Page 38 - gs260802

Insights and Expertise

The missing link For example, an AI model may perform well in testing

when it's asked to flag invoices that don't match a PO. But
between AI adoption once it enters a live finance environment, it has to account
for supplier bank detail changes, duplicate invoices with
slight formatting differences, temporary approval over-
and AI value rides and other exceptions that a controlled test never
fully captures. That gap between model accuracy and op-
By Jean-Jacques Bérard erational readiness is where many AI initiatives begin to
break down.
Esker
I may be everywhere in business conversa- Another pitfall is assuming the work ends at deployment,
tions, but finance teams are still struggling to when in fact that's where it begins. Sustaining value re-
quires continuous iteration. AI models must evolve along-
turn it into operational value. According to
A research from Payhawk, half of finance teams side the business through ongoing monitoring, tuning and
refinement as processes and operating conditions change.
remain stuck in the middle stages of AI maturity (see
https://tinyurl.com/ytf2yxxm).
As organizations focus on deploying and scaling AI, one
critical factor is often overlooked: the people who will
Most of those efforts stall for the same reason: Teams treat actually use it. Even strong AI tools will struggle to gain
AI as a technology initiative instead of a workflow and
control problem. The challenge is making AI useful in the traction if users don't understand when to trust the tech-
nology, when to challenge it and how exceptions should
workflows where invoices are processed, payments are
approved and exceptions are reviewed, all without intro- be handled. The result is a tool that exists in theory, but
not in day-to-day operations.
ducing new risks to governance, compliance or financial
controls.
Moving beyond the middle stage of AI maturity requires
more than building the technology. Organizations must
Many organizations believe they've solved for this in create the conditions for AI to adapt as business require-
theory. In practice, moving from experimentation to op-
erationalization requires process discipline, data quality ments change and become embedded in the invoice-to-
payment processes finance teams rely on every day.
and clear boundaries for human oversight—elements that
remain a work in progress. Approximately 80 percent of Four pillars for turning AI into
organizations report that limited access to data across en- an operational capability
vironments is holding back their AI initiatives, according
to Cloudera research (see https://tinyurl.com/492j66nr). The transition from pilot to production starts with a strong
foundation. These four priorities can help Finance teams
The technology itself is only one piece of the equation. To turn AI experimentation into a reliable operational capa-
create lasting value, organizations need to identify where bility.
AI can improve invoice-to-payment performance, then
build repeatable processes that increase efficiency without 1. Get your processes in order
creating new blind spots in control or compliance.
Before expanding AI across invoice-to-payment opera-
Why AI initiatives are stalling tions, organizations need a clear picture of how those
The root cause of stalling is how finance teams approach workflows function today. Start by mapping the pro-
AI. Rather than embedding it into existing business pro- cess from invoice intake to payment approval and iden-
cesses, they treat it as a standalone project. The focus be- tify where delays, manual rework, unclear ownership
comes proving the model works instead of determining or exception backlogs are creating friction.
how it fits within the workflows, approval structures and
governance frameworks already in place. This step matters because AI will only accelerate what-
ever process it's placed into. If approval paths are in-
This mindset often leads organizations to confuse a suc- consistent, supplier records are incomplete or excep-
cessful proof of concept with an operational solution. tions are handled differently each time, automation
won't magically solve those problems — it'll simply
A POC can show that AI recognizes a pattern or predicts move those problems through the workflow faster.
an outcome, but invoice-to-payment processes are rarely
that simple. They're shaped by changing approval rules, Once those gaps are visible, strengthen the foundation
compliance requirements, fraud controls, supplier up- beneath them. Clean ERP and supplier data, defined
dates, system dependencies and exceptions that don't fol- approval rules, governance expectations and clear au-
low a predictable path. If those realities aren't considered dit trails give AI the structure it needs to operate safely
from the start, AI remains a promising experiment rather and consistently.
than a business capability.
38