The Most Important AI Future Job Is Nearly 3,000 Years Old

Every AI system is only as smart as the data it consumes. That’s common knowledge. Yet companies are racing to hire prompt engineers and fine-tune transformers while systematically ignoring the most critical weak point in their AI infrastructure: the quality of information feeding their models. The result? AI systems that scale misinformation as efficiently as they scale insights.

“AI will accelerate and scale misinformation as quickly as it scales good information,” warns Dougal Martin, head of knowledge at Deel, who oversees a 30,000-article knowledge base spanning global employment law across 160 countries. “It doesn’t know that this piece of information that it’s sharing with you is wrong or is misaligned.”

This isn’t just a theoretical problem. It’s playing out in real-time across enterprise AI deployments, where the cost of bad information compounds exponentially through automated decision-making systems. The solution, according to Martin, isn't snazzy algorithms but better librarians, a profession that was created in the 8th century B.C.

The human layer that can’t be automated

The concept of an “AI librarian” might sound like a corporate buzzword, but Martin’s definition cuts to the core of modern AI infrastructure challenges. “An AI librarian is basically someone who is accountable and responsible for the integrity of information that we use to support our AI tooling,” he explains. It’s a role that’s part data engineer, part domain expert, part quality assurance specialist, and entirely essential.

At Deel, this translates to managing compliance information that directly impacts client decision-making across global markets. When a sales team needed better timeline information, Martin’s team deployed a solution in two hours. But the real work happened in the testing phase, when the AI started surfacing inconsistent information from disparate parts of the knowledge base. “Very quickly, when you start to test it, you go, ‘Ah, timeline doesn’t look quite right,’” Martin recalls. “So right away, you go, ‘Okay, where is it pulling that information from?’”

This is where human intuition becomes irreplaceable. While AI excels at pattern matching, humans excel at pattern-breaking — that is, spotting the anomalies, inconsistencies, and edge cases that probabilistic models miss. “AI is a probability machine,” Martin notes. “Humans are not probability machines. Humans encompass a range of information, both consciously and subconsciously, in any given situation to make decisions.”

Building an architecture of trust

The technical architecture of trustworthy AI systems, according to Martin’s approach, starts with what he calls the “human base layer”. This is a foundation of human-curated, human-validated information that serves as ground truth for everything built on top. It goes beyond just content moderation and is part of what is known as systematic ontology design (the formal classification of information and its relationships).

“We decided that was where we would invest our human power in ensuring that information is correct, up to date and well aligned, so that everything that we build on top of it is going to be correct,” Martin explains. The team includes technical writers, lawyers, and payroll experts—specialists who understand not just the data, but the contexts in which it will be applied.

But here’s the provocative insight: the AI system itself becomes a user that must be designed for. “I really think of our AI tool as one of the end users,” Martin says. “It has its needs, it has its quirks, and it has its way of doing things.” This dual-audience problem, which requires designing information systems that serve both human and artificial intelligence, represents a fundamental shift in how we think about knowledge management.

The scaling paradox

The open secret of AI adoption is that it doesn’t eliminate the need for human expertise. Rather, it changes what humans need to do and amplifies the importance of getting it right. Martin draws a parallel to diagnostic radiology, where AI tools didn’t replace radiologists but made radiographs cheaper and more accessible, ultimately increasing demand for specialists.

“The moment it became more cost-effective to read a radiograph, people took more pictures, and so the demand for radiologists went up,” he observes. “AI is going to make us invest in the things that we care about more, and we’re going to do more of those things.”

For enterprise AI, this creates a scaling paradox: the more you automate, the more critical human judgment becomes at key decision points. The question isn’t whether to use AI, but where to place the guardrails and who will be accountable when the system makes mistakes.

Martin notes that human accountability will remain a feature of AI systems, especially for highly-regulated industries and even as companies tinker with Agentic AI. He advocates for a more measured approach to these AI systems that can swarm data sources to make autonomous decisions.

“It’s difficult to hold an AI accountable,” Martin says. “It’s much easier to hold a human accountable.”

The buck stops with a human

The AI librarian role is essentially a recognition that the future of AI isn’t about replacing human intelligence but augmenting it strategically. As Martin puts it: “There is no substitute for speaking with your end users, and you cannot outsource it to an AI.”

The question for data leaders is not whether to hire more engineers to fine-tune their models, but whether they are investing in the human intelligence needed to ensure those models have something worthwhile to learn from.

Because in the age of AI, the most advanced algorithms in the world are only as good as the librarian who feeds them.

Image credit: iStockphoto/ViktorCap