Despite it being the top buzzword in technology today, there is still a lot of confusion about artificial intelligence (AI). In simple terms, artificial intelligence is an umbrella term for approximately a dozen related technologies all attempting to gain insights from analyzing data or performing tasks based on data that aren’t ideally suited to humans. They include machine learning, deep learning, and robotic process automation.
In the case of cash flow forecasting, these technologies are excellent for doing scenario modeling organizations can act on with high levels of accuracy, or for predicting future trends or results with the greatest degree of probability.
These insights come from two live 2020 Virtual Experience sessions offered by the Association for Financial Professionals. The first was “Getting to Grips with AI and Cash Flow Forecasting,” featuring Marcus Martinsson, the global lead for treasury and cash analytics at Accenture, and David Backhaus, a senior manager in the Treasury Group at Accenture. The second was “Cash Forecasting Solved? Optimizing Working Capital through AI Forecasting and Inventory Principles,” featuring Tommy Cho, director of corporate cash management and sales with Deutsche Bank, David Backhaus, and Kyle Schiller, treasury solutions manager, global banking and liquidity at Accenture.
The message of both sessions is that AI is revolutionizing the process of cash flow forecasting, resulting in greater efficiencies, profitability and productivity.
“When we start talking about cash flow forecasting, and how to optimize the uses of your cash, these prescriptive models are extremely powerful,” stressed Martinsson.
“Our old way of thinking of cash management or cash forecasting was to be very basic, to be honest,” Backhaus said. “We would do a monthly cycle. We would tie a generic kind of cash forecast with some decisions that were very manual, and sort of best guess, and then we would extract some cash. So, we wouldn't take into account a lot of variables. And the accuracy of it was always in question. We often questioned how much value it was actually adding. And it was just a real frustration.”
Compounding the problem: “There does not seem to be one single technology solution or software that can necessarily handle everything and all the nuances that a global company faces,” said Cho. “While there are several third-party software products out there, all of the different companies have their own nuances, which is why so many companies still rely on Excel and manual inputs from their business units.”
In response, Martinsson and Backhaus shared how Accenture has implemented AI models into the firm’s payroll process. Accenture operates in over 100 countries, with countless different currencies and compliance regulations.
“We've implemented AI models into our process,” said Backhaus. “We added a tool that the users can interact with. What we have now is a process that generates a baseline forecast of what we believe is
going to happen in the next 12 months, with a focus of really prompting actions in the next 90 days. That baseline forecast has automatic data feeds that tell our users what's going to be happening and allows them to just change the key things that an AI model might not be able to change. It has saved us tremendous amounts of time. Our users are a lot more engaged with the forecast because it doesn't take half their day, every day. It's really been a huge benefit to us as we started to roll it out.”
“This is the beauty — that you have the data collection happening automatically,” Martinsson explained. “It gets loaded to ERP systems. You have the structured data being stored in a data lake. In turn, it is feeding AI models to generate the baseline forecast. This allows the user to see forecast results within a few minutes.”
According to Cho, the accuracy and availability of data really becomes an issue when trying to create a forecast that can drive business decisions. Success really depends on having a strong data culture. With that in place, the quality of data will improve, data analytics insights will actually be legitimate, and people will start trusting the results.
“What seems to be less evident is the effort needed to access all of the information spread across your entire organization globally, how to bring it together, filtering through all the noise and then, making sure that it's actually searchable and accessible so that an accurate analysis can be completed,” Cho said. “RPA, AI, ML — whatever IT acronym you'd like to throw out there — all these trends are coming together to help make solving the cash forecasting issue easier.”
Schiller described how Accenture came up with an automated solution that enabled “us to reduce cash held in country, which is our ultimate goal here, and have our cash that is centrally held in our in-house bank be as high as it can for strategic initiatives, acquisitions, things like that. The big impact here is, we were able to increase our accuracy based on what we were previously doing in Excel. The other piece was, we didn't have tons of people spending a lot of time putting forecasts together, it just automatically produces them. So, I think that the key takeaways here are that it saves a lot of time, we were able to increase our accuracy, and it frees up our people to do more value-added work.”
As organizations build their models, there are several key themes they should keep in mind, Schiller said.
“You need to have your clear objective and identify what's your pain point,” Schiller explained. Questions to answer include: “What are you trying to solve for? What type of forecast horizon are you looking for? Do you want to just see the next month out, or three months, six months, 12 months, or even further out? What's your accuracy target? What level of accuracy can you be comfortable with?”
Whether it's receipts, collections, disbursements, payroll, etc., Schiller said Accenture had different modeling algorithms it could use with its new system.
“I think we wound up going through 12 of them,” said Schiller. “We tried to determine which one would work best in a given country or for all countries. If we were trying to get the most accurate forecast at the cash type level, we developed what we call this hybrid model. It will actually take a look at each cash type and run it against all 12 algorithms, and it'll tell us which the best algorithm is to use for that cash type in that country. It'll store that and then utilize that set of algorithms for that country and cash type combination, until we actually retrain it.
“We've set up times to retrain these models. Right now, we're doing it on a monthly basis, so it'll go back and automatically look and see if anything should have changed, if any models are working better now.”
In conclusion, Shiller noted: “It's really nice to have the ability to just drill in and see what's going to be driving our cash balance in two weeks, four weeks, six weeks, or 12 weeks. We're able to produce forecasts quicker. We're able to customize the timelines we want to look at, which I think is key. Originally, when we started on this journey, we were looking at a longer-term time horizon. Then we realized that, hey, that's good for directional, but we're not necessarily acting on anything 12 months out, right? So, let's really keep a hard focus on the first three months and look at that accuracy, and that's what we were able to achieve through the AI models that we use.”
For more information around this topic, download the AFP guide, Emerging Technologies Part 2: Artificial Intelligence and Machine Learning, underwritten by Kyriba.