In brief
Smaller lenders such as credit unions and regional banks can compete with big banks by using AI, but only if the hard parts are done for them. Jay Budzik of Zest AI argues that what scales is a repeatable process for building models, not one pre-trained model. That process is paired with pre-purchased third-party data and a guided, automated tool.
The tension is between vendor promises of push-button simplicity and the real work of adopting AI. Host Daniel Faggella keeps pushing back on the MailChimp comparison, noting that model building still involves research and tuning. Budzik concedes the user must still monitor results, but says automation lets business users, not data scientists, run the process.]
Questions
How can a small credit union use AI without a big data science team?
Budzik says Zest AI interviews the lender about its products and goals, sources suitable data, and builds the model automatically. This replaces the 100-person data science team some large banks have.
Why can't a vendor just train an AI on a big bank and reuse it everywhere?
Faggella says that raises ethical and legal issues and the lessons transfer poorly. Budzik adds that populations differ, so a model trained on one group may not fit another.
What if a small lender doesn't have enough of its own data?
Budzik says the lender's data is combined with other data to form a larger, more representative sample. That sample can include people who applied but were denied, or whom the lender hopes to reach later.
Is Zest AI's software really as simple as MailChimp?
Budzik compares its setup wizard and monitoring dashboards to MailChimp. Faggella disagrees, saying the underlying research and tuning are more complex than checking email open rates.
How does Zest AI work alongside other AI platforms?
Budzik says many customers train models on a general AI platform and then use Zest AI's software to check compliance, improve fairness, produce reports and monitor the model. This works because most models are built in Python.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Zest AI helps lenders, from large institutions to credit unions, train algorithms that detect risk and set loan approvals or terms.
“Zest helps to train algorithms to detect risk and determine approvals or terms for different kinds of lending products for all kinds of lenders, from gigantic financial institutions down to credit unions.”
0:01:23 · Daniel Faggella · Lending AI
Jay Budzik, CTO of Zest AI, had appeared on the podcast about three years before this episode.
“Jay was a guest on the show some three long years ago, so it's a pleasure to be able to have him back.”
0:00:32 · Daniel Faggella · Guest history
Zest AI was previously called ZestFinance.
“a company that was once called Zest Finance.”
0:00:32 · Daniel Faggella · Company history
Jay Budzik says there are tens of thousands of small lenders that need help accessing AI technologies.
“There are tens of thousands of them and they need help accessing these complicated technologies in order to win against those competitors.”
0:03:54 · Jay Budzik · Market size
Jay Budzik says large lenders use AI for marketing and loan origination, which lets them make more competitive offers.
“They're using AI for marketing. They're using AI to originate loans. And so they can make offers that are more competitive often.”
0:03:19 · Jay Budzik · Large bank AI
Jay Budzik says smaller lenders have mostly consumed analytics products from vendors such as FICO and Vantage.
“while the smaller ones have been kind of focused on consuming analytics products from vendors like FICO and Vantage.”
0:04:47 · Jay Budzik · Credit scoring
Jay Budzik says Zest AI has learned which data sources are relevant for different lending segments and products.
“we have the map of all the data sources that are relevant and that produce the performance in Lyft.”
0:06:52 · Jay Budzik · Data sources
Jay Budzik says the best data for marketing can differ from the best data for underwriting.
“So a marketing data source might be different than an underwriting data source”
0:06:52 · Jay Budzik · Data sources
Zest AI pre-purchases and aggregates third-party data and makes it available to customers through its tailored modeling process.
“we can aggregate or pre-purchase lots of data and make that data available to our customers through our tailored modeling process.”
0:07:24 · Jay Budzik · Data strategy
Jay Budzik says some large institutions employ data science teams of about 100 people to build and validate models.
“having a 100-person data science team like they do at some of the institutions you mentioned in order to build models and validate them”
0:09:49 · Jay Budzik · Data science teams
Jay Budzik says small lenders often lack enough observations to build robust AI models on their own data.
“They don't often have enough observations or data to be able to make robust AI models.”
0:11:29 · Jay Budzik · Data limits
Jay Budzik says the FICO score is trained on a national population.
“if you think about FICO, it's trained on a national population.”
0:13:36 · Jay Budzik · Credit scoring
Zest AI offers a wizard-like setup process for building a first lending model, which Jay Budzik compares to MailChimp.
“we've developed this wizard-like process, much like a MailChimp, where you can go and set up your first lending model.”
0:15:14 · Jay Budzik · Product design
Zest AI offers an API so larger customers can use its automation inside their own data science environments.
“We also offer an API, much like MailChimp does, for the larger customers to be able to access the power of the automation within their own data science environment.”
0:15:41 · Jay Budzik · Product design
Zest AI's model management system has three applications named build, adopt and operate.
“our model management system, which contains applications build, adopt, and operate.”
0:17:02 · Jay Budzik · Product design
Jay Budzik says most AI models are built in Python, which allows different systems to interoperate.
“most of these models are built in Python. And that provides a way of interoperating between various systems.”
0:19:06 · Jay Budzik · Interoperability
Jay Budzik says many Zest AI customers train models on another platform and then use Zest AI software to analyze them for compliance.
“we have many customers that train their model using a platform company and then use our software to analyze the model for compliance”
0:19:53 · Jay Budzik · Interoperability
Jay Budzik was scheduled to join the AI in Business podcast for a series on measuring AI return on investment.
“Jay is going to be part of a series on measuring AI ROI in the coming two months.”
0:21:33 · Daniel Faggella · Upcoming series
The AI in Business podcast's previous series on AI culture change began with an episode featuring IBM's chief AI officer.
“our kickoff episode was the chief AI officer at IBM.”
0:21:51 · Daniel Faggella · Podcast series