Page 26 - GS211001

CoverStory

brought into the system. The remaining 6 percent needing
Throughout their ongoing additional analysis is possibly due to a subpar scan that
requires a data entry person to manually enter almost the
training, AIs that are unable entire statement, he stated. He also acknowledged that a
significant percentage of incoming statements need cor-
to interpret line items are rections, and if they are not corrected, would be virtually

redirected to experts who impossible to analyze.
help them master specific Talapan concurred that human and machine collabora-
tions boost accuracy, stating a fast 30-second analysis that
items and rules. is 80 to 90 percent accurate is a good starting point, espe-
cially if it gives humans the flexibility to go in and change
a few things. "While the ideal solution would produce
sub-second automated results that are 100 percent accu-
Data integrity rate, in reality, statements are very complicated and most
human analysis is not 100 percent accurate either," he said.
Monica Eaton-Cardone, founder and chief operating offi-
cer at Chargebacks911, a global chargeback management
and remediation firm, noted that interpreting raw data According to Talapan, Fee Navigator partners are mostly
happy with automated analysis, but some of them like to
can be a daunting task for humans and machines alike,
especially when it comes to chargebacks. go in and change a few things. For example, he said, they
may want to manually enter a few data points that were
not retrieved from paper statements, but even with these
"The crux of the problem with chargebacks is preserving add-ons, they still save hours and are able to get back to
data integrity," she said. "If you have bad data, it doesn't
matter what you're looking at; you'll get a bad result. And sign or service the merchant very quickly. Hence, a fast
30-second analysis with 80 to 90 percent accuracy may be
this becomes challenging when not everybody uses the
same words, when there are different languages being a good starting point if it allows a human the flexibility
to spend a few minutes to increase that accuracy to 100
used and different interpretations of rules, which creates
layers and layers of complexities." percent, he added.
Cash discounting
The first goal in preserving data integrity, from Eaton-
Cardone's perspective, is training AI models to interpret Talapan mentioned that cash discount programs, de-
data correctly. Throughout their ongoing training, AIs signed to help merchants recoup processing fees, should
that are unable to interpret line items are redirected to ex- make merchant statements simpler and easier to under-
perts who help them master specific items and rules. stand but that is not always the case. "The industry intro-
duced more cash discounting and reduced the use of tier
"You start with something simple, with the goal of get- and bill-back, which should result in simpler statements,"
ting as much data as possible on a given case, whether he said. "Instead, we have seen markups hidden inside
it's images or digital files or merchant information, and interchange, dues and assessments, as well as additional
then interpreting all the data to make it as standardized fees being added."
as possible," Eaton-Cardone said. "And then you can lever-
age automation, which is not a perfect system but can ac- Noting that hidden fees make statements appear simple
curately interpret 90 percent of cases on the acquiring side when they are not, Talapan said lack of transparency and
and quite a bit on the issuing side as well." paper-based statements that many merchants still receive
by mail can significantly impede statement analysis. How-
Text, image recognition ever, he noted, Fee Navigator scans paper documents and
Shepherd has also seen highly accurate character and im- applies best-of-breed optical character recognition (OCR)
age recognition by his company's AIs but admitted not technologies to extract data and convert it to digital for-
every picture is perfect every time, despite advances in mats and gets excellent results despite these setbacks.
technology. "You don't want to be in a position where you
have to reject a statement because your image recognition Shepherd also observed that merchant statements don't
can't recognize it," he said. "So, in our company, we try to always indicate cash discount savings or credit card sur-
automate as much as we can and we have a team of people charge fees. In some cases, statements from cash discount
doing data entry work when necessary." acquirers do not even show a discount rate, just monthly
fees and overall transaction volume, he stated.
On a typical high-volume day, Shepherd estimated, 94
percent of all statement line items his company receives As a general practice, Shepherd recommended skipping
represent some kind of a fee that text recognition correctly statement analysis when promoting cash discount pro-
grams and relying on it heavily when selling credit card

26