While generative artificial intelligence is the hot conversation topic these days, we must not forget a long and successful history of using nongenerative AI, sometimes called legacy AI, especially for numerical and structured data. Uses such as forecasting of customer demand or revenues or the detection of patterns such as fraud or money laundering are important examples relevant to CFOs and accountants.
These tools and use cases improve in their capability every year and provide tangible business value.
Legacy AI uses
These nongenerative AI systems can also provide significant assistance in meeting compliance and regulatory requirements and preparing analytical reports for those purposes. Matching methods to detect which invoices and payments belong together, especially in cases of partial disparity, are in almost universal usage today and rely on AI.
Many of the more sophisticated management dashboards and systems underlying both accounting and enterprise resource planning software ultimately rely on such AI systems, for example inventory management and planning. Complex processes like just-in-time or just-in-sequence could not function without legacy AI backbones.
Limitations of generative AI
Turning to the oft-hyped topic of generative AI, we acknowledge that many claims are hype. Any tool, for instance, has an intended scope of use for which it is helpful and provides value. Beyond that scope, it is not helpful and may cause harm. Large language models are intended to manipulate language, not numbers, and so are generally not successful at dealing with numbers where we expect absolute accuracy.
A case in point is the analysis of a company's annual report. If we do so using LLMs, we will get answers that are "enhanced" by information extraneous to the report, or we might get numbers that are not grounded in the report. Such uses are not appropriate and misleading. So what can we use them for?
Multimodal uses of generative AI
A step change forward of generative AI is its multimodal facility — the ability to work with text and images at once. Imagine taking a mobile phone snapshot of your latest restaurant bill and it's automatically filed in the travel expense form of your company. What a time and hassle saver! This is quite accurate and thus also prevents human error. The same holds for invoices, receipts and other paper forms.
In case a legacy AI model discovers some sort of mistake — such as fraud or a partially paid invoice — it is generative AI that can convert this discovery into a human-readable message that explains what is going on and what to do about it. We have talked about explainable AI for many years, and it is LLMs that can produce an explanation even if the content of that explanation may need other systems to weigh in.
Natural language dashboards
We have all been in board meetings where one person asks an analytical question to which no one has the right numbers. Oh horror. An analyst will have to be kept busy for a few days, the charts sent, and the result is not actionable for a protracted time. Gone are the days! Generative AI can translate a question from English into the language of databases, SQL, and obtain the table of numbers that results. This table is then translated into the codified language of dashboards and displayed as a graphical image to the human user.
All of this occurs in the blink of an eye. Most importantly, the result is not hallucinated by the LLM but comes directly from the database — the answer can be trusted. This allows further questions to be asked live in the board meeting, eventually getting to an actionable result in a short time. I was present at such a meeting where a sequence of eight pointed questions was asked and answered in less than 10 minutes, leading to novel insights and a board decision. It was an eye-opener.
Support services
Fielding questions by employees, customers and suppliers is a major strain on any accounting division. Generative AI can help by triaging the most common questions and providing correct and sensible answers automatically. From providing help with the dreaded expense reports to filing invoices, AI can largely automate the everyday process of accounting, including matching it to the right expense account and getting approvals.
Security is important, especially when money is involved. Generative AI supplies a new level of sophistication for the detection of a variety of attacks such as phishing and hacking.
Some uses where AI, generative or not, can help in the realm of accounting have been listed here. Beyond the management of a company's finances, the CFO also has to make many decisions for the rest of the company. AI can help analyze scenarios, help find reference data, and contextualize the situations and offerings of competitors or other vendors. It can help to objectify and compare the benefits of multiple options so that the CFO can better decide which to choose.
In conclusion, generative AI delivers genuine business value to the CFO organization after all the hype has been subtracted. The most impressive is the generation of dashboards on the basis of human-language questions. If you do nothing else, have a good look at that.
