AI in Treasury Newsletter – Recap July 2023
Welcome to The AI in Treasury Insight Newsletter! In the ever-evolving world of finance, AI in Treasury is fast becoming a cornerstone of modern practices. This advancement, marked by the integration of artificial intelligence into treasury functions, is revolutionizing how companies manage their financial assets. From the ease provided by AI chatbots in interpreting complex policies to the precision of machine learning in credit assessments, AI in Treasury is reshaping the industry landscape. It streamlines operations, enabling teams to focus on strategy rather than getting mired in routine tasks.
As you read on, you’ll understand how AI in Treasury not only enhances efficiency but also brings real-time insights, such as those seen in credit scoring and corporate ratings. But with these innovations come challenges, including concerns about data privacy and transparency.
To provide a clear picture, this article breaks down the intricacies of these advancements, highlighting their benefits, risks, and applications in understandable terms. It’s crucial to stay updated because, in this fast-paced world, staying ahead means understanding and embracing these transformative tools.
AI chatbots as your inhouse Policy Experts
In this Month’s edition, we will discuss AI-powered chatbots trained on your company policies, and explain you step-by-step how to create your own Policy-Based Chatbot.
And we will do so by taking the example Automation Boutique has walked us through!
Let’s break down how they see AI-powered chatbots and the value it could bring already today to Treasury Departments!
Please note that this edition is not sponsored, we just found their solution interesting and worth talking about!
Introducing AI Powered Chatbots
Treasury, like all corporate functions, has piles of Policies for professionals to constantly be referring to.
Capital Structure, Financial Risk Management, Cash management, Governance management policies… and much more!
All documents that new employees onboarding into your company need to read, digest, embody… but most importantly apply! Costing the organisation hundreds of hours.
How can AI Powered Chatbots help?
In the (very close) future, having to search through hundreds of pages for that one policy item which you need to complete your transaction will be a thing of the past.
This is where AI-powered chatbots trained on your company policies come into play.
Here is a demo made by our friends at Automation Boutique (this newsletter edition is not sponsored, but we definitely recommend you check out what Philip and Jan-willem are doing in treasury automation!)
The Power of Policy-Based AI: Supercharging Treasury Operations
Imagine you train an AI with all of your policies, so when you need an answer, you don’t need to pile through pages of text but rather you just ask your AI expert.
Benefits of Policy-Based Chatbots
The benefits of such a tool are pretty clear, but here’s a summary:
Instant and Consistent Policy Guidance: Chatbots provide instant responses to policy-related queries, ensuring consistent and accurate information is readily available.
Treasurers can rely on chatbots to handle policy-related inquiries, reducing response time and increasing efficiency.
For example: Compliance policies, regulatory guidelines, and risk management procedures.
This helps treasurers ensure adherence to regulations and promote a culture of compliance across the organization.
Improved Customer Experience: Policy-based chatbots enhance the customer experience by providing prompt and accurate information 24/7. Customers can access policy-related details, such as payment terms, compliance requirements, or investment guidelines, at their convenience, leading to enhanced satisfaction and reduced friction in communication.
For example: A policy-based chatbot can assist suppliers or internal stakeholders with questions regarding payment terms, invoice submission, or clarification on specific billing requirements.
It ensures prompt responses, reduces manual intervention, and fosters smoother payment processes.
Streamlined Treasury Operations: Overall, by offloading routine policy-related inquiries to chatbots, treasury teams can save time and resources that can be better utilized for strategic tasks.
Chatbots can handle repetitive queries, allowing treasury professionals to focus on more complex and value-added activities.
So the AI will not replace your treasurers, but supercharge them and make their time more valuable.
Creating a Policy-Based Chatbot: A Step-By-Step Guide
But how do you actually go about building a policy-based chatbot model? I’m sure by now, you have the general idea about how to build an AI: gather data, clean the data, train your model, test it and go!
Data Gathering: Building the Foundation
Collecting policy documents, FAQs, and other relevant resources is the initial step. This data serves as the foundation for training the chatbot model. Remember, garbage in = garbage out. These need to be up to date, accurate and complete.
Natural Language Processing (NLP): Understanding User Queries
Implementing NLP techniques allows the chatbot to understand user queries and extract intent and entities.
NLP helps in identifying keywords and context, enabling accurate responses. That will usually be in the form of a pre-existing one (like GPT4) which you bias with your information and train it.
You would not typically build a new NLP from scratch, as it is a lot of setup work and like using a sledgehammer to put up a picture frame.
Training the Model: Teaching the Chatbot
Using machine learning algorithms, the chatbot model is trained on the gathered data. That usually is done by mapping user queries to appropriate responses based on policy guidelines. Once your mapping is done, the NLP knows what each part of the policy means and which types of questions it relates to.
Testing and Iteration: Ensuring Accuracy and Performance
The trained chatbot model is tested with a variety of queries to ensure accuracy and improve its performance. Iterative improvements and fine-tuning are done based on user feedback and ongoing evaluation.
Integration and Deployment: Making the Chatbot Accessible
Once the chatbot model is ready, it can be integrated into existing communication channels, such as websites or messaging platforms. Deployment ensures users can access the chatbot and benefit from its policy-based guidance.
Machine Learning in Credit Scoring with Datrics
Let’s discuss Credit Scoring with Machine learning. And we will do so by taking the example of Datrics.ai, who has created an end-to-end data science platform enabling the creation of customized AI apps and models for all business aspects, including AI in credit scoring.
Let’s break down their solution and see how AI is used already today.
Credit Scoring: A Quick Refresher
Credit scoring is a statistical method that lenders use to quickly assess the creditworthiness of an individual or business applying for a loan.
This score is based on the data available on their credit history, including their previous and current loans, repayment behavior, and any late payments or defaults.
The financial health of a company also contributes to its credit scoring. Metrics such as Profitability, Liquidity, Solvency, Operational Efficiency, Cash Flows and Business Risk are thoroughly looked at.
Machine Learning: A Quick Refresher
Machine Learning, a subset of Artificial Intelligence (AI), is a method of data analysis that automates analytical model building.
In simpler terms, it is a system that can learn from data, identify patterns, and make decisions with minimal human intervention.
Machine Learning algorithms use computational methods to “learn” information directly from data without relying on a predetermined equation as a model.
When AI meets Credit Scoring
Unlike traditional methods that rely heavily on a borrower’s past performance, AI-based scoring systems pay heed to real-time aspects of a borrower’s financial status, such as current income level, employment prospects, and potential earning capability.
As a result, even individuals with limited or no credit history can gain access to credit, providing they show promising financial potential. Consequently, this allows lenders to make more precise profit predictions based on intelligent AI models.
How can Rating Agencies & Banks leverage that?
If we take this a step further, we can clearly make the parallel with Corporate Rating! Here is how AI can revolutionize corporate credit rating:
Advanced Data Analysis with AI
AI can efficiently analyze vast and diverse datasets, including traditional financial metrics and unstructured data like news articles, social media sentiments, industry trends, and macroeconomic indicators.
For example: It can process real-time news or market sentiments about a company, which might affect its creditworthiness but may not be captured in traditional financial statements.
Real-time Corporate Credit Assessment
Traditional corporate credit ratings are often static and rely on periodic financial statements.
AI allows for real-time updates, factoring in the most recent financial data, market conditions, and even operational changes.
For example: If a company just won a large contract or suffered an unexpected loss, AI can quickly adjust the credit rating to reflect these developments.
Predictive Modeling in Credit Rating
AI uses machine learning to generate predictive models, which forecast a company’s financial behavior based on historical data and trends.
This predictive capability can anticipate potential credit risk shifts and defaults.
For example: AI could predict a downturn in a company’s sector leading to increased credit risk, helping treasurers make more informed decisions.
Customized Risk Profiling with AI
AI provides a more tailored risk assessment by taking into account each corporation’s unique circumstances, industry, and behavior patterns.
For example: A tech startup might have different risk factors compared to an established manufacturing firm, and AI can customize the credit rating process accordingly.
Efficiency Boost with AI
AI application speeds up the credit rating process, leading to faster decision-making.
For example: instead of waiting weeks for manual credit assessment, AI can provide results in a matter of minutes or hours. This not only improves treasury operations but also enables corporate treasurers to manage their credit exposure more effectively.
AI for Better Risk Management
AI’s ability to predict future financial behaviour, identify trends, and provide real-time credit assessments helps corporate treasurers manage their risk more effectively.
For example: by predicting a potential default, AI can help treasurers proactively adjust their credit portfolios and avoid high-risk exposures.
What are the downsides of using AI in Corporate Rating?
- Data Privacy and Security Concerns: AI models require large amounts of data to function effectively. In the process of gathering this data, companies need to ensure they comply with data protection laws and maintain the confidentiality of sensitive information. Any breaches could lead to substantial financial and reputational damage.
- Model Transparency and Explainability: AI models, particularly those based on deep learning, can be “black boxes,” with internal workings that are hard to interpret. Regulators and stakeholders often demand transparency in credit rating decisions, which can be challenging with AI models.
- Data Quality and Accuracy: The effectiveness of AI is highly dependent on the quality and accuracy of the data it’s fed. Inaccurate or biased data can lead to flawed credit ratings, possibly leading to wrong financial decisions.
- Regulatory Compliance: Regulatory standards for credit ratings are strict and often don’t fully account for AI methodologies. Companies must therefore navigate a complex regulatory landscape when implementing AI in credit ratings.
- Dependency and Over-reliance on AI: Over-reliance on AI models can lead to complacency in risk management. While AI can provide valuable insights, it’s important for humans to maintain oversight, corroborate AI findings with independent analysis, and ensure decisions are sound.
- Cost and Implementation Challenges: Implementing AI in an existing credit rating process can be complex and expensive. The cost includes not just the development and maintenance of the AI system, but also potential training for employees and the need for continuous adjustments as business needs evolve.
How to Prompt Like an AI Treasury Expert
Now we will be explaining what prompt engineering is, and how you can become a better prompt engineer to help you maximise the use of Generative AI in your treasury department.
We’ll even give you 8 ChatGPT prompts you can try which will help you become a better treasurer!
What is prompt engineering?
Simply put, prompt engineering is how you ask your NLP (Natural Language Processing) AI model or any generative AI model to give you an output.
The quality of your prompts will dictate the quality of your results. Garbage in equals garbage out!
Prompt engineering has been on the rise as a discipline since the popularisation of AI (largely brought by ChatGPT and Midjourney).
Some people are better at asking AI questions than others. This comes down to playing with the tool, you need to experiment to understand what works for you.
A Few Prompt Engineering tips
Here are some prompt engineering tips based on our experience with learning AI at Corporate Treasury 101.
- Give the AI a title: You are a treasury analyst at a S&P 500 company…
- Give the AI a role: … Your role is to provide me with 5 examples of how hedging can be implemented in a treasury department…
- Set constraints: … Focus the examples on software companies which have a reporting entity in the united states but distribute to EU and Canada…
- Remind it not to hallucinate: “don’t hallucinate”.
Read your response, judge if it meets your expectations and tweak as you go!
7 ChatGPT prompts that will take your Treasury skills to the next level!
- “You are a financial mentor. Your task is to explain financial concepts. Explain the difference between cash flow and net income in simple terms.”
- “You are an investment encyclopedia. Your task is to provide information. Explain the benefits and potential risks of investing in corporate bonds.”
- “You are a financial consultant. Your task is to help me make informed decisions. Provide an overview of the main factors to consider when assessing the financial health of a potential supplier or partner.”
- “You are an economic analyst. Your task is to provide insights. Describe how an increase in inflation rate might affect a company with bonds as their primary debt instrument.“
- “You are a financial history teacher. Your task is to teach about past financial crises. Describe the causes and effects of the 2008 financial crisis.”
- “You are a financial tutorial creator. Your task is to guide me through processes. Walk me through the steps of creating a risk assessment for a potential investment.”
- “You are a financial strategist. Your task is to stimulate strategic thinking. Provide some strategies a company can employ to increase their cash flow.”
- “You are a financial data interpreter. Your task is to make sense of financial data. Help me understand this financial statement: […provide statement…].”
Please note that you need to provide the financial statement in a text format. Meaning, copy pasting the financial statement and not uploading a file directly in ChatGPT.
As always, DO NOT share any confidential information on chatgpt. And always apply your own reasoning to the responses!
Deep Diving into Deep Learning: A Corporate Treasurer’s Swim Among Neural Networks
Let’s break down what Deep Learning and Neural Networks are in Artificial Intelligence!
What is deep learning?
Imagine trying to prepare a complicated recipe for the first time.
Deep learning is like having an expert chef whispering instructions in your ear, getting better with every dish you try!
Computers learning from experience
Deep learning is like teaching computers to learn from experience and understand the world in terms of a hierarchy of concepts.
Just as we learn from the ground up, starting from identifying basic shapes to recognizing complex structures, deep learning allows computers to build layers of concepts.
It’s a subset of AI, which means while all deep learning is AI, not all AI is deep learning. Think of AI as the universe with deep learning being one of its star systems!
How does it actually work?
Let’s get Technical!
- Neural Networks: Deep learning operates using artificial neural networks, which are algorithms inspired by the structure of our brain’s neurons. These networks can learn and make decisions on their own.
Yes, we break down Neural Networks a bit later on in this newsletter 😉 - Layers: Neural networks have layers of interconnected nodes (similar to brain neurons). The depth (number of layers) in these networks is why it’s called “deep” learning.
- Feeding Data: When we feed data (like images) into the network, it passes through these layers. Each layer processes an aspect of the data and passes it on to the next.
- Weights & Biases: As data moves through the network, each connection has a “weight” which adjusts as the system learns. Think of weights like the network’s reasoning while making a decision.
- Training: By feeding tons of data and adjusting weights, the network is “trained”. If it makes a wrong prediction, it goes back, tweaks the weights, and tries again.
- Activation Functions: Within each node in a layer, there’s a decision-making process (mathematical) determining what information to pass on. This helps add complexity to the learning process.
- Backpropagation: This is the magic behind deep learning. If the network makes an error in its prediction, backpropagation adjusts the weights to minimize the error during the next prediction.
We will break down how the system knows it did an error in a future newsletter 😉 - Iterations: Deep learning requires lots of data and iterations. With each iteration, the model gets better and reduces errors in its predictions.
What are Neural Networks?
Neural Networks are like a team of experts in a room. Each expert (or “neuron”) has a specific piece of information. They pass messages (data) to each other, refining and combining this information. At the end of their discussion, they come up with a collective decision (the output).
Just like a treasurer might consult various departments for data before making a financial decision, a neural network consults its neurons to produce a prediction.
Over time, as the network gets more data, these experts (neurons) get better at making decisions, optimizing the company’s financial strategies.
How Does This All Tie Back to Corporate Treasury?
Fraud Detection
- Imagine your financial system guarded by Sherlock Holmes. Neural networks act somewhat like this detective.
- By training on vast amounts of data, they learn to spot irregularities and suspicious activities that might be invisible to the human eye. So, every time there’s a transaction that’s a bit ‘off’, it’s like Sherlock’s intuition kicking in, spotting something amiss in a sea of normalcy.
- This super-detective can discern patterns across millions of transactions, ensuring the company’s assets remain safe from fraudulent activities.
Credit Risk Modeling
- When a company wishes to evaluate the financial reliability of another business, it’s akin to assessing the quality of an apple among thousands.
- Some apples look perfect on the outside but might have flaws hidden away. Similarly, a business might seem profitable but could have underlying risks.
- Neural networks, trained on myriad business profiles, can recognize subtle patterns and nuances. It’s like having an apple expert in the treasury team, ensuring that every business transaction or partnership is with a ‘good apple’ — a company that’s financially sound and trustworthy.
Bonus: High-frequency Trading Strategies
(Not 100% Corporate Treasury related)
- Deep learning, with its intricate web of expert neurons, can assess, process, and predict market movements at a pace and precision no human could match.
- Think of it like having a seasoned chef in your kitchen who doesn’t just suggest which dish to prepare next, but also predicts which ingredient will be in demand next month. In the trading world, this helps in making split-second decisions that can lead to profitable trades.
How AI Will Be Embedded into Your Everyday Tools
Let’s break down why your first real life interactions with AI will be through embedded AI. Why thats a good thing and a look into examples that Google Sheets and PowerBI are already using.
Why Embedded AI is the Future
As treasurers grapple with the complexities of the AI revolution – and although we will continue bringing you insights into it overall – it’s worth noting that the first brush with AI you’re likely to experience will be through embedded AI.
But what is embedded AI?
Essentially, it’s the integration of AI capabilities directly into the tools and systems treasurers already use daily, such as ERP or TMS platforms or even your excel spreadsheet. This is in contrast to stand-alone AI applications which operate independently – like ChatGPT.
It should be said, the most likely interaction/ change in your daily life from AI will come from it being embedded into the tools and applications you already use today.
Let’s go through some advantages of embedded AI vs stand alone AI tools:
Familiarity and Ease of Use:
With embedded AI, treasurers won’t need to navigate unfamiliar systems or drastically change their workflows. AI’s power is harnessed within the familiar interfaces of the systems they already use.
Seamless Integration:
Embedded AI offers the promise of plug-and-play. It’s designed to work seamlessly with existing systems, reducing the need for complicated integrations or alterations in the current tech stack.
Real-time Insights:
Embedded AI can provide real-time insights and recommendations directly within the work environment. This can include everything from anomaly detection in invoices to predicting cash flow trends.
Reduced Implementation Barriers:
With AI integrated into existing tools, treasurers can bypass many of the barriers to AI adoption, such as cost and complexity of standalone AI systems.
In short, embedded AI allows treasurers to enjoy the benefits of AI’s processing power and predictive capabilities without needing to step too far out of their comfort zone, making it an appealing first step into the world of AI.
But what does that look like? How is it different to today? Watch the GIF below.
As one of the most widely used business analytics tools, Power BI has made significant strides in incorporating AI capabilities to help treasurers analyze and interpret vast amounts of financial data.
PowerBI has already integrated AI analysis into it’s tool, allowing you to get help analyzing it as if you have a data analyst talking to you. Now, it doesn’t just visualize the data, it helps you interpret it.
This is a form of embedded AI: an existing tool with AI capabilities built into it. Not a new AI tool.
Here’s how Power BI’s embedded AI can help:
AI Visuals
Power BI introduces AI visuals like the Key Influencers visual and the Decomposition Tree. The Key Influencers visual helps identify factors that drive a metric you’re interested in, such as what influences late invoice payments. The Decomposition Tree allows you to break down a data point in various ways to understand the driving factors behind it.
Automated Machine Learning
With Power BI’s automated machine learning feature, you can build machine learning models directly within Power BI, without requiring any specialized knowledge. For treasurers, this could mean predicting cash flow trends or forecasting financial performance based on historical data.
Natural Language Querying
Power BI’s Q&A feature allows users to ask questions about their data in natural language and receive immediate answers in the form of relevant visuals or tables. This allows treasurers to get quick insights without having to manually sift through data.
Dataflows with AI
Power BI Dataflows now support AI transformations. This means you can clean, transform, and enrich your data using AI insights, all within Power BI. Treasurers can use this feature to clean invoice data or categorize transactions automatically.
You can imagine the same capability in your excel spreadsheet (something Google sheets is already doing) – giving you functionality such as:
- Smart Fill and Smart Cleanup: These features can autocomplete data and detect inconsistencies or errors, ensuring that your invoice data and financial records are accurate and consistent.
- Explore Feature: Just type a question about your data, and the AI will respond with an analysis. For treasurers, this could mean asking about trends in expenditure or identifying top vendors by spend, all with a simple question.
- Pivot Tables: AI suggests Pivot Tables based on your data, which can help identify patterns or insights, such as understanding how expenses are distributed over time.
Understanding Machine Learning In Treasury
Machine Learning is a term that’s been popping up quite frequently in our recent editions.
It’s about time we break it down, wouldn’t you agree? We’ve listened to your feedback and in this week’s newsletter, we will decode the concept of Machine Learning (ML). Rest assured, we’re keeping it treasury-centric and jargon-free!
Unraveling Machine Learning: The What and the Why
Imagine having a colleague who:
- Learns from your company’s historical data
- Identifies patterns
- Uses these insights to make accurate predictions
That’s essentially what Machine Learning is. It’s a subset of AI where computer systems learn from data, improve with experience, and make decisions without being explicitly programmed.
Why should you, as a Treasury Professional, care?
Well, it holds the potential to revolutionize your treasury functions, from cash flow forecasting, risk management, to working capital optimization.
Machine Learning vs. Traditional Methods: What’s the Difference?
Think of it this way. Traditional software is like following a set recipe. You input the ingredients (data), follow the instructions (programmed rules), and you get your dish (output).
Machine Learning, on the other hand, is like a master chef. Given the ingredients (data), it experiments, learns from each attempt, and ultimately creates dishes (outputs) that get better each time… And that you potentially never thought of!
How to concretely apply ML to Treasury?
Let’s take the example of cash flow forecasting, a task that could be quite a headache for most treasurers, and definitely at the top of many’s agendas. Traditional methods often involve manual data input, rigid rules, and a lot of guesswork. With ML, you can automate data collection, learn from past patterns, and predict future cash flows with a higher level of accuracy.
The ‘Learning’ in Machine Learning: A closer look
The ‘learning’ in Machine Learning means that the algorithms can improve over time. Let’s relate this to your foreign exchange hedging strategy. An ML model could learn from historical exchange rate fluctuations, market trends, and the timing of your past hedges to make optimized hedging recommendations. Over time, as it gets more data, its predictions get even better!
Machine Learning: Not a silver bullet, but a valuable tool
Despite its advantages, it’s essential to remember that ML is not a cure-all. For instance, the quality of its output depends largely on the quality of data input. Moreover, like every new technology, integrating ML into your systems would require an initial investment of time and resources.
Still, given its potential to enhance accuracy, efficiency, and decision-making in treasury operations, Machine Learning is definitely a tool worth considering in your treasury tech stack.
What else does ML require?
You may be thinking right now “well let’s implement Machine Learning everywhere! Well, it’s important to know that while the rewards can be great, the initial setup does require some investment.
- The Right Data: One of the fundamental requirements for Machine Learning is a good dataset. The algorithms learn from data, so the quality and relevance of your data are paramount. This might mean you’ll need to clean up your existing data or invest in gathering new, relevant datasets.
- Computational Resources: Machine Learning algorithms often require significant computational power, especially as the volume of data increases. You may need to consider investing in powerful hardware or cloud-based solutions that can handle these requirements.
- Skilled Personnel: Implementing Machine Learning isn’t a plug-and-play solution. You’ll need people with specific skills, such as data scientists and Machine Learning engineers, who can understand your data, select and apply the right algorithms, and interpret the results in a way that’s meaningful for your business.
- Software and Tools: There’s a range of software and tools available for Machine Learning, from programming languages like Python and R, to libraries and frameworks such as TensorFlow and PyTorch, to cloud-based Machine Learning platforms. The right choice will depend on your specific needs and the skills of your team.
- Time and Patience: Machine Learning is an iterative process. The first model you build won’t be perfect, and you’ll need to spend time refining it based on its performance, which requires a certain level of patience.
Embracing Machine Learning is a journey, not a sprint, but it’s one that can lead to significant rewards in terms of improved accuracy, efficiency, and decision-making capabilities in your treasury operations.
The dangers of using ChatGPT in your Treasury Department!
Let’s break down how ChatGPT can be a threat to your company’s data and Treasury Department.
Leveraging Generative AI Models in Treasury: Potential and Pitfalls
Generative AI models (like ChatGPT) offer huge advantages to treasurers. You can use them for research, train them to give you treasury advise and even use them as junior analysts to assist you in making conclusions from your data.
However, the buzz around AI isn’t all positive. And you need to know the dangers, pitfalls and potential risks to your business of using these tools. There have even already been very public examples of big companies such as Samsung who have had data leaks via ChatGPT.
Samsung Bans Employee Use of Generative AI Tools Citing Security Concerns
Samsung Electronics has imposed a ban on employee use of generative AI tools such as ChatGPT following an incident where sensitive code was unintentionally uploaded. The company has expressed concern over data security, as information shared with AI platforms can be difficult to retrieve, delete, and may be disclosed to other users.
A recent internal survey revealed that 65% of Samsung’s staff believe such AI tools present a security risk. While Samsung works on creating its own internal AI tools, the use of external generative AI systems is prohibited, and any violation of this policy may result in termination of employment.
A Closer Look at Publicly Available AI Models: ChatGPT and Bard
We’re talking today specifically about the publicly available NLP/ Generative AI’s out there, such as ChatGPT (OpenAI) and Bard (Google). We’ll also talk about how to overcome them!
The main risks around using ChatGPT are around the use of your data and how you use the answers.
Navigating the Data Dilemma: Understanding AI Data Usage
ChatGPT and other generative AI models use your chat data to improve its model. Whilst this is a great way for the AI to continuously improve, it means any data you share with it will be saved in the model for all 100 million of its users. If you input sensitive financial data or policy data into it, it may show up in someone else’s responses in the future! Never share confidential data with these models.
The AI Hallucination: Between Reality and Fabrication
Publicly available generative AI models may exhibit the risk of hallucination, where they generate deceptive or misleading information that appears realistic but is actually fabricated or inaccurate.
This is due to limitations in training data or algorithmic biases, and the models will interpolate between data points it already has and try to give you an answer.
This is why at Corporate Treasury 101 we always say that AI always needs an expert treasury professional to check the responses.
Data Security and Confidentiality: Keeping Your Data Safe
When using external generative AI models, treasury departments may need to share sensitive financial data, such as cash flow projections, supplier information, or investment strategies.
This raises concerns about data security, confidentiality, and the risk of unauthorized access or breaches – as your data could be intercepted as it leaves your secure company firewalls, goes to the ChatGPT servers and comes back.
Example: Sharing cash flow projections with an external AI platform could expose proprietary financial information, potentially compromising the company’s competitive advantage.
Compliance and Regulatory Considerations: Navigating the Legal Landscape
Treasury departments must adhere to strict compliance and regulatory requirements, such as KYC (Know Your Customer) or anti-money laundering regulations. Using publicly available AI models may introduce compliance challenges, particularly when dealing with sensitive financial data. The data you share on ChatGPT is not private.
Example: Sharing customer transaction details with an external AI model could potentially violate data privacy regulations or financial compliance standards.
Lack of Customization and Control: Finding the Perfect Fit
Publicly available AI models may not be tailored to the unique needs and intricacies of treasury operations. This lack of customization and control can hinder the ability to fine-tune the models according to specific treasury requirements or integrate them seamlessly into existing systems.
Example: Applying a generic AI model to cash management may not capture the nuances and intricacies of treasury processes, leading to suboptimal cash flow forecasts or risk management decisions.
AI for Account Payable Automation
Let’s explore how AI can automate and simplify your invoice data entry and Account Payable (AP) solution. We will be taking the solution proposed by many existing vendors, one of which is Aavenir – a Saas company offering. Let’s break down their solution and see how AI is used already today.
AP Management Problem-Statement
Unfortunately, there is no one template for all vendors to submit their invoices in.
This leaves the back-end, AP department of most companies with highly manual tasks of going through each individual soft or hard copy of each document, extracting key information and entering it into their ERP to ensure vendors are paid.
A manual, low value and repetitive task – some of the key points for automation to come in! Today’s manual invoice processing systems are riddled with issues
- Complex Invoicing: Large companies often deal with a barrage of invoices arriving via various methods—EDI documents, PDFs, email, or even paper copies. Consolidating these formats manually is time-consuming and error-prone, with issues such as duplicate or incorrect payments often arising.
- Data Inconsistencies: Invoice data must be cross-checked against purchase orders, goods received notes, and contracts—a tedious, error-prone process. Even the most diligent team can miss key details, which can lead to overpayments and other risks.
Hidden Costs: Manual invoice processing isn’t just labor-intensive—it’s costly. With expenses ranging from manpower to materials, processing can cost between $12 and $30 per invoice. Hidden costs, like missed discounts or late fees, can add up, too.
AI-Powered Invoice Processing
AI-powered invoice processing is an innovative solution that uses machine learning (ML), natural language processing (NLP), and optical character recognition (OCR) to automate and streamline the invoicing process.
Here’s how:
- Data Capture and Digitisation With OCR: OCR converts invoices in various formats (paper-based, PDFs, electronic invoices) into digital format – so that the computer can read it as an input. It takes the image, and identifies patterns on it that look like letters, words and numbers and writes them onto the image in text format. Advanced OCR even handles different languages and formats
- Classification and Extraction of Information: Post-digitisation, the AI system classifies information and extracts relevant details like vendor name, invoice date, line items, totals, etc. This is where NLP and ML shine, interpreting human language in invoices and learning from each processed invoice to improve accuracy over time.
- Verification and Validation: The system cross-references the extracted data with your existing database to ensure the data’s correctness. Sophisticated AI can even flag potential duplicates or fraudulent invoices.
Integration with Accounting Software
After data validation, the AI integrates with your ERP or TMS, automating the update of financial records and minimising the risk of errors.
Machine Learning Feedback Loop: AI’s power lies in its ability to learn and improve over time. Every correction made by humans feeds back into the system, improving future accuracy.
AI in Dynamic Discounting with Taulia
Taulia is a financial technology business that provides supply chain finance and dynamic discounting services. They have had AI implemented in their solution since 2019!
Harnessing the power of AI
- Their AI starts by analyzing data available inside the company’s ERP about their payment terms and past transactions with vendors and customers.
- It then combines this with large third parties data sets where they have trends related to payments behavior under different payment conditions.
Creating Intelligent Insights
- Based on the insights gained, companies can determine the most suitable payment terms for individual suppliers and the level of flexibility required in managing those terms.
- This analysis, conducted alongside AI, can help determine the appropriate APR to offer suppliers.
- “What is APR?”, you may be wondering…
- APR stands for Annual Percentage Rate. It refers to the annualized interest rate charged on short-term borrowing or earned through short-term investments.
A Practical Example
- For example, a buyer implementing a self-funded dynamic discounting program might find that offering an 8% APR results in a 20% likelihood of a specific supplier accelerating payments, whereas a 6% APR increases the probability to 25% but with a lower net yield
- Conversely, increasing the APR to 10% may result in only a 1% chance of supplier acceptance.
Optimizing with AI
By leveraging AI and carefully considering APR options, buyers can optimize their strategies to strike the right balance between acceptance likelihood and financial outcomes.
Fine, but what is happening behind the scene?
Behind the scenes, model selection plays a pivotal role in building effective AI solutions for working capital and supply chain financing.
Here is how you can go about implementing AI models in your treasury department.
What are we solving for? Understanding the Objective
Group treasurers should clearly define their objective, such as optimizing cash flowforecasting for dynamic discounting in supply chain financing.
For example, the goal might be to accurately predict supplier acceptance rates for early payment offers.
What do we base ourselves on? Assessing Available Data
Evaluating the quality and availability of data is essential. For dynamic discounting in supply chain financing, treasurers would assess historical payment data, invoice details, supplier profiles, and market trends. Sufficient and relevant data is crucial to train the AI model effectively
Example: Treasurers analyze payment history, invoice attributes, and supplier characteristics to identify patterns influencing supplier acceptance rates for dynamic discounting.
AI But which one? Exploring Model Options
Treasurers should explore various AI models suited to their objective, such as:
- Regression models
- Decision trees
- Neural networks
- etc…
Each model offers unique capabilities in capturing different patterns and relationships within the data.
Example: Decision trees can help identify the key factors, such as invoice amounts, payment terms, or supplier relationship, that influence suppliers’ decisions to accept dynamic discounting offers.
Are we not overcomplicating it? Balancing Interpretability and Performance
Striking the right balance between interpretability and performance is crucial. While more complex models like neural networks offer high accuracy, you can never see the “thought process” the model took to get there. Simpler models like decision trees provide transparency and interpretability.
Example: Neural networks can capture intricate patterns in supplier behavior for dynamic discounting, but decision trees offer clear decision paths based on easily interpretable rules for early payment predictions.
Does it work? Validation and Testing
Prior to implementation, thorough validation and testing of the selected AI model is necessary. This ensures that the model performs well on unseen data and aligns with desired outcomes.
Example: Treasurers validate the AI model by comparing its predictions against actual supplier acceptance rates for dynamic discounting offers over a specific period, ensuring accuracy and reliability.
Conclusion
The journey through the realm of AI in Treasury has unveiled a plethora of opportunities and challenges. What’s undeniable is the transformative power AI holds for the treasury sector. From streamlining mundane tasks with chatbots to harnessing deep learning for intricate credit evaluations, AI in Treasury is setting the stage for a new era of financial management. Its applications, such as real-time insights in credit scoring and dynamic discounting, are instrumental in providing more accurate and timely financial decisions.
However, as with any technological leap, caution must be exercised. Ensuring data privacy, maintaining transparency, and addressing regulatory compliance are hurdles that need meticulous attention. By navigating these challenges effectively, organizations can truly harness the full potential of AI in Treasury. It’s not just about integrating advanced tools; it’s about understanding their implications, training teams adequately, and iterating solutions for optimal results.
In conclusion, the fusion of AI and treasury practices is not a mere trend; it’s a paradigm shift that’s here to stay. The future of financial management lies in the balanced integration of human expertise and AI-driven solutions. As AI in Treasury continues to evolve, so will the tools and strategies at our disposal. Embracing this change, while remaining vigilant of its challenges, is the key to a prosperous and efficient financial future.
