I often reference in my episodes how I used to spend four hours transcribing my interviews. Hours of my day were spent sitting at my desk, listening to a few seconds, pausing, typing, and playing the next few.
It was slow and repetitive until AI helped me produce transcripts within seconds. Initially, they were unreliable, but over the years they have become a formative part of my workflow.
Consider: I can record more conversations, publish more work, and meet more people. I no longer get anxiety when I see the numbers on the timer get higher and higher, aware of the work that lies ahead of me.
By almost every professional measure, I’m better off. But I’m aware of one consequence I didn’t anticipate: I can still recall details from interviews I transcribed years ago, yet some recent conversations are already fuzzy.
Yes, I gained time, but sometimes I feel I lost something else.
That concept came up during my conversation with Sagi Eliyahu, CEO of KMS Lighthouse, an Israeli company operating in knowledge management.
Knowing the right answer
Before chatbots could compose paragraphs or agents could act on our behalf, large organizations had a more basic problem. Bank employees, insurance representatives, or hospital staff needed to understand customers, policies, or equipment. Delays or mistakes could have consequences for patients waiting for treatment or those looking for solutions to problems.
KMS Lighthouse helps companies organize, approve, update, and retrieve that operational information. Eliyahu described repositories with hundreds of thousands of documents, some current, some superseded, and some contradicting one another. Its software distinguishes curated, approved knowledge from less-structured material, tracks versions, and controls who can access what.
Today, the same information is being fed to AI systems and LLMs. “They understand that AI will be only as good as the knowledge underneath it,” Eliyahu said.
This is where the promise of AI becomes awkward. We are building systems that can answer questions convincingly, and increasingly take action, without necessarily knowing whether the underlying instructions are correct. A confident answer based on an obsolete insurance policy is still wrong. An agent that acts on bad data could be fatal.
LLMs are algorithms that rely on data. And that data is only as good as the information it relies on. Computers may be able to compile plenty of information, but knowledge and wisdom are still human traits.
Eliyahu’s proposed remedy is governance: approved sources, version control, appropriate permissions, and a human who decides when two documents genuinely conflict. If the system lacks a reliable answer, he wants it to say so. As he put it: “A human needs to take a call. It’s not AI.”
“You manage him like an employee,” he added. “If not, it’s going to be a mess.”
The experience we can’t automate
In my case, transcription wasn’t just a chore. It was an unplanned second listen, a way to absorb someone’s argument and remember the conversation. Automation removed the friction, but perhaps also part of the learning. Sometimes I feel it affected how much knowledge I’m left with once the interview is published.
The same tension appears in software development. Experienced engineers are better placed to catch AI-generated mistakes than juniors because, of course, they are. They’ve spent years learning what good work looks like. “A junior will not notice,” he admitted.
So if AI handles beginner tasks, how do new beginners become experts?
This isn’t an argument for making journalists type every word, or programmers write every line. But we should at least differentiate between work that consumes time versus work that builds on our expertise.
The consequences become more profound when we think about our children. I shared my belief that education may need to put greater weight on traits such as curiosity, communication, and judgment, rather than on learning technical skills. He agreed that learning to use AI won’t be enough; people will need to challenge it.
“Your only chance to survive in this new world is by using it,” he said.
To challenge a machine, you need some basis for knowing when it’s wrong. And building that basis often requires doing things badly, slowly, and repeatedly before you can do them well. Mistakes are ok, and should be a part of the learning process. But so long as we keep automating those tasks early in our careers, we risk losing what gives experts their initial foothold into industry.
We discuss this and more in the full episode. For now, below is a preview.
[Preview: The AI Paradox: More efficiency, but less experience]
