Joe Rose, president of strategic technology provider JBS Dev, wants to bust one of the myths about using generative and agent AI systems. “It’s a common misconception that the data has to be perfect before running these types of workloads,” he explains.
As noted in a recent article in the AI Fieldbook Overview, vendors and consultants are rightly suggesting the need for huge data lakes and multi-year data transformation programs, respectively. Therefore, management is struggling with this issue. The reality is a little different. “The tools to handle low-quality data have never been better than they are now,” Rose says. “It’s almost amazing what an LLM can pick up on a half-written prompt.”
That makes sense. If you have access to such tools, it’s worth putting appropriate guardrails in place and using them to your advantage. The inherent unpredictability of the model means that bad outputs need to be handled. This is where humans join the loop. Text or categorical data is resilient. “People are…used to, ‘We build it, it works, and we forget about it,’” Rose says. “That’s not how these systems work.”
Regarding incomplete data, Rose cites the example of a client in the medical field whose goal was to move to a different claims adjustment system. The records were mixed. Some were in PDF format, others in image format. Sometimes the procedure is done in the doctor’s name, and sometimes the doctor’s name is in the patient’s name. From OCR to image to PDF text extraction, Gen AI was able to scope clean data from a simple prompt. A more agent-like approach was then utilized, such as comparing customer records with insurance policies to ensure they were being charged at the correct rate.
“You start stacking different use cases on top of each other,” Rose says. “That doesn’t mean everything will work. Humans still need to be involved. But what you want to do is say, ‘Start with 20% automation, then 40%, then 60, 80%,’ and scale that up over time.”
Going forward, Rose expects the discussion around these models will be about cost and portability. “I think we’re going to see these fundamental leaps and shifts away from model capabilities and towards, ‘How can we make the costs more sustainable so that we don’t have to build data centers at the rate that we’re building data centers?'” he says.
“The last mile is, ‘How can I run these functions on my laptop or my phone instead of running them in a data center?'” The model was trained on a set of data: all the pages on the Internet and whatnot. It’s not like there’s a lot more data that hasn’t been thrown in that could lead to some kind of breakthrough. ”
Rose is looking forward to the conversation at the AI & Big Data Expo, where JBS Dev is participating. And another controversial opinion he makes is telling people to stop buying from SaaS vendors when they can buy it themselves. “It’s not as difficult as you think,” he says. “Almost everyone uses the cloud in some way, and that’s where I start, especially since cloud tools from the big three have everything you need to start implementing agent workloads tomorrow without the need for new software licenses or new training.”
Once that’s in place, JBS Dev takes the next step.
Watch Rose’s full interview below.
Image by Gerd Altmann from Pixabay
