AKI · ENTERPRISE KNOWLEDGE

The missing talent is the company’s domain knowledge

To bring AI into production, technical skills are needed, but to make it useful, I believe company’s domain knowledge must become part of the organization’s heritage.

Part One · The talent required to deploy AI into the company

Anthropic is investing one hundred million dollars to train ten thousand engineers to deploy models into production. Meanwhile, McKinsey reports that 89% of companies adopt AI, but only 6% derive significant value from it.

Though these are two different pieces of news, I tried reading them side by side, and I believe they converge toward the same skill to develop: the value of enterprise AI grows when the knowledge that makes it useful within a specific organization becomes explicit, transmissible, and preserved. Simply put, when know-how is collected in a shared, understandable form and kept up to date over time.

Per realizzare questa trasformazione non basta adottare un modello di AI; abbiamo bisogno di professionalità distinte: chi conosce il modello e il suo impiego; chi possiede il significato di ciò che l’impresa decide ogni giorno.

What evidence is needed to make a decision? Who can approve an exception? In which cases is a specialist’s judgement required?

It is in questions like these that the company’s knowledge takes shape, and we need to ensure that it becomes a shared asset of the organization, rather than remaining confined to the memory of a few individuals.

Part Two · The signal to recognise

“When we train people in AI, are we training them to use a model, or also to make what our company knows reusable?”

On October 2nd, Anthropic announced the Claude Frontier Academy: an initial investment of one hundred million dollars to train ten thousand Frontier Deployed Engineer by the end of 2027. According to the announcement, the programme includes an intensive in-person course based on a simulated enterprise deployment, from choosing the use case to security review and handover, followed by twelve weeks in which each engineer leads a real project in their own organization.

Reading this announcement, I think the stated reason deserves particular attention. Anthropic said that talent is one of the most urgent issues in AI implementation, adding that it is difficult to find people able to make AI work inside a real company and that a small group of highly skilled people can produce a disproportionate share of the value. I find even more interesting the way it summarised to CNBC what customers are asking for: people who combine business context with a high level of AI proficiency.

Two skills, one goal.

We should keep in mind that technical competence allows a model to be deployed into production, but the company’s knowledge is essential to give it context: evidence, responsibilities, rules and exceptions. These are two distinct and complementary kinds of work.

Part Three · The second news, read alongside the first one

The day before Anthropic’s announcement, David Pralong, senior partner at McKinsey, published an analysis in Fortune of the firm’s latest research. It revealed that organizations using AI in at least one function increased from 88% to 89% in a year, but the number of companies attributing at least 5% of their operating results to AI remained stable at 6%.

 

Individual productivity is obviously increasing: 15% more cases resolved per 15% more cases resolved per hour in a sample of approximately 5,200 customer service representative, and around 26% more tasks completed by developers equipped with a coding assistant. Despite this progress, Pralong notes that enterprise value does not follow the same trend.

Pralong propodes an organizational explanation, and this caught my attention. Among organizations in the early stages of AI adoption, those that had redesigned workflows were 5.3 times more likely to report enterprise-level value: 32%, compared to 6%.

I also find instructive the example Pralong chose: a bank uses AI to read small business loan documents, saving time in the review process. However, the benefit to the customer remains limited if the process does not clarify what evidence is needed to approve a solid loan, who has the authority to decide, and which exceptions require a specialist’s review.

We can conclude that speeding up reading is one thing, but redesigning the process is another: it means clarifying decisions, responsibilities and next steps.

Pralong encourages us to look at how work is organized, although this does not mean that redesign alone guarantees value in every context.

These two news items, read together, tell us about the transition we are living through: on one hand, there is a shortage of people who know how to bring the model into production, technicians and AI engineers able to implement the best infrastructure to support business processes. On the other hand, they show us that, when the model is adopted, its value also depends on how the company organizes work and makes evidence, authority, and exceptions explicit.

Two different skills are therefore needed.
If the first can be trained (the aforementioned Frontier Deployed Engineers), the second (company-specific knowledge) is not a ready-made package to transfer. It must be brought to light with the contribution of those who know the context and with the support of the necessary technologies.

Part Four · From digital to cognitive

As we have read, an engineer can follow training programmes to learn how to select a use case, design an agent, handle security reviews, and deliver a working system. This is a valuable and, in part, scalable skill.

But knowing “the reason why behind things” offers a completely different reading of the events that happen in a company.

What evidence makes a loan solid in that bank? Who can approve an exception in that plant? What exception distinguishes a strategic client from an important one?

This knowledge is not standard: it lives in the minds of the most experienced people, partly in systems, and partly in unwritten rules. It cannot be transferred simply through a course. It must be collected, made explicit, tested, validated, and maintained.

Those who deploy the model can be trained technically.
Those who keep the company knowledge heritage must be listened to, and what they know must become an asset of the organization: an explicit, verifiable, and adoptable resource for any AI model in use.

It is in this Human + AI convergence that the epochal shift occurs, because it is no longer enough to make data available through digitized processes. To use it in daily decisions, it must be traced back to the specific domain and clarify what it means for the company.
In other words, knowledge has to be put into action.

My reading is that Pralong’s example describes precisely this shift, from digital to cognitive. Evidence, levels of authority and exceptions are part of the meaning of the credit domain. For an AI agent to apply them correctly, they must be represented and retained.

Part Five · Understanding where to start, what remains in the company and how to measure results

“How much of the knowledge that makes AI useful in our company is explicit, verifiable and reusable today, even when people or tools change?”
Answering this question precisely is likely complex, and it was for us at Altea Federation as well. But as Newton said, “If I have seen further, it is by standing on the shoulders of giants.” This reminds us that innovation happens in stages, building on solid foundations. Consequently, we should not think of cognitivizing the entire company in a single project. However, we do need a holistic vision of business processes to identify the most strategic area to start from.

The smartest approach is to start with a well-defined domain, where we observe decisions that recur daily and where responsibilities are already clear and applied.

Some conditions can guide us along the path from digitization to cognitivization.

Who holds the meaning.

For that domain, can we currently write down what evidence is needed, who has the authority to decide, and which exceptions require review? If the answer exists only in the memory of a few people, you have identified an excellent starting point.

Where learning is deposited.

When a trained engineer changes role or the AI model is replaced, what remains with the company? Ontologies, rules, test cases, and the provenance of decisions should remain in a readable and portable format, not only in a tool’s prompts.

What we measure.

I agree with Pralong’s approach, which suggests measuring the time needed to reach a decision, the error rate and the cases the company can handle responsibly, not just the hours saved reading documents. We can record these values before starting and then compare them on homogeneous cases.

Part Six · Training people, preserving knowledge

I believe that ten thousand engineers trained by 2027 is a good news for the market: they will have followed a programme to bring models into organizations’ systems and processes. However, for these skills to translate into value, I also believe it is necessary to make the meaning of what the company knows explicit and create a governed context in which people and AI can work together.

This transformation within a company develops different skills and professions, following parallel tracks.

The first skill is the race toward upskilling people, with education and AI usage paths. The second skill, however, must be built more deeply, starting from the core of company knowledge: it is a capability that grows with those who possess distinctive knowledge, proceeding one domain at a time, even by introducing new professional figures.

New roles are indeed emerging: in a sense, we might call them “the guardians of knowledge”.

The Knowledge Steward brings out and curates domain knowledge, with the support of elicitation agents that help interview experts and structure what emerges in meetings.

The Ontology Engineer formalizes and structures that knowledge into a domain ontology for the company, using open standards, rules, and verification cases.

The Inference Engineer works to make reasoning reliable in operation, with attention to sources, conditions, traceability and responsibilities.

The result is knowledge that does not depend on a single person or a single model in use. It is knowledge made explicit, governable and reusable, even when tools and key people change.

This is the missing talent for companies: transforming their own knowledge into the most enduring strategic asset and making it a heritage of the entire organization, thanks to a cognitive infrastructure like AKI, which makes it usable, protected, and governed over time.

At the next AI training meeting, let’s try adding a second question alongside “How many people know how to use the model?”: “How much of our knowledge is currently in a form that a model—and an auditor—can read?”

Because making that knowledge explicit is an important step towards reducing the gap between the 89% of organizations that use AI in at least one function and the 6% of companies that attribute at least 5% of their operating result to it.

Andrea Ruscica

Chairman & CEO, Altea Federation

Glossary

Domain knowledge

The knowledge related to a specific area of the company: concepts, rules, relationships and exceptions that help interpret data and make decisions within that context.

Ontology

The formal representation of the concepts within a domain, the relationships between them and the rules that govern them. It serves to make explicit how the company organizes the meaning of its knowledge.

Elicitation

The process through which the knowledge of experts is brought to light through interviews, questions, and the examination of cases.

Digitalized domain and cognitive domain

A digitalized domain collects and makes data and information available, while a cognitive domain also represents the meaning of that data, including relationships, rules, authority levels, and exceptions.

FAQ – Frequently Asked Questions

A model is an important component, but on its own, it lacks knowledge of company-specific elements such as criteria, roles & responsibilities, and exceptions. To be genuinely useful in daily operations, it must operate within a structured context that defines what data means and which rules apply.

It refers not only to documents and procedures, but also to the implicit know-how that guides people’s decision-making. For instance: which evidence is relevant, who holds approval authority for an exception, and why an expert chooses one solution over another in specific scenarios.

It means structuring knowledge into a shared format so it doesn’t remain trapped in the heads of a few key individuals. It ensures this knowledge is understandable and reusable for anyone joining the process later, while keeping it updated over time, even as personnel or tools change.

Data simply describes what happened or what was recorded. Understanding its meaning, however, involves knowing how to interpret it: identifying which evidence matters, which business rules apply, and when an exception calls for further review.

Accelerating a task, such as reviewing file documentation, improves a single stage or boosts individual productivity. Redesigning a process, on the other hand, means clarifying fundamental elements: what evidence is required, who holds decision-making authority, how exceptions are handled, and how overall outcomes are measured.

The most effective starting framework isolates a specific domain characterized by repetitive decisions and established, real-world responsibilities. Using this method, organizations can identify necessary evidence, decision-making authorities, and exceptions, ultimately measuring benefits such as reduced time-to-decision, lower error rates, and increased volume of responsibly managed cases.


Why we create AKI:
the infrastructure for Cognitive Organizations.

Bibliographic Note

SOURCES FOR THIS ARTICLE

The details of Anthropic’s programme are those announced by the company. The figures on adoption and productivity are reported by Pralong in Fortune, based on McKinsey research and the studies cited in his commentary; they are not Italian data or AKI results. The banking example is illustrative. The link between the two news stories and enterprise knowledge is the author’s interpretation.

Share this article!