Ron Green, co-founder and CTO of KUNGFU.AI, has been building AI systems for almost 30 years. He explains why the biggest enterprise wins come from your own data, not from generative AI alone.
A financial company had hundreds of people reviewing loan decisions, 9 to 5, with a 24 to 48 hour turnaround and a fraud problem. KUNGFU.AI took that company’s own historical data and built a system that now automates about 60% of its $2.8 billion in annual loan factoring decisions. The turnaround went from two days to nine seconds. Fraud and chargebacks went down. And when the company announced it, its market cap jumped by a quarter billion dollars that week.
I wasn’t on this episode of Deconstructing Data (Adam Fitzgerald, our enterprise sales director, co-hosted with Jessie), but after going through the conversation with Ron Green, that story is the one I keep coming back to. It wasn’t a chatbot. It was proprietary data put to work.
Why isn’t generative AI enough to create enterprise ROI?
Ron Green has been in AI since the 1990s, during what’s called the second AI winter, when almost nobody was investing in the field. He did his master’s in AI at the University of Sussex and specialized in deep learning and neural networks, the exact techniques powering everything happening now.
His concern today is that people have started treating ChatGPT and language models as if they’re the only kind of AI there is. Generative AI is incredibly powerful, but Ron describes it as “jagged.” It can write a poem and still fail at counting the letters in “strawberry.” That unpredictable surface is why almost every generative AI solution KUNGFU.AI builds still needs a human in the loop, and that limits how far it can scale.
The bigger opportunity, according to Ron, is the data a company already owns: supply chain data, customer data, fraud data, transaction history.
“If you go and you train a model on open source data, anybody can do the same.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
That’s a big deal. If your AI is built on the same public data as everyone else’s, you don’t have an advantage. You have a subscription.
Back to that loan system:
“We went from 24 to 48 hours to 9 seconds and they went from a business hour stance to 24/7. AI doesn’t get tired. AI doesn’t get bored.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
Ron Green listed other places where proprietary data pays off for almost any company: product recommendations (including the cold start problem of what to show someone you know nothing about), real-time inventory decisions, churn and sales prediction, and document AI. That last one covers all the workflows where a human still has to check whether a document was signed or a driver’s license is valid. Until recently, AI couldn’t reliably read handwriting or tell a 7 from a 2. Now it can.
How do you know if your data is ready for AI? KUNGFU.AI’s four-stage process
Ron says some early-stage clients still ask whether AI can just fix their data, generate it, or magically understand it. Not yet. Most AI systems today rely on supervised learning, which means they learn from labeled historical examples. If you want to automate loan decisions, you need a history of loan decisions to learn from.
“With AI, you need the data and you are beholden to the quality of your data.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
So KUNGFU.AI follows a four-stage methodology, and I think any company could use it:
- Build the business case. Quantify the ROI if it works, and the performance metric that gets you there. For example, 95% accuracy would let you automate 30% of a decision process and save X million dollars a year.
- Understand the data. Do you have enough? Is it high quality? Is it accessible, or scattered across data stores that are hard to join? Then do feature engineering to see if there’s enough signal to build on.
- Prototype the model. Find out which KPIs are realistic before you commit, so you can bring stakeholders the data, the techniques, and expected performance.
- Build and deploy.
Ron says stages two and three can usually be done in under a month. The mistake he sees most often is skipping straight to the end.
“The number one thing we see businesses do where they make a mistake is they just jump to step four.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
Jessie made the point that marketers do this too. We launch experiments without deciding what metric defines success. Same lesson.
Where are we in the AI revolution?
Ron Green doesn’t think we’re early. He thinks we haven’t started.
“I love to say we’re at day zero. We’re not even at day one.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
His reasoning: there are still so many obvious ideas nobody has had time or money to try, and the pace is accelerating. KUNGFU.AI used to go three or four years before a state-of-the-art shift justified rebuilding a client system. Now it’s sometimes a year and a half. He’s seen companies make $100 million technology purchases and shelve them two years later because cheap APIs could do the same job.
And research is crowded around the same ideas.
“It’s like watching six-year-olds play soccer. The ball’s in the middle and they’re just surrounded by kids and they’re all kicking the ball and the ball’s not going anywhere.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
He also described three stages he sees among companies: AI curious (lots of analysis, little action), “give everyone a chatbot and tell the board we have an AI strategy,” and finally the realization that real value means using proprietary data to automate processes, add predictive capabilities, or create new products. Adam framed that last stage as moving from efficiency to decision intelligence, and Ron agreed.
On agentic AI, Ron pointed to research showing that the length of tasks agents can handle reliably has been doubling about every seven months for five years. If that holds, by 2030 agents could take on work that currently takes a person a week or a month.
“AI will never be as dumb as it is right now.”
— Ron Green, Co-founder and CTO, KUNGFU.AI
What happens when AI runs out of data?
Frontier models were trained on essentially the whole public internet, and the scaling laws worked beautifully until there was no more free data left. Ron’s point is that proprietary data isn’t affected by that ceiling, because those models never had access to it.
He’s also excited about a new lever: letting models think. Reasoning models use chain of thought, checking and revising their own output, and that ability emerged from reinforcement learning rather than human examples. Coding and math come first because answers can be verified. Eventually, Ron expects models to generate their own training problems, solve them, find their weaknesses, and keep climbing.
Ron Green’s tech stack
His favorite thing about AI tooling is that the best tools are free and open source. He recommends PyTorch (from Meta) over TensorFlow (from Google), along with NumPy, pandas, and the broader Python ecosystem.
The bigger picture
Here’s where it gets interesting for anyone in the data business. Ron Green’s whole argument comes down to this: the companies that win with AI will be the ones with data nobody else has, and that data is clean enough to learn from. Public data is a commodity now. Your own data, and the quality of it, is the moat. That’s exactly why we’re so obsessed with data quality at BDEX. AI amplifies whatever you feed it, good or bad.
Connect with Ron Green: Learn more at kungfu.ai, find Ron Green on LinkedIn, and check out KUNGFU.AI’s podcast, Hidden Layers.
Want cleaner data to power your AI and marketing? Visit bdex.com and click “Talk to an Expert” to see how BDEX identity data can help.
Watch the full episode: What’s Next in AI: Beyond Generation and Hype on YouTube
This article was adapted from an episode of Deconstructing Data, BDEX’s weekly podcast on data-driven marketing. Tune in live every Thursday at 4:15 PM Eastern on LinkedIn.
About BDEX: For companies that need clean identity data to power their products, BDEX offers unmatched quality and execution. Visit bdex.com and click “Talk to an Expert” to get started.
