Enterprise AI: 6 Governance Guidelines and MIT's Findings

The MIT Media Lab highlights a gap between massive investments in enterprise generative AI and the actual value created. An executive in charge of AI transformation in retail proposes six concrete guidelines for moving from pilots to governed execution: prioritized principles, graduated trust, unified context, and exception-based management, with human judgment as the central resource.

The Evolving Human Role: Judgment Becomes the Scarce Resource

Resistance to AI adoption often stems from professional identity and the question of each person's role when machines take over execution. The approach starts from an observation: the core of the business has always been judgment. AI enhances the value of judgment, which, when well-articulated, can guide thousands of decisions each day. Taste differs from judgment: taste allows for the evaluation of the relevance of a question and the selection of the best option among many possibilities, while a machine can generate 100 options without selecting the most appropriate one. The transition to AI thus values qualities such as systemic thinking, precision in requirements, and clarity about failure modes, which are present in many IEEE members. From this perspective, governance places judgment and taste at the center of work, and those who embrace this evolution will define the profession in the age of AI. This mode of operation should enable an organization to grow without proportionally increasing its workforce.

The GenAI Divide: Many Investments, Little Measured Impact

The MIT Media Lab estimates that companies have invested between $30 and $40 billion in generative AI. The authors note that the majority of organizations studied have not demonstrated measurable impact on profits and losses, estimating that about 5% of integrated pilots create substantial value. They refer to this gap as the GenAI Divide. From the ground perspective, the main obstacle is generally not the technology, often shared with the few successes, but the lack of people capable of leading the systems and supporting the outcomes.

Moving Beyond "Human Middleware" and Restoring Judgment's Place

Many professionals serve as an interface between tools and teams, a role of "human middleware," a trap of the administrator that relies more on architecture than on individuals. AI agents effectively transmit information but do not decide which figures to consider, what risks are evident, or what acceptable compromises are. In customer-facing roles, virtual assistants can answer common questions and guide towards products, services, or steps, but the arbitration of complex situations remains reserved for human judgment.

Transitioning from Rules to Principles and Clarifying Decision Rights

When systems make thousands of decisions per hour, the recommendation is to abandon detailed rules in favor of prioritized principles. Examples of traditional rules, such as managerial signatures beyond a threshold or systematic code reviews, work at human speed but fail at scale. Principles such as not harming the customer, telling the truth even at the cost of a sale, or protecting the economy before acting guide the arbitration. This framework relates to the notion of decision rights, or formal authority over each decision, making the drafting of a principles library an act of leadership for technologists and business leaders.

Coding Culture: Constitution, Doctrine, and Playbook

Another proposal is to embed company values and policies directly into the code, in the form of machine-readable instructions. The system consists of three layers: a constitution that sets inviolable rules, such as the prohibition of stating unsupported facts; a doctrine that specifies how to compete and the acceptable trade-offs, for example, preferring a long-term relationship over an immediate sale; and a playbook that gathers operational tactics for each task.

Graduated Trust and the CCRAG Loop for Explainable AI

Rather than a binary switch, trust is envisioned as a thermostat: each agent decision receives a score relative to the principles. Above a threshold, the agent acts; below it, a human decides, and their response retrains the system. This transparency, referred to as a "glass box," requires a complete view of the context. The CCRAG loop breaks down as follows: connections that ingest raw data; a context that links them into a graph providing a form of memory; reasoning that decides; actions that reinject the decisions; and governance that ensures alignment with intent. Organizations often prioritize reasoning and under-invest in context and governance, while context enriches with each interaction. In this framework, the directive is to manage by exception: the machine leaves routine flows intact and escalates low-confidence, ambiguous, or critical cases to human decision-making.

A Framework from the Retail Ground

These proposals come from an executive leading AI transformation at a Fortune 100 retailer and who co-authored The Enterprise Brain. The stated goal is to provide the customer with relevant expertise at the right moment rather than promoting AI. In this approach, technical roles include defining what systems can decide autonomously, what must be escalated, and what remains prohibited. This evolution is described as a "governor change," where operators, without necessarily coding, set, for example, the price exceptions that an agent can approve or must submit to a human.