welcome to the cfo's guide to Ai and machine learning a seven chapter audio guide from Oracle netsuite that provides the cfo's perspective on where they think AI fits in a finance practice whether you are a CFO or a business leader sit back relax and enjoy the guidance on how various Ai and machine Learning Systems function and ideas to harness them for your business [Music] benefit artificial intelligence or AI is a computer science discipline that started in the 1950s and has progressed in fits and starts borrowing from mathematics statistics economics and even philosophy to get where
it is today after what seems like Decades of Promise artificial intelligence now presents a reality that both offers more than we expected and yet seems more dangerous than we foresaw within the short availability of chat GPT it's managed to shake up Academia and Intrigue and frustrate Business Leaders while chat GPT has caught the attention of the masses other instances of AI and more often machine learning have made businesses more productive with the buzz near its peak we surveyed CFOs to learn where they think AI fits in a finance practice their feedback provided interesting ideas and
laid bare misconceptions about what the technology offers and how it works chapter one early Innovations for business some of the most useful AI Innovations are downright dull for instance algorithmic approaches to optical character recognition Peak near 95% accuracy on typical business documents with handwriting recognition far less accurate when these systems looked at 3,000 characters on a typical page the algorithm got 150 of them wrong that left a lot for humans to fix so much for relieving the drudgery of moving data from paper into the digital realm today m machine learning techniques have improved algorithms better
Optical readers also help with the aid of deep learning some optical character recognition OCR Technologies now claim 99.8% accuracy depending on the source material that near 100% rate makes automating business processes like Bill capture a Nob brainer for many companies at the other end of the complexity Spectrum AI tools aimed at Enterprises with gigabytes to terabytes of data may be enhancements to data warehouses and business intelligence systems or they may be Standalone products particularly in the latter case adoption can be a huge commitment one that requires a substantial return on investment or Roi promise no
one implements these products just to see what AI is all about but for companies with a enough data about customer Behavior or other key metrics AI assisted systems can uncover unique insights and correlations not easily found any other way for firms smaller than the fortune 2000 or not in highly data intensive Industries it may seem like the basics of reading and classifying business documents and transactions has been the extent of ai's benefits and that indeed is a huge payoff as it digitizes data and lets leaders start to automate rot tedious tasks however there are intriguing
use cases that involve training AI systems on huge data sets and then applying the learning to individual businesses a good example of this is systems that predict how changing the price of a product will affect retail sales and customer satisfaction determining competitive pricing used to be reasonably straightforward even if it took some leg work you could go a long way just by visiting competitor stores and seeing what stock was being added or removed and when sales were run now with online marketplaces and branded e-commerce sites it's much more difficult to determine what customers will do
in the face of price fluctuations across channels machine Learning Systems can gather competitive intellig Ence by scaring the internet and use it along with your own data on buyer Behavior to determine price elasticity and predict customer Trends sometimes at a personal level what will the market pay for a popup performance in an intimate venue a hand painted bathtub or a 50-year-old bottle of Scotch some machine Learning Systems even help retailers set Dynamic prices maximizing the revenue for goods or services with finite Supply like concert tickets or limited edition items chapter 2 behemoths of the cloud
many consumer AI Innovations have come from the largest tech companies and we've all watched them improve over time Siri debuted with the iPhone 4 in 2011 and Amazon released Alexa in 2015 Google Assistant followed a year later search engines have benefited from AI in several ways not least of which is the ability to correct our Collective poor spelling now email apps are spotting grammar issues and offering fixes or completing sentences on the Fly these are examples of AI systems that learn and get more accurate over time as an area of research AI has been greatly
aided by vast and Powerful cloud computing environments every major cloud provider now has infrastructure offerings and AI software libraries that serve as the basis for creating new AI products and facilitating AI systems training at a more palatable cost still most commercial AI systems require a lot of computing horsepower so much so that commercially viable AI products need to provide a major return on investment in the voice recognition examples we provided the return is millions of people engaging on a regular basis and providing lots of personal data in science the return needs to be major discoveries
that couldn't have practically been made otherwise or assistance with tasks that are too timeconsuming and expensive for humans to do alone for instance pharmaceutical researchers now use AI systems to simulate millions of chemical compound interactions in hopes of developing new drug therapies the AI systems involved are big complex and expensive but they can model countless interactions in minutes and therefore are a logical alternative to the usual lab testing of a much smaller number of compounds selected based on current known science by quickly finding the 10 or 20 protein combinations that warrant human trials AI has
led to the creation of drugs that may never have been found otherwise it's a big return on a very big investment chapter 3 machine learning or artificial intelligence it's not just AI because someone says it is because of the current fascination with artificial intelligence or AI systems there's a temptation to label smart algorithms as AI media Outlets bear some responsibility for calling technology AI when it isn't with others then repeating those claims this is a problem if return on investment depends on the system scaling in a certain way or improving as it's given new data
asking questions about what data was used to train the system and how it will learn from your data will often help identify whether you're dealing with a true machine learning or AI system machine learning or ml is a huge step towards artificial intelligence but it's not the same thing you can show a machine Learning System a few hundred thousand of anything from x-rays to French to English translations the more good and bad examples you show the system the more it refines its understanding of the subject a machine learning system uses models and probabilities that are
refined as it sees examples to become proficient in a task the key tenant of machine learning is that the system adapts and improves as it's given more examples to analyze once Engineers create an ml system for a given purpose they train it by providing examples of the sorts of items that they want the system to evaluate Engineers tweak and perfect the algorithms and eventually end up with a system that's very useful in evaluating say X-rays providing Radiologists with ML tools can lead to faster better and more economical diagnosis in machine learning the system produces better
results as it sees more samples but the algorithms that do the learning don't change unless humans Tinker with them the system we've described doesn't know anything about cancer research but it does know statistically what lung image aberration look like the algorithms use pattern matching and probabilities to guide findings and they're highly effective in that learning process ml systems need a lot of data problems that have many samples and classifiable outcomes are good candidates for machine learning to solve those with computer readable examples are especially ideal take spam filters coming up with training data is as
easy as digging through any raw email stream hitting a busy organization servers if you have a few thousand email users a machine Learning System could become relatively good at spotting spam by looking through several months of messages sorted by spam and not spam one challenge for businesses is determining whether the problem you're trying to solve creates enough data for an AI system to adequately learn and whether that data accurately always describes the condition you want to test the characteristics of spam might change some what but good emails will mostly continue to look like good emails
and bad ones will be relatively easy to spot major email vendors now claim a 99.9% success rate in identifying spam but let's say you want an ml system to tell you whether your company's electric bill is higher than it should be when you start programming perhaps the system has access to bills from the previous 12 months to learn from from so the system will have some idea of how costs vary depending on seasonality what bad data would you give it perhaps bills that are 20% above or below the previous year's bill but each month would
be considered bad but then your business grows you add more equipment people computers and the cost of electricity changes Maybe by more than 20% the system has no basis to understand this context so it flags all subsequent bills for human review your monthly electricity bill doesn't generate nearly enough data for a machine learning algorithm to provide insights you'd be far better off with an Erp system that could automatically look at last year's Bill and see if this year's Bill differs you could set a business rule that says to notify facilities if a bill varies by
more than 20% from the previous year it's still going to flag the bill after you've added new equipment and people but you've only invested 10 minutes to set a business rule versus spending significantly on an ml system that was never going to yield insights you didn't already have and you can easily adjust the business rule to reflect expected growth chapter 4 Ai and ml in Finance as Ai and ml technology matures it's more commonly being embedded in business applications as we've discussed for CFOs one of the first benefits is the ability of systems to recognize
interpret and classify business documents storing their contents as data accounts receivable and accounts payable automation systems use these Technologies to digitize and classify paper and digital invoices the process is considered robotic process automation RPA because once documents are read they are further processed according to business rules that you set it's important to recognize the distinction between process-driven automation which can include everything from spelling and grammar Checkers on the desktop to those image recognition optical character recognition systems we mentioned earlier and datadriven automation which depends on ML and AI to offer insights and guide decision robotic
process automation happens when Finance team members use a system that knows and follows business rules to accomplish tasks for expense reports if you're using a system that captures receipt images classifies expenses and enforces reimbursement rules you've got a good example of RPA in our findings from CFOs many want to leap frog from limited automation to artificial offal intelligence but you can't skip the process automation phase because this is where you teach systems the business rules to follow with the data that they take in without automation there's no way for the AI systems to know what
to do it needs both digitized data and knowledge of your process to accomplish the desired job on its own here's an example many teams spend a lot of time executing repe itive tasks whether it's managing accounts receivables or payables three-way matching expenses running payroll closing the books or any of the other monthly functions automation is an affordable way to save tremendous resources while enforcing your business processes not only is that helpful for finance efficiency it's essential for AI since AI systems can work only with digital data right now automation is the place to start if
you're looking to channel Finance team resources to more strategic tasks like scenario or demand planning fpna and other data analysis the quickest way to reduce days sales outstanding DSO is to automate most of the work that goes into accounts receivable billing and collection actions happen more quickly and predictably the system generates the data you need to get constant updates on how your DSO is tracking and you get early warnings on accounts that are pushing DSO in the wrong direction so you can deal with problems early on once a finance team has defined processes and digitized
business data automation tasks and follow the business rules required to complete them now it makes sense to start considering opportunities to use datadriven machine learning and AI to further improve or even automate operational decis decisions but before diving into datadriven intelligent automation it's important to pause and take stock to avoid mistakes you may have made in the past the market for datadriven AI is shaping up much like the business applications Market has over the past 30 plus years vendor a might excel at AI assisted Supply Chain management vendor B is great at customer sentiment analysis
vendor C can apply AI to Finance and Accounting vendor D has a great solution for AI assisted Inventory management you get the idea engaging with all those vendors leads to lots of Standalone systems that each optimize exactly one piece of the business management puzzle going challenge by challenge to find ml systems that can help is daunting and probably no wiser than it was to go Problem by problem problem to find applications to help solve business challenges but don't discount obvious wins particularly if your business is slowed by one intractable problem just bear in mind that
datadriven intelligence is often best served by collecting many data streams into a single data warehouse that's capable of complicated whatif data analysis and offers techniques like data visualization along with ML tools combined in finance data with operational data web analytics lead generation data store or Warehouse traffic information customer satisfaction Matrix and other business interights lets analysis uncover Trends unique to your company that wouldn't otherwise be discoverable the key is to keep the advantages of process automation As you move into Data driven analysis ideally speciality systems that help with a particular aspect of business can easily
and that is without a lot of custom code or extract transform and load ETL work use the same data store in which you've chosen to keep your business data look for those that do here's where chat GPT stirs the imagination it's expert in lots of things maybe enough so that it or technology like it could subvert the whole evolutionary Journey from paper to digital to process automation to datadriven AI Nirvana enter the world of deep learning chapter five go deep in ml the algorithm remains constant unless a programmer intervenes while the system improves as it
is shown more examples through guided learning the next step is to let the algorithm develop as it learns this is the domain of artificial neural networks that seek to use huge arrays of computers to mimic the function of the human brain in a process known as deep learning in a neural network the idea is to create layers of artificial neurons that solve a problem by successfully getting into more detail pretty much how the human brain works think of it as each layer of the network analyzing one feature of the data its learning if it's looking
at pictures of humans the first layer might determine edges and Lines within the picture while the next layer understands more about Contours eventually you get far enough into the layers of the network that it can identify eyes noses and ears if the goal is facial recognition the algorithm refines itself so that it can identify individuals after seeing many millions of varied images of millions of people just like our own brains the facial recognition process needs to work on massive amounts of data in parallel to get results quickly enough to be useful the CPU of a
typical computer like your laptop or phone might have eight cores meaning it can process eight streams or run eight algorithms simultaneously it does so very quickly but in a task like image processing data needs to be processed in parallel not eight streams but hundreds of thousands in a human brain millions of neurons may be working at the same time to allow us to instantly recognize the things that we see a revelation in 2012 was that we could use graphical processing units gpus rather than general purpose CPUs to analyze pictures that technology combined with very large
data sets and some algorithmic advances particularly in the learning phase led to vastly improve accuracy and identifying images after training gpus have hundreds of thousands of cores and can access memory very fast meaning they can efficiently process lots of data in parallel in many ways they're ideally suited to neural net designs they're far better suited to analyzing the very large training sets than are CPUs data sets can be analyzed in Days by gpus that would take years with the same number of CPUs over the past decade Nvidia and other vendors have refined their architectures to
create chips specifically designed for deep learning applications now deep learning is leading to the sorts of mindblowing results that we see from chat GPT the generative AI systems can pass medical and bar exams because everything it needs to know to do so can be found on on the internet and in other learning data sets it's also a very good coder and writes in clear English with few errors chapter six reduce to practice for business chat GPT works on a huge scale Microsoft reportedly spent hundreds of millions just to create the Computing infrastructure where the system
was trained and now lives to put this in context the King James version of the Bible runs 4.13 megabytes meaning the algorithm studied text amounting to more than 800,000 King James Bibles in that training set are medical and legal texts along with many many programming language examples and explanations on the web plus a lot of general knowledge content and while version 3 was trained only with text version 4's training will include images you can have a conversation with chat gpt3 it will answer questions and beyond that build on previous responses for purposes of answering basic
questions and writing content for business or other purposes it's fairly good if you need a refresher on business practices say the pit fulls of intercompany transactions it can provide that this version of chat GPT can't audit your books and flag where interc compan transactions haven't been handled properly because it doesn't have access to your books and won't inherently know how to read them finance and other Business Leaders would love an AI system that checks things like interc compan transactions or better yet one that checks by following the fasb and GAP rules that chat GPT seems
to understand asking chat GPT to describe how to comply with accounting rules is one thing having a system like it that applies those rules to your books is unfortunately another it understands regulation so creating a system that does the work seems well within reach chat GPT is already advising you not to try it without software tech companies are scrambling to release products that leverage chat GPT to build systems that access the technology through apis or that use similar technology those systems could read and interpret your books and apply an aid driven accounting tool to comply
with fasb rules and your business practices though the technology used probably won't be generative AI your AI accountant is somewhere on the horizon the question will then be how much do you trust your AI accountant and how will you review view its work to ensure the rules that you've set for managing your books are followed CFOs and controllers will start out with caution so understanding how to monitor the work AI does for you will be a big part of adoption chapter 7 ethical and practical concerns ethical concerns abound particularly as AI systems start making decisions
that affect lives in substantial ways or create output that encroaches on the copyrights and intellectual property ownership of humans discussions of these issues are important complicated and nuanced ethicists lawmakers and technologists are raising warnings about potential issues that could arise from a headlong Rush towards widespread use of AI in all sorts of applications an open letter encouraging a pause in development of giant AI experiments starts out with these two sentences AI systems with human competitive intelligence can pose profound risks to society and Humanity as shown by extensive research and acknowledged by top AI Labs as
stated in the widely endorsed ayoma AI principles Advanced AI could represent a profound change in the history of life on Earth and should be planned for and managed with commensurate care and resources the letter has been signed by more than 28,000 people so far including many leading thinkers in technology business and academic research many of the signers have pointed to the unintended negative consequences on Society of unregulated social media and postulate that unregulated AI could have a much greater effect how will this affect your industry here's a short list of areas to consider AI in
healthcare as in many areas some of the lead concerns in healthcare revolve around whether existing legal structures will be sufficient to assess responsibility for the actions of AI systems AI in law enforcement two questions tend to dominate law enforcement concerns the first is around privacy AI has the potential to use huge amounts of data to let law enforcement track what citizens are up to there's an obvious potential for abuse of such power the second concern revolves around implicit bias training sets may not include enough data about minority groups as a result false positives including in
facial recognition are more likely in those groups possibly leading to to improper actions by law enforcement AI ethics in the legal profession lawyers have a professional obligation to fully understand and agree with and stand behind the output of their practice whether it's creating contracts offering opinions on disputes or citing case law that would guide actions on behalf of a client AI offers a way to reach legal opinions more quickly however particularly in deep Learning Systems the tools operate as a black box which is to say the algorithms used to create output is purely mathematical and
therefore not intelligible to humans in a way such that we can know if the output was derived in the right way as a result lawyers can't just trust what AI tells them vetting is required AI ethics in banking here again the blackbox nature of systems leads to concern AI will advise Bankers on issues like whether clients are suitable for loans or whether the bank's risk portfolio is well managed but without understanding how conclusions were reached it's impossible to know whether implicit bias played part in more complex matters of risk management AI systems could propose portfolio
structures that are so complex that bankers and Auditors cannot assess whether they'll meet the risk goals of the institution this very short list provides a sampling of the questions that must be addressed industry by industry and use Case by use case for more reading on the topic Stanford offers a paper entitled ethics of artificial intelligence and Robotics and Harvard Business Review offers another on how AI might be regulated titled AI regulation is coming the bottom line we're only beginning to understand the ramifications of AI both good and bad there are many many questions to be
answered about its use but the promise is alluring vendors won't necessarily have all the answers so smart decision makers will carefully consider their choices and watch efforts to restrict what AI is allowed to do curious to see what netsuite can do for your business head over to Nets s.co to discover more and get in touch today
