Page 39 - gs260802

Insights and Expertise

payment timing and cashflow. As teams gain confi-
Measuring AI after the pilot dence in these lower-risk use cases, they can gradually
expand AI into more complex workflows while main-
taining the controls needed to manage risk.
Getting an AI initiative into production answers
one question: Can the technology work within the 3. Define clear decision-making boundaries
organization's actual workflow? A more impor-
tant question follows: Is it making that workflow Fraud prevention and review controls should be built
better? For invoice-to-payment operations, useful into AI workflows from the start. Define which signals
measures can extend well beyond model accuracy. require review before payment approval, such as un-
Finance teams can compare performance before usual invoice patterns, suspicious supplier or bank ac-
and after deployment across several areas: count changes, or vendor verification discrepancies.
• Processing time: How long does an invoice
take to move from receipt through approval? At the same time, human oversight should remain
Measuring cycle times can reveal whether AI firmly in place for exceptions and higher-risk deci-
is eliminating bottlenecks or simply shifting sions. This applies particularly when supplier records
them elsewhere. are modified, invoices are disputed or transactions
raise fraud or compliance concerns.
• Manual intervention: Track the percentage
of invoices that require employees to correct By identifying risk before funds are released, AI gives
data, resolve exceptions or reroute approvals. teams a better opportunity to investigate and prevent
A system that automates routine work but costly errors from moving further through the pay-
generates excessive exceptions may deliver ment process.
less value than expected.
• Errors and duplicate payments: Automation 4. Establish ownership and accountability
should reduce preventable mistakes rather
than introduce different ones. Error rates, AI initiatives need clear ownership, starting with an
duplicate detection and payment corrections executive sponsor or AI champion. This person under-
provide tangible measures of performance. stands both the technology's potential and the reality
• Exception quality: More alerts aren't neces- of day-to-day invoice and payment operations, allow-
ing them to identify where automation is appropriate
sarily better. Teams can examine whether AI and where human control must remain.
is directing employees toward exceptions that
genuinely warrant attention or overwhelm- Ownership can't sit with one person alone, however.
ing them with false positives. Leaders should provide sponsorship and support so AI
• Control and compliance: Audit findings, un- doesn't remain an isolated experiment. And as AI be-
authorized changes and the ability to recon- comes more deeply integrated across finance systems,
struct how decisions were made can help de- treasury platforms and operational workflows, this ac-
termine whether efficiency gains are coming countability will only become more important.
without weakened controls.
• Employee adoption: A technically successful Taking AI from pilot to production
system delivers little value if employees rou- For organizations that feel stuck between experimenta-
tinely work around it or distrust its recom- tion and operationalization, the answer is not to add more
mendations. Usage and override patterns can tools. It's to strengthen processes, clarify ownership and
reveal where additional training or system take a deliberate approach to where AI can deliver value
refinement is needed. and where human judgment should remain in control.
When those foundations are in place, AI shifts from an
experiment to a dependable business capability that sup-
2. Start with high-volume, low-risk workflows ports the workflows, decisions and controls that keep fi-
nance operations moving.
AI should first be applied where it can deliver clear val-
ue without compromising control. In invoice process-
ing, that often means repetitive, rules-based tasks such Jean-Jacques Bérard is chief product and technology officer at Esker. In
as invoice data extraction, PO matching, approval rout- this role, he implements product strategy and oversees product plan-
ing, duplicate detection and exception flagging. These ning and development. Jean-Jacques joined Esker in 1995 as project
workflows are ideal starting points because they con- leader for the SQL team and later moved to research and development
sume significant time but often don't require strategic manager in 1997. In 1998 he advanced to his current role. Prior to Esker,
judgment. Automating them can reduce manual work Jean-Jacques was R&D team manager at Arthur Andersen Consulting in
and provide better visibility into approval bottlenecks, Lyon. Contact him via LinkedIn at linkedin.com/in/jean-jacquesberard.

39