What isn't AI, anyway?

December 12, 2023 5:00 PM

Today there is a veritable universe worth of products that claim to use artificial intelligence or be powered by AI or are, themselves, an AI solution. But how true is that? Because while "What is AI?" is an important question (see previous story), another one just as vital is, "What isn't AI?" It is difficult to understand what counts as artificial intelligence without also having an idea of what doesn't count. Without this boundary, there's precious little people won't call AI, which only serves to confuse the issue.

When we asked our panel of AI experts, though the particulars did vary, the strongest points of commonality were that, for something to count as artificial intelligence, it needs to be dynamic and adaptive, which excludes most deterministic rules-based systems like robotic process automation. With few exceptions, most of the experts we spoke to said that automation alone does not make AI — it needs to be able to approach novel problems and learn from them.

Yet even within this framework there remain differences in what they chose to highlight. Some, like Aaron Harris, CTO of Sage, believed that AI cannot be deterministic, saying that AI — like human beings — operates on a more probabilistic basis, which serves as both an advantage and a disadvantage. Others, like Samantha Bowling, managing partner of GWCPA, pointed to the need for human intervention as a marker of something which decidedly is not AI. And some, like Wesley Hartman, founder of Automata, are skeptical that even the things we call AI are AI at all.

What the following slideshow demonstrates is that while there is a general sense for what shouldn't count as AI, the specifics can vary as people focus on different things. While AI will no doubt have a major impact on the accounting profession, just how this will happen, and with what kinds of AI, remains to be seen. (See what our experts think AI is here.)

Ellen Choi

COO, Aiwyn

Rules-based algorithms and procedural automation and predefined "if x then y" automation technology do not pass the bar for being AI to me. While useful automation that could feel like AI, those systems follow hardcoded constraints. A core requirement for AI is a level of reasoning capacity where the system continuously learns relationships between inputs and outcomes, takes in context, and generates answers, that changes over time — improving over an initial training regimen rather than staying static like automation.

An example of automation is expenses categorization automation that follows robust, set rules without learning. Accountants can get very far with improving their productivity with automation. An AI application would be generating a predictive forecast of expenses by analyzing and learning from financial data, including expenses, and increasing its accuracy over time.

Automation and AI are symbiotic, and using both can yield greater productivity than one on its own.

Hitendra Patil

President, global F&A outsourcing services, Datamatics Business Solutions

Smart automation (rules-based), chatbots that follow programmed responses through data-matching algorithms, benchmarking against market data, etc., are not AI systems. Any system with no capability to "learn" from data unless a human does software programming to create software capable of "connecting the dots" learning isn't AI.

Avani Desai

CEO, Schellman

AI should not be applied to everything that involves a computerized process — it's important to distinguish between true AI and automation or rule-based systems. AI, at its core, involves machines that can learn, reason, and adapt autonomously, thereby mirroring human intelligence, while these other technologies, even though they're still helpful, lack the capacity for learning and decision-making that defines AI.

Adam Orentlicher

Senior vice president and chief technology officer, Wolters Kluwer, Tax & Accounting North America.

AI encompasses a wide range of technologies, including machine learning, which is a subset of AI. ML systems learn from training data, either supervised, unsupervised, or semi-supervised. Rule-based systems, which operate based on predefined rules, are sometimes classified as AI but are not considered ML. In contrast, systems like ChatGPT, which are large language models trained on public datasets, continually learn and improve as new data is inputted. This is an example of ML.

Pascal Finette

Co-founder and CEO, Be Radical

Simple rule-based systems that lack the ability to learn or make decisions beyond their explicit programming are often incorrectly labeled as AI. True AI should possess adaptive learning capabilities.

Chris Griffin

Managing partner, transformation and technology, Deloitte & Touche

Automation and AI are not interchangeable.

Automation is not a new concept. Automation is prevalent and involves machines or software performing repetitive tasks by moving data from one place to another, typically at scale. Automation is typically rule-based and follows predetermined instructions.

The difference between automation and AI is that AI interprets data, recognizes patterns, and generates outputs — in certain cases mimicking human intelligence.

Wes Bricker

Vice chair, PwC

There are certain technologies that are mistakenly referred to as artificial intelligence — an example would be machine learning models. ML is a subset of AI, but not all ML should be considered AI because ML algorithms are used for specific tasks without exhibiting broader intelligence.

Paul Goodhew, Richard Jackson

Global assurance innovation & emerging technology leader,
Global AI assurance leader
Ernst & Young

Richard Jackson, EY Global Artificial Intelligence Assurance Leader

The differences between AI and automation are important to consider. While they are closely connected, they are not the same. AI is an IT approach that can help with aspects such as harnessing and learning from human behavior to help generate intelligence, including outputs or activities. For example, generative AI might help craft the first version of a memo that a human then reviews. While AI can assist with automating activities to improve efficiency with certain tasks or activities, it should not be considered purely as an automation technique.

Steve Chase

Vice chair of AI and digital innovation, KPMG

AI should demonstrate a level of both learning and adaptation. As such, computerized decision-making, process automation, and statistics are not always considered artificial intelligence. While AI is the most revolutionary field of tech today, it isn't a panacea. It won't take over the world or replace humans; you still need them. While it is a catalyst for progress, it also has its own limitations and risks that must be responsibly managed to ensure safety and engender trust.

Blake Oliver

CEO, Earmark

Don't confuse AI with basic automation or rule-based systems that lack adaptability and learning capabilities. Data entry tools or macros, although automated, do not qualify as AI, as they don't involve complex data analysis or learning from patterns.

Abigail Zhang

Professor, University of Texas, San Antonio

The following AREN'T AI:

Enterprise resource planning systems;
Rule-based data analytics (e.g., identify high-risk transactions by looking at those that happen at irregular times);
Data visualization;
Rule-based automation (e.g., those that follow if-then rules);
Blockchain;
Rule-based text mining techniques (e.g., bag-of-words approach); and,
Data analytics based on statistical methods (e.g., Benfords' law, summary statistics, t-tests, etc.)

Jason Staats

Founder, Realize

So much can fall under the umbrella of artificial intelligence that it's tough to gatekeep, and ultimately doesn't matter. The most commonly fumbled vernacular is the notion that, "We've trained our own model," when it's simply prompt engineering or fine-tuning around a publicly available model.

Enrico Palmerino

CEO, Botkeeper

Automation is not artificial intelligence unless it is done with actual artificial intelligence.

Shane Westra

Chief product officer, Canopy

What isn't AI:

Simple automation and scripting: Tasks that follow a set, unchanging sequence of actions without learning or adapting. These are basic automations, like scripted responses or routine processes, that don't involve any learning or decision-making.
Standard software programs: Traditional software that operates under strict rules and algorithms without learning or evolving from their experiences. They perform specific tasks as programmed but do not exhibit any form of intelligence or cognitive abilities.
Statistical tools without learning capability: Statistical models or tools that analyze data and provide outputs based on predefined algorithms, but do not learn from new data or adapt their functioning over time.

Danielle Cheek

Vice president of strategy & industry relations, MindBridge AI

Most people don't realize that when making a decision, they are doing two things almost simultaneously: predicting various potential outcomes and making a judgment call. The expanding role of AI in supporting humans requires us to be more intentional about separating prediction from judgment. AI excels at identifying patterns and enhancing prediction capabilities. However, judgment must involve a "human-in-the-loop" approach. This technical term refers to integrating human oversight and intervention into AI systems to maintain control and accountability.

Vsu Subramanian

Senior vice president of engineering and head of AI, Avalara

We sometimes talk about or include automation or decision-making systems in conversation around AI, but these are not true artificial intelligence because they don't exhibit adapting or learning behaviors we associate with AI. Anything that repeats the same task in the same manner every single time is not AI. These systems are instructed or programmed to do a task, and they only do that. They are not self-learning systems that can learn, adapt and get better.

Jeremy Sulzmann

Vice president, Intuit QuickBooks Partners Segment

Automation is often confused with AI because of the data elements upon which they both rely. AI is adaptive and can learn from data inputs, which can evolve automation and improve its accuracy. However, the function of automation in and of itself cannot learn and evolve on its own from data inputs.

Kacee Johnson

Vice president of strategy and innovation, CPA.com

I think oftentimes OCR (optical character recognition) gets mislabeled as AI. OCR reads documents and images, then converts them into editable and searchable data. AI has the ability to learn, problem-solve and, in some cases, even make decisions.

Aaron Harris

Chief technology officer, Sage

Any technology with deterministic behavior is not artificial intelligence. For example, rules-based systems, pre-defined decision trees, scripted automation, and forecasting that relies on traditional statistical methods like regression analysis would not be considered artificial intelligence. This determinism is the fundamental constraint to traditional technology — it means software must be coded to foresee every possible input, which in many cases is impossible. AI unlocks a broad set of use cases that have traditionally required humans. The trade-off is that AI, like humans, is not 100% accurate and in some cases is less accurate than a human.

Samantha Bowling

Managing partner, GWCPA

Anything that requires human intervention to make a decision or operate — rules-based coding of transactions in bookkeeping software is not AI.

Wesley Hartman

Founder, Automata Practice Development

I would argue that ChatGPT is not an AI even though it feels like it is the closest to achieving that status. I like to call ChatGPT the most advanced autocomplete in the world. It has been trained on the internet of data to take a prompt and use that to predict what the next word should be in a response. These functions are already permeating our daily life. In writing the answers to these questions Microsoft Word is, at times, recommending the next word or auto correcting my misspellings. This is not AI, but is just Word making its best guess and what word comes next in my writing. ChatGPT is just the best at doing it so far.

Jin Chang

CEO, Fieldguide

The most common misconception about AI is that it is synonymous with automation. In fact, the only shared elements between AI and automation are their reliance on data and their goal of streamlining convenience. One example to differentiate the two is "rules-based automation" versus modern AI. The first is driven by explicit business rules, whereas the second is driven by thousands (and in some cases millions) of real-world examples.

To liken AI to automation beyond those commonalities is to ignore the greater power of an AI system. While automated systems must be manually configured to execute monotonous, repetitive tasks, AI systems are often independently adaptive once they have data to process, meaning that they learn as they go without continuous monitoring. AI leverages aspects of automation, but it goes beyond simple execution of tasks by learning to make decisions on its own, mimicking human behavior.