AI Report Digital - Flipbook - Page 13
“I think it’s really important to have a dedicated policy
framework around [AI], notably because of the amount of
evolution that’s going to happen in the space.”
Russell Barrett, Valley National Bancorp
data into an AI tool or model that could later be extracted
One benefit of partnering with existing providers is the famil-
and shared externally.
iarity the latter have with prudential regulators’ expectations, Moody’s Sarkar says. These vendors are likely familiar
As regulators have made clear, financial institutions are
responsible for the actions and outcomes of their third-party
vendors as if they had taken the action themselves. In the
with model validation processes and oversight expectations,
and can produce documentation the institution can share
with regulators. For instance, as part of its responsible
absence of regulatory guidance specific to AI, they should
AI approach, Microsoft lets clients peer into its models to
revisit and potentially update and strengthen their model risk
explain why they may generate certain outcomes and uses AI
governance and validation as they conduct due diligence.
to identify data or model drifting, the gradual change in the
As far as technology goes, there are several data science
techniques that organizations can use to explain the opera-
data properties used to train a model that can degrade performance over time.
tions of machine learning-powered models and address con-
It’s also essential that financial firms stay focused on their
cerns about these models’ fairness and reliability, according to
identified use case and desired outcomes during due diligence,
a July 2023 paper from the nonprofit FinRegLab that looked
says Bud Financial’s Maslaveckas. They should make sure
at machine learning in credit underwriting. Banks and credit
they’re using the right technology on the right problem, and
unions may want to familiarize themselves with these tech-
not be swayed by the compelling capabilities of AI if it’s not
niques as they conduct due diligence on vendors and inquire
appropriate for their application.
about models.
“Not everything needs a large language model or genera“If I was a community bank, I would be very conservative in
tive AI. Sometimes it just needs some really basic machine
my approach. I would not want to be on the bleeding edge,
learning. Sometimes you just need a rules-based system doing
and only look for things [that] have been tried and tested,”
math questions for affordability calculations,” he says. “Don’t
Parmar says.
get blinded by the shiny thing.”
ARTIFICIAL INTELLIGENCE: A REAL-WORLD APPROACH | 11
