AI and the future of work: the fear and the paradox 

Earlier this year, Pew Research Center surveyed adults across 37 countries. In 34 of them, more people expect AI to reduce jobs than to create them, and the concern runs higher in wealthier countries. As a high-income country, Italy fits the pattern: concern about AI is widespread, and large shares of people worry about its effects on jobs and inequality. 

What is really surprising? Adoption keeps growing. According to Eurostat, 19.95% of EU enterprises with at least 10 persons employed used at least one AI technology in 2025, up 6.47 percentage points compared with 2024. Among large enterprises, the share reached 55.03%. 

However, not every enterprise that considered AI goes ahead with it. Among those that stopped short, 44.37% report difficulties with the availability or quality of the data they need. 

That is the paradox. AI is feared because it could do without people. Yet it cannot do without the data it works with. And not all of the missing data sits in databases; some of it was never written down: it is what experts know about their work. 

Talking to an agent is easy. Knowing what to do when its answer falls short is not. The fear is that AI will know enough to replace people. Often, the problem is the opposite: it does not know enough. 

What is tribal knowledge? 

Tribal knowledge is the know-how behind a process, product or service that lives only in the experience of the people who do the work. One person may carry it, or a small team. But it always has one constant: no one has written it down. As long as it stays in people’s heads, it is only as useful as the people who hold it and the colleagues they choose to tell. 

It is the rule everyone “just knows” but no one has written down. The exception handled a hundred times over a twenty-year career. The procedure that only works when one specific person does it. 

At Syllotips, we see this asset as tacit knowledge: knowledge an organization already has, but cannot yet reuse systematically. Part of the data enterprises say they lack is already here. 

Tribal knowledge 
Undocumented operational knowledge held by one or more people within an organization.  

On this topic, iSixSigma notes that companies may meet resistance when they try to document this knowledge: some employees see it as protection against layoffs, or as leverage in salary negotiations. 

For the company, the risk is twofold. When someone leaves, the knowledge often leaves with them, as iSixSigma points out. When someone stays but keeps what they know to themselves, the team depends on them. 

Two forces push in the same direction. The first is structural: knowledge disperses because no system captures it at the moment it is used. The second makes it worse: as the fear of being replaced grows, holding on to it looks like the safer choice. 

How can experts share what they know without losing value? 

Syllotips started from that tension. Asking people to hand over what they know, with nothing in return, does not work. So Syllotips does not extract knowledge from people. It puts experts in charge: agents bring them questions, experts decide what is correct, and every correction stays traceable to them. 

It runs as a four-phase loop: Detect → Route → Review → Write Back. 
Every agent response is evaluated, and when a gap or an unreliable answer emerges, it is routed to the expert who knows the subject. The expert corrects it once. The correction is written into governed memory, and every later conversation can draw on it. 

An illustrative example, based on a hypothetical lender policy. A customer on a fixed-term contract expiring in 18 months asks whether they can take out a five-year payroll-deduction loan. The agent says yes: the general maximum term is 120 months. The answer is flagged as unreliable and goes to an expert, who clarifies that under this lender’s policy, a loan for a customer on a fixed-term contract cannot run past the contract’s end date. 

That rule was not in any of the agent’s documents. The validated rule is now available for the agent to retrieve in later conversations, without asking the expert about the same question again. 

This is how the loop answers both sides of the data problem. Routing questions to experts makes undocumented knowledge available. Their validation turns it into knowledge the agent can rely on.  

From answering to teaching: the expert as manager of AI agents 

This way, what the expert gains is a different role. The questions to ask about an AI agent are the ones every manager already asks about a team. Only the scale changes. 

Evaluate. With people: objectives, KPIs, performance reviews. With agents, every response can be evaluated: is it correct, is it grounded in the right sources, does it match the company’s tone of voice? 

Correct. With people, a manager often learns about a mistake from a complaint. With agents, flagged cases can be routed automatically to the right expert for review. 

Train. With people: shadowing, courses, refreshers. With agents, a correction can become validated knowledge, reused immediately, without interrupting the service. 

Today, experts answer the same questions every day, one at a time, and what they know stays with them. When agents learn from them, experts teach once, decide what is correct, and spend their time on the cases that matter. The workload shifts: repetitive questions go to the agents, and new ones stay with those who can answer them. 

This can reduce the incentive to hold knowledge back. If sharing it means losing an advantage, people have a reason to keep it to themselves. If it means becoming the person who teaches the agents and vouches for their reliability, knowledge stops being an individual insurance policy and becomes a role. That is the principle behind Expert-in-the-Loop: the more an organization relies on agents, the more it needs the experts who can correct them. 

Experts move from repeatedly answering to teaching.  

“Real-world tasks require domain expertise.  
We found a way to give it to AI agents: putting experts in the loop.” 
Simone Silvestri, CPO and Co-Founder, Syllotips 

A real-world case: Leonardo Assicurazioni 

Leonardo Assicurazioni introduced knowledge agents to support its network of advisors. Questions left unanswered were routed to the company’s experts, and their validated answers became reusable knowledge for every later interaction. 

Today, 72% of conversations reuse knowledge validated by Leonardo’s experts. The system does more than search documents better: it scales the experts’ judgment. 

Work is changing. That does not mean it has to change without the people who know it best. AI agents need managers who can teach them the trade and provide the knowledge they need to do the job well. In every company, those people are already there. 

Your experts know what your AI agents are missing.  
Find out how they can teach it. 

→ Write to us at info@syllotips.com 
 

FAQ 

Will AI replace experts? 
AI agents can handle recurring questions. They cannot produce the undocumented knowledge their reliability depends on. That knowledge belongs to experts, whose role shifts from answering questions to teaching agents. 

What is tribal knowledge in a company? 
Operational knowledge that exists only in people’s heads: exceptions, unwritten procedures, interpretations built through experience. 

Can tribal knowledge be documented? 
Yes. Unlike artisan skills, tribal knowledge can be codified, as iSixSigma notes. The best moment to capture it is when it surfaces: when an AI agent meets a question it cannot answer. 

Why isn’t RAG enough to manage tribal knowledge? 
RAG retrieves information from existing documents and data. This is one of the core limitations of RAG: tribal knowledge is undocumented by definition. Capturing it takes a mechanism that carries it from experts to the knowledge base. 

What does an expert who manages AI agents do? 
They evaluate answer quality, correct the uncertain cases routed to them, and turn each correction into validated knowledge the agent reuses immediately.