Judging by the hype, the subject of AI feels like the panacea to everything. Got a problem with your accounting? AI. Need images for marketing? AI. Can’t crunch all those millions of data points in real-time? AI!
Indeed, the recent breakthroughs in generative AI, brought about by the transformer architecture that underlies most of the current advances, have been staggering. (Without getting too technical, transformer architecture allows neural networks to process immense amounts of data in parallel, opening the door to much more contextual understanding of complex topics.)
We’ve seen this type of excitement before, and it happens with every significant tech breakthrough. We saw it in the Dot Com boom just before the bust. We saw it in cryptocurrencies, the metaverse, and even in the early days of Facebook—before the data breaches and egregious privacy violations that sobered many of us up regarding how wonderful social media is.
Facebook is still around, and it remains one of the most important platforms for online advertisers, but no one wears rose-colored glasses about it anymore.
Similarly, we’re starting to recognize AI’s limitations. Understanding those limitations is key to leveraging it as much as possible.
Understanding AI’s limitations so you can work around them
We’ve already seen cracks in the glass surrounding AI. Generative AI “hallucinates”—a technical word meaning that it generates output that isn’t related to its input. This hallucination probably won’t ever go away because of the way transformer networks are designed. Getting rid of the hallucination means getting rid of what makes generative AI work in the first place.
Research published on Bloomberg Law found that general-purpose language models hallucinate 75% of the time when it comes to core court rulings. And at least two lawyers were fined for submitting fake legal research in court that was produced by ChatGPT.
Another major concern with generative AI is copyright infringement. A recent study found that ChatGPT produced 44% copyrighted text in response to prompts, making it a legal nightmare for companies that want to use it to automate content creation.
Still, the benefits of AI can’t be understated. Implemented correctly, with programmatic guardrails to significantly reduce issues such as copyright infringement and biased answers, AI can automate tasks in seconds that previously took hours to accomplish.
In an automation, you can also add a human fact-checking step to verify generated output—the overall time taken for the task is still faster than without the AI.
For example, tools now exist to analyze massive datasets inside an Excel spreadsheet. Although humans should always verify AI’s outputs, you can save hours of manual labor by sending a spreadsheet to an AI tool and asking it to establish connections between the data. Once those connections have been established, you can then develop formulas or charts to visualize the connections.
It’s also possible to constrain generative AI answers to a specific dataset, such as internal company policies.
Some AI has existed since the 1980s—and it works excellently
Another thing to know is that AI has existed for many decades. The phrase “artificial intelligence” refers to any computer system that carries out tasks that have previously been considered only possible by humans. That might include translating languages, computer vision, or decision-making based on data analysis.
The banking and insurance sectors have used sophisticated AI since the 1980s for credit scoring and determining a loan’s potential risk.
In some ways, mathematical AI might be considered “superior” to generative AI because it operates on more predictable variables that result in definite answers. In software development, we call that a deterministic program: Given the same set of inputs, a deterministic program will always come up with the same answer.
Generative AI isn’t deterministic. When you give it a prompt, it answers differently each time.
Deterministic models have their place in automations, as do non-deterministic models. An automation that requires 100% mathematical precision every time might be harmed by using generative AI. However, a solution that requires immense flexibility and scalability might be better suited to using generative AI.
Both forms of AI have their place, and both offer increased efficiencies. However, using the wrong one for a given task can lead to complexity and errors.
A software partner can help you implement AI effectively and gain the advantage
How AI fits into an automation workflow you have in place isn’t always easily answered. So many tech options exist that a comprehensive overview of technologies, platforms, and systems is required to figure out the most efficient automation workflow for your business processes.
The best way to gain immediate access to that comprehensive knowledge is to partner up with a software development company.
When dealing with AI, implementing safeguards against faulty output is crucial. Many of those safeguards don’t exist yet because the technology is so new. However, waiting for the guardrails to appear before implementing AI in your automations means you’ll likely miss out on first-mover advantage over your competitors.
For example, the lack of a solution to generative AI’s hallucination problem hasn’t stopped the likes of JPMorgan from forging ahead and using ChatGPT for investment advisory automation. They’re doing this by developing their own tool, not by waiting for external tools to make the task easier.
Partnering up with an experienced software development company such as JWilliams & Associates is your best way forward to:
- Discover the potentials of AI in all possible automation workflows at your business.
- Implement sophisticated AI workflows for which tools don’t yet exist—JWilliams & Associates can develop these tools for you.
To learn more about how we can help you implement an AI automation solution in your business, sales@jwausa.com
Jeffery Williams – CEO – JWilliams & Associates