Using AI Capabilities in Payables

https://www.youtube.com/watch?v=BbtjwXslmpY – oracle pervasive AI

1. Understand how to leverage Oracle’s pervasive AI
approach in Oracle Cloud Applications

Pervasive AI at Oracle Open World

September 27, 2019 | 5 minute read
Emma Hitzke
Senior Product Marketing Director, Emerging Tech
Since joining Oracle earlier this year, I’d been hearing a lot about Oracle Open World (OOW) and
how this year’s OOW was going to be quite different, and I am not just talking about the fact
that the streets around Moscone Center were not blocked to the traffic!
While there were a lot of announcements and great customer stories to pay attention to, I want
to take a moment to reflect on artificial intelligence and machine learning (AI/ML). More than
200 (!) sessions and demos made it clear that AI/ML is no longer hype. It’s here, it’s pervasive,
and our customers are using it. While there’s so much to cover about AI/ML, I’m highlighting
what caught my interest during a whirlwind week in San Francisco.
AI/ML is elevating everyday actions across customer businesses
Let’s start with the session “How to Turn Change into Opportunity with Oracle Cloud Apps“ led
by Steve Miranda, Executive Vice President for Oracle Applications product development, who
said: "Our Machine learning helps you become more efficient and realize data insights you may
not even do before. Ultimately you have people still driving those organizations and business
decisions. Every step of the way, adding Machine Learning into all your transactions to make
them much more effective.”
 I
like this quote not only because it makes clear how people, even when interacting with AI,
remain at the heart of processes and decisions, but also because it boils AI/ML down to true
business benefits, and this is the level where Oracle has focused its development efforts by
providing a dual path for adoption. Other technology companies either focus solely on AI-based
business solutions or on custom development of ML models but Oracle offers both.
Oracle customers can speed time-to-value with AI-embedded applications that deliver optimized
business recommendations; intelligent user interfaces that deliver more personalized
experience; and digital assistants that extend the value of enterprise applications. Oracle
customers can also build custom ML solutions by using our integrated PaaS and IaaS services.
And at the infrastructure level, Oracle is embedding ML in its cloud services for improved
customer experiences by offering recommendations to lower cost though automating
maintenance, streamlining workflows, and improving security with automated patching and
predictive security analytics.
Why have we taken this approach? Because our focus is on getting customers to
the benefits of AI/ML faster—no matter their route or resources. I witnessed the
power of this flexible approach when I spoke with Cammila Yochabell, CEO and Founder
of Jobecam. She shared that “Oracle has changed our [company’s] life” with AI, digital assistant,
better data security and storage, to support her vision: how to alter a person’s image and voice
during video interviews so that unconscious bias is removed and candidates are chosen for
positions on their merits alone. She even got a chance to share her story with Larry Ellison
himself!
I heard another compelling story from Miranda Nash, Vice President of Product Management at
Oracle. Miranda recounted how an Oracle customer had a major “a-ha” moment grasping the
continuous transformative power of machine learning. The customer was attending a
demonstration of AI Apps for ERP and realized that as the Accounts Payables team was
completing tasks in the application, the intelligence embedded in the application was learning at
the same time from the human decisions and was adjusting future recommendations and alerts.
It’s like having two teams of people working at the same time, complementing each other, and
jointly feeding improvement and efficiency back into the finance process.

Some takeaways from keynote and sessions

For anyone who has been to OOW, you know that there are so many great sessions that it’s hard
to get to all that you want to attend. Here are some snippets from my favorite sessions and links
to the recording or presentations:
 Ask the right questions about AI: Senior Vice President of AI Apps Clive
Swan’s session covered the key questions to ask when embarking on an enterprise AI
project. Failing to take the time to answer them, and you may risk to feel the impact of
"what does fast cost you?" He also warned of the traps of AI hype.

 From Data Tsunami to Smart Data: "It has become impossible for humans to handle
and interpret huge volumes of data. That's where AI comes into play," explained Melissa
Boxer, Oracle VP Product Management, in her session about The AI-Powered Marketer.
 Hungry for meaningful and accessible insight: Too many companies suffer from
“data rich but insight poor” syndrome. "Tell me something I didn't already know" is the
underlying promise of how emerging tech applied in HR can lead to new insights,
predictions, and make work more human, according to Kai Petzelt, VP of Product
Marketing HCM at Oracle. Applied AI/ML is already built into Oracle’s IoT apps to provide
insights (e.g. anomalies, predictions, trend detection) and actionable recommendations.
"Now you have data science in a box!" exclaimed Bhagat Nainani, Group Vice President
for IoT and Blockchain Applications Development. For example, Titan International is an
Oracle customer benefiting from the access to Oracle Cloud and advanced technologies
to improve their inventory process and get on their path to smart manufacturing.
  AI contributes to business agility within the finance function: "We are building AI
and analytics into our EPM applications with regular updates to enable a more agile
organization,” said Jennifer Toomey, Senior Director, EPM Cloud Product Marketing in her
“Value Realization of EPM in the Cloud” session. Our customers can drive better decisions
through analytic insights and automate manual processes. Brian Pawlus, Senior Director
of Product Management for Oracle AI Apps, gave an overview of how AI and data-driven,
ready-to-go applications for ERP help you manage supplier health, and build intelligent
and self-learning dynamic discounting programs using trusted data sources, decision
science, and a modern cloud infrastructure. Steve Cox, Group VP for ERP/EPM Product
Marketing, challenged the audience to forget the 4.8 day benchmark for financial close,
and aim for continuous close! By leveraging emerging tech into streamlined business
processes, you can simply revolutionize the finance function.

Embedded AI capabilities in Oracle Fusion Cloud Applications

Organizations can quickly and easily adopt the latest AI capabilities to help improve productivity
and decision-making, automate some complex tasks and end-to-end business processes, and
help reduce costs. Explore the AI features by business function below.

Leverage intelligence and automation with embedded AI to support supply chain decision-
making and optimize supply chain management and manufacturing processes. All delivered as
standard in Oracle Fusion Cloud SCM.

Functional
Use case Business benefit
area

Order Predict ETA and identify at-risk shipments for

Predict shipment
Management customers and provide total transportation lead
transit lead times
and Logistics time when order is placed.

Order
Predict order fulfilment Automatically calculate average order cycle
Management
cycle times times and show impact vs required ship dates.
and Logistics

Classify requisitions, purchase orders, invoices

Procurement Spend classification and expenses using AI for faster, more granular
analysis of procurement data.
Functional
Use case Business benefit
area

Categorize suppliers and enroll them into a

dynamic discounting programme. Early payments
Procurement Dynamic discounting are rewarded with larger than normal trading
discounts to enhance working capital and
optimize ROI on cash.

Broaden your supplier pool through an auto-

New supplier
generated list of new suppliers. Add them to your
Procurement recommendations for
supply base and increase competition for your
negotiation invitation*
business so you can agree better deals.

Generate negotiation Simplify the creation of content used in supplier

Procurement
cover page* negotiation documents.

Supply Chain New product Forecast demand of new products learning from
Planning introduction the past product launches.

AI powered forecasting for demand

Supply Chain encompassing a wide range of data including
Demand sensing
Planning history, price, seasonality, intermittency, demand
shocks, anomalies and external events.

Cope with supply variability and improve supply

Supply Chain
Lead time insights planning decisions by analyzing actual supplier
Planning
lead times.

Help customers find your products faster by auto-

Product
Create item generating product descriptions that emphasize
Lifecycle
descriptions* key features and benefits and make the customer
Management
experience clearer and more satisfying.
Functional
Use case Business benefit
area

Predict asset health and communicate issues

Manufacturing
Product anomaly quickly to facilitate timely action. Monitor assets,
and
detection fleet and manufacturing equipment to detect
Maintenance
anomalies.

Manufacturing Generate list of

Fast-track the creation of a quality inspection
and inspection
plan.
Maintenance characteristics*

Manufacturing
Generate shift notes Facilitate the smooth hand off of work to the next
and
for operator hand off* operator.
Maintenance

Manufacturing
Generate repair Streamline the process of repairing multiple
and
summary* assets on a single work order.
Maintenance

2. Use AI to automate intelligent account combination defaulting

Oracle ERP Payables Adaptive AI IACD feature not training as expected
Summary:
In our NonProd, we configured the Aaptive AI feature , Intelligent Account
Combination Default feature, with Promo code in the evaluation report.
However, we see that the combination is not defaulting fully except one
segment (which is always 0000).

1. At first glance, our Business is not happy. We are looking to give full
demo after we train the model for few suppliers so that the CC will default
as it learns when we type in Non-Prod. To train the model:
In Non-prod instance, I created 2 to 3 invoices with same supplier/Supplier
Site and provided the correct Code combinations. I expected that with this
training, system (as document mentioned real time), it pulls the data and
learns from recent invoices and the next 4th invoice will generate with
defaulting code combination. But it didn't happen.

Could you answer the below points:

a. How can we train the model once it is configured? any documents to go
through it.
Though it is NonProd(where we configured) and system uses the Prod
data(after refresh date) to train and generate evaluate report, does the
training (creating invoices) needs to do in Production though the model
configured to Non-Prod?

b. Even in training the model after its implementation, is there any

frequency (like daily once) that the Adaptive AI pulls the data from ERP
Cloud to learn/train?

c. Does these Adaptive AI apps needs to train independently for each

instance. Or can we move the same trained model of Non-Prod to Pre-Prod
and Prod.

d. Any architecture diagram for this ERPCloud to Adaptive Intelligence Apps

(Where does these Adaptive AI apps resides).
Link for the faq’s
https://support.oracle.com/cloud/faces/DocumentDisplay?
_afrLoop=122461215322058&_afrWindowMode=0&id=2917668.1&_adf.ctrl
-state=ecmmf1ozg_263

https://docs.oracle.com/en/cloud/saas/adaptive-intelligent-apps/aigsd/faqs-
for-intelligent-account-combination-defaulting.html#When-does-the-
Intelligent-Account-Combination-Defaulting-feature-predict-values%3F
I will try to address the questions that you have raised in your message
above.
 a. Training the Model
 How can we train the model once it is configured?


 The model is trained automatically every 24 hours once the feature is

activated if new data is available. However, in a test environment, the
model only uses data prior to the P2T Refresh date. This is primarily so
that the model is only trained on Production data and not test invoices
and the associated patterns in test.
 If you did not have sufficient relevant data prior to the P2T Refresh
date, is there any other more recently copied environment that you
can use or test after the next P2T Refresh?
 We do not recommend this, however if you want to demonstrate the
functional flow only, you can create test invoices in a test environment
and set the P2T Refresh date after the invoice creation date so that
the model can learn on these invoices. Please allow up to 12 hours
after setting the new P2T Refresh date until the model is retrained.
 Frequency of Data Pulls for Training:


 What is the frequency with which the Adaptive AI pulls data from ERP
Cloud
for training after the model has been implemented? Specifically:
o Does the data get sent from ERP Cloud to the model each time
an invoice is validated (either online or via a batch process)?
o Or is there a scheduled frequency (e.g., daily) at which this data
transfer occurs?
In Production, the model is retrained every 24 hours if new validated data is
available. In test, the same still applies however only new data prior to the
P2T Refresh Date is considered.
c. Training Independence for Each Instance
1. Independence of Training Across Instances:
 Do these Adaptive AI applications require independent training
for each instance (Non-Prod, Pre-Prod, Prod)?
 Alternatively,
is it possible to move a trained model from Non-Prod to Pre-Prod
and
subsequently to Production? Please provide guidance on this
process.
Models cannot be moved from one instance to another. However the model
is always trained on the historical data available in the respective
environment (up to 18 months of history).
d. Architecture Diagram
1. Architecture Overview:
 Could you provide an architecture diagram illustrating the
interaction
between ERP Cloud and Adaptive Intelligence Apps? Specifically, I
am
interested in understanding where these Adaptive AI apps reside
within
the overall architecture.
AI Apps environments are provisioned to customers in the same region as
the related Fusion ERP Cloud environment. The environment are
customer/pod specific and are only integrated with the respective customer
pod. The data ingestion mechanism is automated and currently leverages
BI Publisher (this may change in the future). The data (and model) is stored
in a customer/pod specific OCI object storage and prepared for model
training. No data is shared across customers or environments. An API call
from the Fusion environment invokes the prediction from the AI Apps Model.
e. Support for Custom Invoice Sources
1. Support for Various Invoice Sources:
 The FAQ mentions that "all invoice sources are supported."
Previously, I
noted that support was limited to MANUAL/IDR sources. Could
you confirm
whether this indeed includes custom invoice sources as well?
The models are trained on models from all invoice sources with the
exception of the invoice source "spreadsheet". Also invoices that have more
than 25 lines or invoice lines that are booked against a distribution set or
have multiple distributions are excluded from model training. All other
invoice lines within the last 18 months can be considered in the model
training.
a. Training the Model
Today is 01-Oct-24, and in our Non-Prod environment, we set the P2T
refresh date to 01-Jul-24. For training the model, it considers the last 18
months of invoice history. Specifically, it includes:
1. From Prod: It reads invoices with invoice_date later than 01-Jul-24
through to 01-Oct-24.
2. From Non-Prod: It reads invoices with invoice_date from 01-Mar-23
(18 months prior to today) to 30-Jun-24.
i. To train the Non-Prod model (as part of convincing the business for
approval), if we create invoices in Non-Prod today with invoice_date ranging
from 01-Mar-23 to 30-Sep-24, these new invoices should be included in the
training from Non-Prod. is it?
ii. And if I change the P2T refresh date to today in NonProd then all 18
months invoices, to train the Non-Prod model, will be considering from Non-
Prod only . is it?
iii. And also I believe it takes only the validated invoices (though not
accounted).
iv. I believe the new data for training is based on the invoice_date and not
the creation_date. Can you confirm if this understanding is correct and
provide any additional comments?
iv.
d. Architecture Diagram
Your explanation was very helpful. We encountered an issue with assigning
'Dynamic Discounting' Adaptive AI application roles to a user. Even our
Admins, who possess the necessary access, are unable to see these
application roles in 'Oracle Cloud Services' when navigating to the Adaptive
AI Apps page from ERP Cloud.
Since the AI Apps exist in the same POD, I assume the admin access for
these PODs is equivalent to the admin access for SaaS ERP Cloud. We
logged an SR two weeks ago but have not received a satisfactory response
yet. If you have any insights or can share related information, that would be
appreciated. Alternatively, I can start a separate thread for this issue.
e. Support for Custom Invoice Sources
Thank you for confirming that this feature applies to all invoice sources
except for one. To elaborate on its practical application:
 We can create only Manual Invoices from the UI where it defaults.
 For IDR, similar defaulting behavior might be applicable.
 For other custom sources, typically created using FBDI, how does
defaulting work? Specifically, should we omit the Expense account for
lines so that the importing program defaults them? Will this approach
ensure that the segments default correctly and the statuses remain
valid?
f. Default from Invoice Distribution Combination or Accounted
Combination (Post-SLA)
1. Invoice Accounting Combination:
 We create invoices (using FBDI) for custom sources with a
specific code combination for the expense account. Later, while
executing 'Create Accounting,' SLA rules apply, updating the
 initial invoice distribution code combination with the final, SLA
rule-based code combination.
 Does the IACD train the model based on the initial Invoice
Distribution Code Combination, or is it the final accounted Code
Combination post-SLA application?
2. Training Attributes:
 While defaulting the Code Combination, do the training attributes
include POET information along with Supplier/Supplier Site,
Business Unit, Legal Entity, Header-Line description, etc.?
a. Training the Model
Today
is 01-Oct-24, and in our Non-Prod environment, we set the P2T refresh
date to 01-Jul-24. For training the model, it considers the last 18
months of invoice history. Specifically, it includes:
1. From Prod: It reads invoices with invoice_date later than 01-Jul-24
through to 01-Oct-24. - No, the AI Apps features only ingest data from
the related Fusion environment for model training. Hence only data
from non-prod is ingested to train the model in non-prod
2. From Non-Prod: It reads invoices with invoice_date from 01-Mar-23
(18 months prior to today) to 30-Jun-24. - yes only the data from prior
to the P2T Refresh date will be used up to 18 months from today's
date
i. To train the Non-Prod model (as part of convincing the business for
approval), if we create invoices in Non-Prod today with invoice_date ranging
from 01-Mar-23 to 30-Sep-24, these new invoices should be included in the
training from Non-Prod. is it? No, the date criteria for ingestion and model
training is the creation date, not the ingestion date. If you want to train a
model on new data that you generate in test, then you will have to move
the P2T Refresh date after that date. So if you enter data on the 2nd
October, set the P2T Refresh date to 3rd October. However, this is not
recommended as test data will also be included in the training set. Also the
new data that you create will only be in the evaluation but not in the
training set of the model. The feature is optimised for Production and to
test based on Production data. The feature is not optimised to support to
test on new training data as this typically is not a true reflection of the
model that customers will have in Production
ii.
And if I change the P2T refresh date to today in NonProd then all 18
months invoices, to train the Non-Prod model, will be considering from
Non-Prod only . is it? Yes correct, see above
iii. And also I believe it takes only the validated invoices (though not
accounted). Correct
iv. I believe the new data for training is based on the invoice_date and not
the creation_date. Can you confirm if this understanding is correct and
provide any additional comments? The date criteria is based on the
creation date of the respective invoice distributions
iv.
d. Architecture Diagram
Your
explanation was very helpful. We encountered an issue with assigning
'Dynamic Discounting' Adaptive AI application roles to a user. Even our
Admins, who possess the necessary access, are unable to see these
application roles in 'Oracle Cloud Services' when navigating to the
Adaptive AI Apps page from ERP Cloud.
Since the AI Apps exist in
the same POD, I assume the admin access for these PODs is equivalent to
the admin access for SaaS ERP Cloud. We logged an SR two weeks ago but
have not received a satisfactory response yet. If you have any insights
or can share related information, that would be appreciated.
Alternatively, I can start a separate thread for this issue.
AI Apps have a separate IDCS stripe and only the default IDCS
Administrators are provisioned with the respective IDCS Administrator roles
in this new stripe. If you send me the pod name in a Direct Message, I can
check if the engineering team can find the respective resources. However,
if the default IDCS Adminstrators are not known, raising a SR is the correct
process
e. Support for Custom Invoice Sources
Thank
you for confirming that this feature applies to all invoice sources
except for one. To elaborate on its practical application:
 We can create only Manual Invoices from the UI where it defaults.
 For IDR, similar defaulting behavior might be applicable.
 For
other custom sources, typically created using FBDI, how does
defaulting
work? Specifically, should we omit the Expense account for lines so
that the importing program defaults them? Will this approach ensure
that
the segments default correctly and the statuses remain valid?
The prediction service can currently only be invoked from the invoice line
and is not invoked during the invoice import. Hence the spreadsheet source
is excluded from model training as it will be more efficient to enter the
account code combination in the FBDI spreadsheet. Also the feature is
optimised that values are only defaulted if the dynamically calculated
confidence score threshold is reached. Hence for some invoice lines the
segment values are not defaulted. The feature is intended to drive a high
level of accuracy of the data that is captured in the Account Combination
but support the user by defaulting the values with sufficient confidence.
f. Default from Invoice Distribution Combination or Accounted
Combination (Post-SLA)
1. Invoice Accounting Combination:
 We
create invoices (using FBDI) for custom sources with a specific
code
combination for the expense account. Later, while executing
'Create
Accounting,' SLA rules apply, updating the initial invoice
distribution
 code combination with the final, SLA rule-based code
combination.
 Does
the IACD train the model based on the initial Invoice Distribution
Code
Combination, or is it the final accounted Code Combination post-
SLA
application?
The model is trained based on the initial code combination and not post
SLA. This is to default the same values as are visible in Payables. Hence in
your scenario, if the SLA rules derive the correct code combination in all
instances, it will be more efficient to maintain the existing process
1. Training Attributes:
 While
defaulting the Code Combination, do the training attributes
include
POET information along with Supplier/Supplier Site, Business Unit,
Legal
Entity, Header-Line description, etc.?
POET information is currently not used. The full list of attributes used in

model training is available here: Using AI Apps for Financials

See question: What data does the machine learning model use for providing
predictions using Intelligent Account Combination Defaulting

Question 2 – Enable Intelligent Account Combination Defaulting in

Invoices REST Endpoint
Description (Required):
For non-PO AP invoices, the Distribution Combination needs to be
entered manually, and it's resource intensive for the users to enter
them to system.
Oracle Fusion Payables includes a feature called "Intelligent
Account Combination Defaulting", which utilizes AI to predict the
correct account combination from historical data. The predicted
value is entered as a default, which the user can then modify if
required.
Currently, the feature is enabled only when manually entering the
Distribution Combination for AP invoice line.
However, utilizing the AI model is not supported when importing
AP invoices from external sources. Especially in the Nordics, AP
invoices are almost always imported from external sources, and
they are very rarely created manually by the users.
Use Case and Business Need (Required):
Create a built-in feature to the Invoices REST Endpoint to predict
the Distribution Combination using the same AI model than used in
the "Intelligent Account Combination Defaulting" feature.
For instance, the feature could work so that if the Invoices REST
Endpoint is called to create an invoice, and the invoice line doesn't
include either Distribution Combination or PO reference, then the
AI model is called automatically to create the default Distribution
Combination.
The benefits of implementing this feature include:
 AP invoice would include Distribution Combination suggestion
directly after import, which the users can then verify through
Approval workflows
 Users can complete all their actions from the Approval
notifications, and no need to go to the full AP invoice user
interface
 Significantly reduce invoice entry time and accelerate the
process of making invoice ready for payment
 Reduce “think time” for Accounts Payable analysts
 Allow user to focus on data accuracy rather than data entry
 Improve the user experience
  Allow AI to improve the prediction accuracy for segments by
learning from the user edits

This would be a huge process improvement. My company is 90%

non-PO transactions and even with the IACD feature, all invoices
still need to touched. If, at a minimum, those invoices where a full
combination is predicted it can be automatically added to the
invoice after it is imported through image processing, this would
significantly reduce time spent on keying and would also be cost
benefit as we have third parties keying invoices due to the volume.
=============================================
==========================
AI feature like Invoice Document recognition
We would like to know if there is any AI feature that is readily available
where data is recognized from the document received and created in oracle
applications. For example like 'Invoice Document Recognition (IDR)' in
Accounts Payables where invoices are created in Oracle application based
on the PDF sent by the supplier.
This AI document recognition feature that we are asking for and would like
to explore, can be from any module and not restricted to only
Payables/financials. It can be from any other modules as well like Supply
chain, HCM, etc.,
Also, if the requested feature is not readily available to explore, we would
like to know if there any such features in the pipeline which might be
available/implemented in future.
----------------------------------------------------------------------------------------------------------
--
We have built this solution for Fusion apps. The solution can be found on
the Oracle Cloud Marketplace
OAN Business Docs (Content Management)
https://cloudmarketplace.oracle.com/marketplace/en_US/listing/161024285
Here is the video of the Fusion manage attachment feature (After extracting
the data an linking it to ERP MDM or Transaction):
https://oappsnetdocmdev-oappsnetgroup.cec.ocp.oraclecloud.com/
documents/link/LD49FF713C7D94652C256A65FEE0EE5C97CE959D7D00/
fileview/D0E27CA48ACC32E0EF82393CE7A720D7052BBA857B5F/
_OAN_Business_Docs.mp4

Here are two product specific solutions found on the Oracle Cloud
Marketplace
OAN Accounts Payable for Fusion Apps
https://cloudmarketplace.oracle.com/marketplace/en_US/listing/154474834
OAN Sales Order Automation
https://cloudmarketplace.oracle.com/marketplace/en_US/listing/146393995