ENTREPRENEURIA | IN-DEPTH ARTICLE
Ford, the Return of Tacit Knowledge and the Silent Transformation of the Enterprise
By Pascale Caron | July 12, 2026 | EntrepreneurIA
By recruiting several hundred veteran engineers to strengthen its design reviews and improve its artificial intelligence tools, Ford has not turned its back on automation. The automaker has revealed a reality that many companies are discovering in turn: the value of an AI agent depends on the quality of the organization, knowledge, and responsibilities that surround it.
Ford Makes a Countercurrent Choice
While technology companies announce new agents capable of performing an increasing share of human work every week, Ford has made a less spectacular but perhaps more instructive decision. Over the past three years, the automaker has recruited or reintegrated approximately 300 to 350 experienced engineers, nicknamed the “gray beards.” Some had already worked for Ford. Others came from its suppliers. Their mission was to strengthen design reviews, train younger teams, and pass on experience accumulated over several generations of vehicles. [1][2][3]
The media narrative has sometimes summarized this decision in an effective formula: Ford would have rehired its seniors to save a failing artificial intelligence. The reality is more nuanced. Available sources do not demonstrate that these engineers had all been laid off in favor of AI, nor that Ford would have abandoned its automated systems. Rather, they show a company that has understood that technology alone could not recognize all the consequences of a design choice. Nor could it anticipate a rarely documented defect or reconstruct the lessons from problems that occurred twenty years ago.
This reintegration of human experience is part of a broader transformation. Ford has brought together previously siloed functions, involved its suppliers earlier, automated more software testing, and deployed AI-assisted vision systems in its factories. In June 2026, the brand ranked first among mainstream manufacturers in the U.S. J.D. Power Initial Quality Study, with 152 problems reported per 100 vehicles, compared to 193 the previous year. [3][4][5] It would be imprudent to attribute this progress to veteran engineers alone. The indicator covers the first ninety days of use and does not measure long-term reliability. Nevertheless, the Ford case reveals an essential shift: performance did not come from replacing humans with machines, but from recomposing the work system around a common objective.
Automation Meets Corporate Memory
The Ford story is of interest far beyond the automotive industry. It shows what happens when automation meets the real memory of an organization. For years, companies have invested in standardization, procedure documentation, and data centralization. They have sought to transform experience into explicit rules, in order to make work reproducible and less dependent on a few individuals. This approach has improved quality and deployed operations at scale. It has also maintained the idea that all useful knowledge could eventually be recorded in a manual, document base, or software.
AI agents make this assumption much more fragile. They can retrieve a procedure, connect documents, query multiple systems, and trigger a sequence of actions. They become less reliable when the decision depends on a rarely recorded exception, a compromise between several constraints, or a weak signal that only an experienced professional knows how to recognize. The problem does not necessarily come from the model. It comes from the fact that the organization does not always know what it knows.
This distinction is decisive. A company can possess millions of documents without having usable memory. The reasons for a decision have sometimes disappeared. Discarded alternatives have not been preserved. Incidents have been corrected without their lessons being transmitted. Part of the intelligence remains scattered in emails, presentations, informal conversations, and the experience of collaborators who will leave one day. The AI agent does not create this weakness. It makes it visible.
The Return of Tacit Knowledge
This situation brings an old notion to the forefront: tacit knowledge. In The Tacit Dimension, published in 1966, Michael Polanyi summarized this reality with a formula that has become famous: “we know more than we can tell.” A skill is not reduced to a set of instructions. It is built through observation, error, repetition, comparison of situations, and the gradual learning of what deserves attention. [6]
An experienced engineer can sense that an assembly will cause problems without being immediately able to transform their intuition into an equation. A physician combines clinical data with a perception of context that is not entirely in the file. A seasoned salesperson detects hesitation before the customer formulates it. A craftsperson adjusts their gesture because the material does not react exactly as expected. In each case, the knowledge is real, but it is not fully explicit.
Generative artificial intelligence has reinforced the idea that all knowledge could be extracted, indexed, and then made accessible through a conversational interface. This vision confuses information with understanding of the profession. An agent can summarize a feedback report. It does not necessarily possess the sensitivity built by twenty years of confrontation with exceptions. It can apply a rule with regularity. It does not always know how to recognize the moment when this rule ceases to be relevant.
The risk is then twofold. The company may overestimate what its system actually knows. It may also discover too late that the departure of a collaborator has taken away part of its ability to decide. Automation therefore does not make tacit knowledge disappear. It finally forces us to consider it as a strategic asset.
From Individual Knowledge to Collective Intelligence
The work of Ikujiro Nonaka offers a second framework for understanding. In his 1994 article on organizational knowledge creation, and then in his research with Hirotaka Takeuchi, he describes knowledge as a continuous movement between what remains tacit and what can be made explicit. An organization learns when experience is shared, formulated, combined with other information, and then reintegrated into new practices. This process, often summarized by the SECI model, reminds us that a document base is never an end in itself. [7]
The Ford case illustrates this circulation. Veteran engineers do not simply transmit answers. They observe designs, question assumptions, recount past failures, and make visible the reasoning that allows their repetition to be avoided. AI can then monitor more points, search for similar configurations, and repeat certain controls with a regularity difficult to achieve manually. But the machine only extends this knowledge when the organization has created the conditions for its transmission.
This perspective profoundly changes knowledge management. It is no longer enough to archive meetings or pour documents into an internal search engine. We must preserve the reasons for decisions, abandoned options, alert thresholds, exceptions, and warning signs. We must also preserve spaces where the results produced by the machine can be challenged. A company’s memory does not reside only in what it stores. It depends on its ability to circulate experience between generations, professions, and systems.
AI Agents Reveal the Company as It Actually Functions
This is where Weiwei Hu’s analysis takes on its full force. In “Redesign Work Before You Add More AI Agents,” published on July 8, 2026, in Towards Data Science, she describes teams enthusiastic about AI, but still dependent on old workflows. Information comes from several sources, is copied into Excel, goes through successive validations, and sometimes ends up in a system only after numerous manual exchanges. Most of the reasoning remains in the heads of collaborators or in conversations that no one knows how to exploit sustainably. [8]
Adding a chatbot or agent on top of this organization does not eliminate its fragmentation. It can make the interface more modern and speed up a few tasks, without changing the path followed by the decision. An agent that prepares a commercial response does not necessarily reduce delays if the file must still go through four validations. An agent that consolidates data does not create a reliable source when three departments use different definitions. An agent that automates a decision does not clarify who will have to answer for it in case of error.
AI agents thus become a powerful organizational revealer of the 21st century. They highlight the distance between the official organization chart and the way work actually gets done. They reveal silos, hidden dependencies, ambiguous responsibilities, and the multiple compensating operations that collaborators perform every day to keep the system functioning. As long as these adjustments remained human, the company could fail to see them. As soon as it seeks to automate them, it must name, explain, and govern them.
This revealing function distinguishes agentic AI from previous waves of digitization. A copilot can assist a user in a task. An agent can consult data, call tools, and intervene in several steps of a process. The more its autonomy increases, the more costly organizational inconsistency becomes. A contradictory rule no longer simply produces an imprecise response. It can trigger an erroneous action, repeat it, and spread it on a new scale.
The Race for Use Cases Masks the Concentration of Value
Many companies have nevertheless begun their AI strategy with an inventory of use cases. Each profession has proposed its ideas, innovation departments have launched pilots, and suppliers have multiplied demonstrations. This phase has allowed learning. It has also created dispersed portfolios, where the easiest projects to present have sometimes received more attention than the processes whose transformation would have required a real organizational change.
McKinsey cites the case of Johnson & Johnson, which had identified nearly 900 potential uses of generative AI. According to this analysis, 10 to 15% of the initiatives concentrated about 80% of the estimated value. The company therefore redirected its resources to the most promising areas. [9] The data, initially reported by the Wall Street Journal, is not an audited financial result. Nevertheless, it reflects a known reality of technological transformations: experimentation is diffuse, but value is concentrated.
A more mature strategy therefore begins with a value mapping. Where does the company lose time, margin, or quality? What delays deteriorate the relationship with the customer? What decisions are slowed down by the search for information? In which processes does dependence on a few experts create a major risk? Where can AI open a service, product, or business model that was not possible before? These questions are less attractive than an agent demonstration. Yet they are the only ones that allow deciding where to invest.
The BCG AI Radar 2026 shows the intensity of expectations. The companies surveyed expected to almost double the share of their revenue devoted to AI. Nearly 72% of CEOs considered themselves the main decision-makers on the matter, and nine out of ten believed that agents would contribute to a measurable return. [10] This confidence is not proof of results. On the contrary, it makes the selection of processes and measurement discipline even more necessary.
The Workstation Gives Way to the Workflow
Traditional organization describes work through functions and job descriptions. Agentic AI imposes a more precise unit of analysis: the workflow. A commercial, industrial, or administrative activity consists of steps of research, verification, decision-making, execution, coordination, and control. Some can be automated. Others can be augmented. Still others must remain under human responsibility, not because technology would be incapable of intervening, but because the decision involves judgment, a relationship, or responsibility that the company cannot delegate.
This decomposition leads to a more fruitful question than that of replacing professions. Who should do what in the system that produces the result? An agent can prepare data and spot anomalies. An expert can interpret ambiguous cases. A manager can arbitrate exceptions and modify rules when the context evolves. The challenge is not to choose between human and machine, but to design cooperation in which each intervenes where they create the most value.
The Microsoft Work Trend Index 2026 describes the most advanced users, called Frontier Professionals. They represent 16% of respondents. They are distinguished less by the quantity of prompts produced than by their ability to use agents in multi-step processes, to question workflows, and to establish shared standards. [11] The study is based on 20,000 users in ten countries and on anonymized signals from Microsoft 365. It does not demonstrate that these practices mechanically cause better performance. However, it shows that maturity is moving from individual mastery of the tool to collective design of work.
The manager’s role changes accordingly. They no longer simply distribute tasks among collaborators. They orchestrate a hybrid system, define escalation thresholds, verify the quality of decisions, and organize learning from errors. McKinsey describes this evolution as the transition to skill partnerships between people, software agents, and robots. [12] The managerial function is gradually becoming an architecture function.
What the EntrepreneurIA Interviews Already Show
Since the launch of EntrepreneurIA, interviews conducted with over a hundred entrepreneurs, experts, and executives reveal a constant. No serious project creates lasting value through the quality of the model used alone. Initial gains are often visible: time saved, accelerated research, better documentary quality, or targeted automation. The transition to scale then depends on much less spectacular elements: data quality, integration into the information system, adoption by teams, governance, and management’s willingness to modify processes.
Arnaud Pinte, founder of iPepper, emphasized the need to start from a concrete problem, to test quickly with the business units, and to measure the result before industrializing. Christophe Gaultier, at OpenText, highlighted the risks of Shadow AI and the need to govern unstructured data. In heritage professions, AI opened another perspective: documenting gestures, indexing site feedback, and transmitting expertise that still resides largely in the experience of craftspeople. These examples do not belong to the same sector, but they describe the same transformation. AI becomes useful when the company knows how to connect technology to real work.
Luc Julia’s remarks also provide necessary caution. A system’s ability to generate a response or execute an action should not be confused with a general understanding of context. This distinction becomes more important as agents gain autonomy. A machine can act without understanding as a human understands. The organization must therefore retain responsibility for determining limits, interpreting exceptions, and answering for consequences.
EntrepreneurIA thus reveals a paradox. The more accessible tools become, the more differentiation shifts to what cannot be easily purchased: business knowledge, trust between teams, the quality of arbitration, and the organization’s ability to learn. The AI agent does not erase these assets. It increases their importance.
The Leader Becomes Architect of the Augmented Organization
The deployment of AI agents can therefore no longer remain a portfolio of projects managed at the periphery of the company. It touches on the distribution of responsibilities, internal control, skills, customer relations, and sometimes the business model. It becomes a matter for the executive leadership.
An executive team must first be able to formulate the desired result. It must then identify the complete process that produces it, the tacit knowledge on which it depends, and the decisions it accepts to entrust to an agent. It must determine the situations that require human validation, the owner of the economic result, and the person responsible in case of incident. As long as these answers remain imprecise, adding agents mainly increases complexity.
The leader therefore does not just choose a technology. They design an augmented organization. They decide which steps should disappear, which information should become common, which experts should be mobilized, and how learning will be transmitted. They must also accept that an AI project may lead to eliminating a procedure, merging two responsibilities, or challenging a historical indicator. Transformation begins precisely where the company stops adapting AI to its old functioning and accepts to redesign that functioning.
Govern Autonomy Before Generalizing It
This transformation cannot be separated from governance. An agent that prepares a note presents a limited level of risk. An agent that modifies a file, contacts a customer, triggers an order, or recommends a decision directly engages the organization. Access rights, action limits, logging mechanisms, human validations, and escalation thresholds then become architectural choices.
The 2026 Deloitte study, based on more than 3,000 leaders involved in AI programs, indicates that only 21% of respondents declare having a mature model for governing autonomous agents. [13] Concerns particularly relate to confidentiality, security, compliance, intellectual property, and explainability. This delay is not surprising. Many companies are still experimenting with agents as individual tools, while their integration into complete workflows requires mechanisms comparable to those that already frame operational actors.
Governance must therefore not intervene at the end, in the form of legal authorization. It must be integrated into the process design. The company must be able to reconstruct what the agent consulted, the rules it applied, the action it triggered, and the reason why a human did not intervene. The more autonomy progresses, the more this traceability becomes indispensable.
Measure the System, Not the Demonstration
Measurement finally constitutes the test of truth. A company can display numerous agents, licenses, and active users without creating demonstrable value. An agent can reduce the time to write a document, while increasing verification time. It can speed up a response, but multiply exceptions transferred to teams. It can decrease the cost of a task and degrade the result of the complete process.
Evaluation must therefore focus on three levels. The first concerns the agent: accuracy, reliability, speed, cost, and quality of escalation. The second concerns the human-agent system: judgment capacity, distribution of responsibilities, quality of cooperation, and learning from errors. The third concerns the company: cycle time, service cost, quality, customer satisfaction, revenue, or risk reduction. Performance arises from the combination of these dimensions, not from the power of the model alone.
The Ford case regains all its interest here. The automaker did not simply add cameras or algorithms. It modified design reviews, brought teams together, involved suppliers earlier, and placed experience back at the heart of control. AI participated in this evolution. It was not the sole cause. This analytical caution should apply to all projects: just because an agent is present in a workflow does not mean it alone explains the value obtained.
The Real Competitive Advantage Will Be Organizational
AI agents do not constitute a new workforce that one would simply plug into the existing company. They force us to look unflinchingly at the way decisions are made, knowledge circulates, and responsibility is exercised. They make visible the cost of silos, inherited procedures, and information preserved in the memory of a few individuals.
The Ford case does not signal the failure of artificial intelligence. It recalls a more demanding truth: the machine can only exploit the knowledge that the organization knows how to transmit to it and the context it knows how to preserve. Human experience is not a residue destined to disappear as models progress. It constitutes the material from which systems are trained, corrected, and maintained in contact with reality.
For over a century, companies have sought to optimize human work by breaking down functions, standardizing procedures, and measuring productivity. They must now learn to design organizations where people and AI systems together contribute to producing and transmitting knowledge, without diluting responsibility. Leaders who consider AI as an additional tool will perhaps gain a few productivity points. Those who will use it to question processes, preserve expertise, and rebuild decision-making mechanisms will create an advantage more difficult to reproduce.
The most successful companies will probably not be those that have deployed the greatest number of agents. They will be those that have transformed dispersed expertise into collective intelligence, eliminated unnecessary work, and organized cooperation in which human and machine correct each other. AI agents are perhaps the first organizational revealer of the 21st century. They already show that the next battle of artificial intelligence will not only be fought in models or computing power. It will be fought in the quality of the organization that will give them meaning.
[1] ZDNet France, “Mirage de l’automatisation : quand Ford réembauche ses seniors pour sauver son IA,” July 3, 2026. Online source
[2] Sasha Rogelberg, “Ford on why it hired 350 ‘gray beard’ engineers,” Fortune, June 29, 2026. The source specifies that these are former Ford employees and experts from suppliers, recruited over three years. Online source
[3] Ford Motor Company, “Ford is No. 1 Mainstream Brand in J.D. Power Initial Quality Study for First Time Since 2010,” June 25, 2026. Ford mentions approximately 300 veteran engineers in its own summary. Online source
[4] J.D. Power, “2026 U.S. Initial Quality Study,” June 2026. Study based on 78,514 buyers or lessees of model year 2026 vehicles after 90 days of use. Online source
[5] Nora Eckert, “US automotive quality increased industrywide last year, with Ford taking top honors,” Reuters, June 25, 2026. Online source
[6] Michael Polanyi, The Tacit Dimension, University of Chicago Press, first edition 1966, reissue 2009. Online source
[7] Ikujiro Nonaka, ‘A Dynamic Theory of Organizational Knowledge Creation’, Organization Science, vol. 5, no. 1, 1994, pp. 14-37. Online source
[8] Weiwei Hu, “Redesign Work Before You Add More AI Agents,” Towards Data Science, July 8, 2026. Full text provided for the preparation of this article. Online source
[9] Bengi Korkmaz, Sandra Durth, Vincent Bérubé, and Lucia Darino, “Rewiring Talent to Value in the age of AI,” McKinsey & Company, June 18, 2026. The Johnson & Johnson case is attributed to the Wall Street Journal. Online source
[10] Boston Consulting Group, “As AI Investments Surge, CEOs Take the Lead,” BCG AI Radar 2026, January 15, 2026. Online source
[11] Microsoft, “2026 Work Trend Index: Agents, human agency, and the opportunity for every organization,” May 5, 2026. Online source
[12] McKinsey Global Institute, ‘Agents, robots, and us: Skill partnerships in the age of AI’, November 25, 2025. Online source
[13] Deloitte, “The State of AI in the Enterprise 2026: The Untapped Edge,” January 2026. 21% of organizations surveyed declare having a mature governance model for autonomous agents. Online source
