Learning by Doing: Why Active Use Remains the Only Real AI Strategy

Direct engagement with artificial intelligence remains the single most effective way to understand its capabilities and steer its integration into professional life.

For the past several years, the global conversation around artificial intelligence has swung wildly between breathless hype and apocalyptic warnings. Technology executives promise fully automated enterprises, while regulatory bodies attempt to construct guardrails around systems whose inner workings remain difficult to predict. Amid this noise, a more grounded thesis has taken root among researchers who study how work actually changes: the only reliable way to understand AI is to put it to work every day.

This principle forms the backbone of the work by Ethan Mollick, a professor at the University of Pennsylvania’s Wharton School and faculty director of its Generative AI Labs. When Mollick released his first major book on the subject, Co-Intelligence, the technological landscape was defined by the early novelty of large language models. The software was impressive but erratic, acting like a brilliant, distractible intern whose every output required painstaking line-by-line verification. In his subsequent work, including Co-Existence, the focus shifts from basic familiarization to structural integration. The technology has evolved from isolated text generators into persistent systems capable of taking actions, chaining reasoning steps, and reshaping organizational hierarchies. Yet the underlying prescription remains unchanged: individuals and leadership teams cannot evaluate AI from a distance.

Beyond the Spectator Sport

Many organizations approach technological adoption as an administrative exercise. Committees are formed, enterprise risk assessments are commissioned, and consulting groups are hired to produce multi-year roadmaps. In the case of modern machine learning, this detached approach consistently misreads the technology. The core mechanics of large models do not conform to traditional enterprise software rollout schedules. Their interfaces are deceptively simple natural language prompts, yet their capabilities change every few months as base models gain multimodal inputs, larger context windows, and improved reasoning paths.

Reading about what a system can theoretically achieve is not the same as discovering where it breaks in practice. When users interact with models across varied, everyday tasks, they quickly run into the uneven boundaries of machine capability. A system might translate complex legal jargon into plain language with astonishing precision, only to stumble over basic arithmetic or invent factual citations with total confidence. Observing these contradictions in real time dispels both unearned awe and dismissive cynicism. It teaches practitioners how to formulate queries, how to structure verification steps, and when to abandon automated assistance entirely.

Passive observation also creates a false sense of security. Leaders who only consume high-level summaries tend to believe that deployment is merely a technical upgrade to be managed by the IT department. In reality, generative systems affect the cognitive core of knowledge work. Deciding how to distribute tasks between humans and algorithms requires first-hand familiarity with what the models do well and where their reasoning falls apart.

Mapping the Jagged Frontier

One of the most persistent challenges in operational AI is what researchers call the jagged technological frontier. Unlike human competence, which generally scales predictably across related tasks, machine intelligence expands unevenly. A professional who writes high-quality code can usually be trusted to write clear technical documentation. AI models do not adhere to this continuity. A model may generate an advanced architectural plan for a software deployment yet fail to identify a trivial syntax error in an accompanying configuration file.

This jagged frontier cannot be mapped using executive summaries or vendor marketing claims. It shifts constantly with every model update, fine-tuning pass, and system prompt revision. The only way an operator or team lead can navigate this terrain is through persistent, iterative testing. By testing everyday problems against these models, users build an intuitive map of the system's current boundaries.

This hands-on calibration also alters the nature of expertise. Deep domain knowledge does not become obsolete when an AI enters the workflow. Instead, domain knowledge becomes the primary filter for quality control. An inexperienced worker using an AI model will often accept a plausible-sounding but flawed output because they lack the background to spot the error. A seasoned professional, by contrast, uses their hard-won knowledge to interrogate the system, challenge its assumptions, and push it toward a workable result. The human role shifts from raw generation to rigorous editing, direction, and oversight.

The Trap of Abstract Policymaking

When enterprise leadership lacks direct experience with AI, corporate policy tends to drift into two unhelpful extremes: blanket bans or uncritical mandates. Both approaches fail to account for how work actually happens on the ground.

Broad prohibitions against the use of public or commercial models rarely prevent employees from using them. Instead, they drive adoption underground, giving rise to unmonitored shadow AI. Workers facing demanding deadlines will quietly paste meeting notes, source code, and drafts into external tools to save hours of manual labor, bypassing corporate security protocols precisely because no official, practical path has been provided. The organization absorbs all the privacy and security risks without capturing the collaborative benefits.

Conversely, top-down mandates that require teams to incorporate AI into every deliverable often generate performative compliance. Employees use models to generate bloated, generic prose or redundant diagrams to satisfy a metric, degrading overall work quality without improving operational efficiency. Useful integration occurs organically when workers have the autonomy, the enterprise-grade tools, and the basic literacy to solve specific bottlenecks in their daily routines. Practical policy emerges from observing what working teams discover, formalizing effective workflows, and providing clear boundaries around data privacy and attribution.

Designing Pragmatic Workflows

Moving past the novelty phase requires treating generative systems as collaborative partners that require structured management. This means establishing clear divisions of labor between human intuition and machine processing. Routine data extraction, initial drafting, code translation, and synthetic stress-testing can often be delegated to models, provided a qualified professional reviews the output.

Furthermore, the technology is rapidly expanding beyond single-turn conversational interfaces into autonomous, multi-step agents. These systems can execute complex sequences: searching a knowledge repository, cross-referencing findings against an internal database, drafting a technical summary, and submitting a ticket for human approval. Managing these agents requires a different skill set than traditional software operation. It demands clarity in defining objectives, anticipating unintended side effects, and creating verification checkpoints.

As highlighted in a recent review by Bloomberg, Mollick argues that individuals and organizations cannot afford to remain passive observers while these shifts take place. Waiting for the technology to mature or for industry standards to solidify ensures that others will dictate how work is structured. Those who actively experiment today are the ones who will establish the norms, workflows, and competitive baselines for their industries tomorrow.

Why It Matters

For infrastructure operators, data centre managers, and field technicians, these lessons are particularly urgent. As generative and multimodal models move out of white-collar offices and into physical engineering environments, hands-on literacy becomes a safety and operational priority. Technicians troubleshooting complex fiber networks or power distribution systems in remote facilities cannot rely on generic vendor guidelines. They need the practical skill to query on-device AI assistants, parse technical manuals against live telemetry, and rapidly spot model hallucinations before implementing physical fixes. In high-stakes telecommunications and facility operations, understanding the jagged frontier is not an academic exercise. It is the difference between a fast, safe restoration of service and a costly, system-wide outage.