Whether you term it AI-led, AI-powered, AI-infused, or simply AI automation, it’s an undeniably pervasive technology. In the last 24 hours of writing this article, at least six new AI-led automation products were launched across various industries. But what is AI-led automation?
In a previous article, we defined AI as ‘the property of computational systems that can learn’. Marrying AI and automation, theoretically, creates self-learning systems that complete tasks with minimal human intervention.
Automation refers to using computers to complete business processes otherwise done by human beings. It should be self-sufficient, low-maintenance and unpredictable (in contrast, Frankenstein’s monster is frequently described as philosophical and ‘animated’ with ‘fiendish rage’).
Memes like the one above often capture the public's fear of automation. After all, the relationship between automation and human labour is complex. Automation is designed to alleviate human efforts so employees can turn towards more value-adding, fulfilling activities. Yet, automation is also responsible for lessening the demand for traditional roles (such as travel agents, bank tellers and data entry clerks). The pull of automation’s monotony-mitigating effect is countered by the push of fear of becoming obsolete (FOBO).
Automated systems generally cost less to maintain than human salaries and can provide a more consistent level of service. One of the main disadvantages of non-AI-led automation is its rule-based framework. Traditional automation often breaks when confronted with unforeseen circumstances. Developers usually can’t program an automation system to respond to every single potential scenario. But, with AI, that’s no longer a concern.
Firstly, AI algorithms can understand the relationship between unforeseen information, much like a human. For example, AI can now immediately recognise that ‘nuclear fuel’ is a tangible asset on a balance sheet, even if it’s never been supplied with the balance sheet of a nuclear company.
Then, whenever AI misinterprets information, it learns from its mistakes. Through continual learning, AI can quickly onboard new information without requiring new training. AI systems should be fully flexible; their continuous learning should not compromise future learning.
AI’s continuous learning ability is invaluable for automation products requiring adaptability. For example, financial automation systems may need to be modified for new regulations. Yet, not all AI can continually learn. For example, the latest version of ChatGPT, GPT-4o, has a knowledge cutoff of October 2023. Consequently, GPT-4o cannot generate responses based on recent events—a hindrance for users looking for timely information.
But, maintaining continual learning is only one potential limitation of AI-led automation.
AI can’t just be added to any system or product for immediate success. Instead, developing or onboarding AI-led automation requires careful consideration to achieve the desired performance.
Training AI algorithms can take different forms. For example, training our data extraction product (Transcribe) is as simple as pointing and clicking. The convenience of training Transcribe results from our algorithms' ability to learn from mistakes rapidly.
Training an AI-powered automation model as a developer requires a high volume of diverse data (even up to millions of pages), which is not always accessible. Training can also be time-consuming, taking weeks (though, with a specialist vendor like Evolution AI, training can be condensed to approximately two days).
Likewise, the speed and accuracy of the model will depend on its training. Historically, older versions of AI-led automation faced a trade-off between speed and accuracy. For example, lightning-quick automation meant letting a few mistakes slip through the net. As AI-led continues to advance, the speed-accuracy debate is becoming less relevant.
Bill Gates once said, ‘Automation applied to an inefficient system will magnify the inefficiency’. AI-led automation must integrate seamlessly into your business operations, enhancing productivity rather than causing bottlenecks. For example, an automated loan underwriting system that doesn’t automatically contact customers post-underwriting is functionally incomplete. Therefore, initially mapping out the required outputs for the automated system will prevent scope creep (i.e. the time-consuming addition of extra functionalities to the system).
The benefits of correctly implemented AI-led automation are predictable – a high return on investment (ROI), reduced labour costs and higher data accuracy.
To effectively implement generative AI-powered automation, start by assessing which business processes in your organisation are automatable. For example, Evolution AI offers automated, AI-powered data extraction from financial documents, a tedious back-office task that many firms in financial services still use employees to complete. We’ve worked with global leaders like NatWest, Hitachi Capital and Dun & Bradstreet to automate data extraction from documents. To find out more, please email hello@evolution.ai or book a demo.
*These five AI-led automation products are Jitterbit’s operational platform, Zenerate’s design software, Cymbio’s image AI, Kore AI’s contact centre AI, Aspire’s analytics solution and Schneider’s energy management centre AI.
