiences for any investor, and we are happy to be providing strong data to power these models.
If somebody wanted to follow up on this, could they?
I would love it if you visited Benzinga.com/apis, or you can add me on LinkedIn. Andrew Lebbos on LinkedIn. I am always on there.
Thank you, Andrew. It has been a pleasure.
Keep exploring the show
Hear how more business owners are putting AI to work in their day-to-day.
