# The five dimensions of AI adoption > A five-dimension framework - literacy, engineering, champions, governance and playbook - to turn generative AI adoption into lasting business value. - Date : 2025-11-02 - Lecture : 10 min - Catégorie : strategie-ia - Tags : AI, Data, Automation, Testing - URL : https://www.adservio.fr/en/insights/articles/les-cinq-dimensions-de-l-adoption-de-l-ia ## TL;DR - The initial excitement around generative AI has often given way to disappointment: a year after rollout, only 20% of employees were actually using these tools. - A five-dimension framework structures AI adoption: AI literacy, sound engineering practices, champions and community, AI governance, and an AI playbook. - AI literacy rests on prompt engineering, AI evaluations, critical thinking, and architectural thinking. - Lightweight governance (an innovation lab, a technology radar, AI ADRs) helps balance innovation against risk. - AI champions and a shared AI playbook accelerate collective learning by capitalizing on both successful and failed use cases. ## Introduction: why the generative AI hype fades so fast For technology leaders, when it comes to AI for software delivery, several key questions need to be considered: what do I want to accomplish by adopting AI? How do I balance innovation and risk? How does this impact my team, including myself? How do I measure the real impact of AI on performance and quality? When generative AI became a major trend with the rise of ChatGPT, and later the agentic coding assistants that followed, many organizations wanted to jump in without waiting. Bold claims about productivity gains, GitHub Copilot announcing developers 55% faster, some vendors promising to halve delivery timelines, only fueled that excitement further. That initial optimism has nonetheless given way to disappointment in most organizations. A year after rollout, most found that only 20% of employees were actually using these tools in their daily work, far from the promised value. In 2026, that share is climbing but still lags well behind the initial marketing announcements, largely because adoption was driven by tools rather than by business use cases. It is time to move away from a purely tool-centric view. With this in mind, we developed a five-dimension framework to structure AI adoption sustainably: AI literacy, sound engineering practices, champions and community, AI governance, and an AI playbook. Let's examine each of these dimensions in detail, and what they really mean for an organization trying to turn initial excitement into lasting value. ## AI literacy: the four skills that make the difference Generative AI has had a significant impact on software development, both in how we build and how we maintain systems. But for it to be used effectively, creators, developers, business analysts, QA, experience designers, need to develop solid AI literacy, which cannot be improvised and is built through repeated practice. ### Prompt engineering and AI evaluations In order to use generative AI to produce high-quality, maintainable and secure code aligned with good engineering practices, it is essential that users know how to properly formulate requests to a model. This requires mastering the context given to the model, breaking complex tasks down, and iterating on instructions rather than accepting the first response as-is. Large language models remain non-deterministic and inherently unpredictable. Testing using evaluations ("evals") is essential for the technology to stay consistently useful and reliable over time. Tooled evaluation frameworks now make it possible to track accuracy, drift, and regression metrics with every model or prompt change, rather than relying on a subjective impression of quality. > Related read: [How to Evaluate an LLM System](https://www.adservio.fr/en/insights/articles/comment-evaluer-un-systeme-llm): Evaluating an LLM system in production: quality metrics, eval datasets, LLM-as-judge, prompt regression testing and continuous drift monitoring. ### Critical thinking and architectural thinking The ability to think critically and independently, minimizing automation bias and anchoring bias, is becoming a professional skill in its own right. Accepting an AI suggestion without questioning it exposes the organization to silent regressions, particularly once the volume of generated code exceeds the team's capacity for human review. With AI taking on specific tasks, professional roles are likely to be elevated by an additional layer of abstraction. The role developers play, for example, is shifting toward one that places greater emphasis on technical leadership, architecture, and validation, rather than line-by-line production. ## Sound engineering practices: AI as an amplifier With AI acting as an amplifier, good engineering practices become even more important. A poorly tested pipeline or a team without review discipline will see its flaws amplified rather than corrected by AI, the tool accelerates whatever already exists, for better or worse. ### Test-driven design and fast feedback Test-driven design, fast feedback, and automation of repetitive tasks form the foundation on which AI can be safely layered. A solid test suite acts as a safety net that lets teams accept AI-generated suggestions with confidence, knowing any regression will be caught quickly rather than discovered in production. ### Countering AI noise with critical thinking Critical thinking, the ability to analyze, organize, consider counterfactuals, and form an opinion, has emerged as the key skill for countering the risks of hallucination, bias, and the creation of "AI noise" in the form of redundant, poorly structured, or falsely idiomatic code. The most mature teams put systematic reviews of AI-generated code in place, with explicit criteria for readability and technical debt, to prevent generation speed from masking a gradual degradation of the codebase. ## Champions and community: bringing the whole organization along AI impacts every part of an organization, not just technology. Adopting AI tools requires informed input from legal (intellectual property protection), finance (future costs and audit traceability), HR (strategic workforce planning), sales (customer reaction is critical), operational risk, and even the board of directors. Its impact can extend as far as cyber-insurance risk and regulation, especially with the growing weight of AI-related compliance obligations in Europe since 2025-2026. ### The ADKAR framework applied to AI Change management must be prioritized to ensure any major change is sustainable. The best results are achieved when people are genuinely placed at the heart of any change initiative and a coherent narrative is shared with every stakeholder. The ADKAR framework, awareness, desire, knowledge, ability, reinforcement, remains a useful tool to structure this journey, particularly for identifying what stage each team is at and adapting support accordingly. > Related read: [How to Build Trust With AI-Powered Coding Assistants](https://www.adservio.fr/en/insights/articles/comment-cultiver-la-confiance-avec-les-assistants-de-codage): Why adoption rates for coding assistants stall, and what actually builds a team's trust in the tool. ### The four traits of a good AI champion AI champions play a crucial role in this dynamic. Four traits are required to be a good champion: learning continuously, since the whole ecosystem is evolving rapidly, while encouraging the wider team; experimenting, since much remains unknown about how AI will redefine software; advocating for AI day to day, leading by example and sharing relevant lessons; and defining, by helping shape the future way of working and the balance between people and AI, as well as the new skills required. ## Lightweight AI governance: balancing innovation and risk Since the start of the LLM revolution, some organizations are already starting to pay down technical debt from earlier decisions made under pressure, such as building their own models without real business justification. How can we make sound investment decisions in this fast-moving world? How can we make sure we balance innovation and risk? How can we build the confidence organizations need to move at the right speed without exposing themselves to undue risk? ### Three concrete levers: lab, radar, ADR An AI innovation lab is a production-like environment where teams get space to run experiments with appropriate guardrails and governance controls. Beyond the environment itself, it is useful to apply work-in-progress limits to cap the number of experiments allowed without disrupting business as usual. An AI technology radar can be an effective catalyst for architecture conversations and act as an information radiator. An AI-specific edition provides visibility into the tools and techniques to adopt, trial, or retire, which significantly reduces the risk of unapproved tool usage within teams. An AI ADR (architecture decision record) makes it possible to effectively communicate AI-related decisions to internal and external stakeholders, reducing the risk of local decisions with negative global consequences. ## Use-case-driven experiments: avoiding the seven anti-patterns It is crucial to avoid certain recurring anti-patterns when running AI experiments. The scattergun approach involves trying many tools or models without a clear hypothesis, leading to wasted effort. Siloed experiments, run without sharing learnings, prevent collective growth and reinforce "not invented here" syndrome. Ignoring feedback loops turns experiments into academic exercises rather than drivers of real impact. Premature scaling, trying to generalize a solution before its value, risks, and operational requirements are fully understood, is just as risky as the lack of clear success metrics, which makes it impossible to evaluate an experiment's effectiveness. Static experimentation, which stops after a single trial, deprives the organization of the continuous iteration needed to keep pace with fast-moving technology. Finally, focusing solely on technical metrics, accuracy, latency, at the expense of business value and user experience durably skews conclusions. > Related read: [Organizing for AI: from experimentation to industrialization](https://www.adservio.fr/en/insights/articles/s-organiser-pour-l-ia-de-l-experimentation): Why 88% of AI POCs never reach scale and how to organize: a dedicated squad, distributed squads, a lean center of excellence and an industrialized AI Factory. ### Mapping the value stream to find where to experiment A good starting point for identifying where to experiment with AI is to use value stream mapping, from story ideation to production and support, identifying friction points, then using AI to reduce system waste through automation or process restructuring. ### A dashboard to avoid duplicated effort Use-case-driven experiments, centered on fast feedback loops and a clearly defined hypothesis, build organization-wide transparency around what has been learned. An experiment dashboard, showing past, current, and future status, guarantees this transparency and prevents two teams from separately rediscovering the same limitations of the same tool. ## AI playbook: capitalizing on collective learning An AI playbook is the organization's single source of truth for everything related to AI, from principles to guidelines and practices, with a list of proven use cases, both successful and unsuccessful. It paves the way for guided adoption rather than every team continually reinventing the wheel. It's important to build a culture of transparency around failed experiments, emphasizing learning rather than punishing failure. The value of an AI playbook lies in being grounded in operational reality rather than theory. It accelerates learning and adoption by providing a collection of proven use cases across the organization; it mitigates risk and ensures ethical AI through clear guidelines; it fosters innovation by drawing on past and most recent insights; and it builds a collective organizational memory that, over time, becomes a genuine strategic asset. ## Expected outcomes and roadmap With successful adoption addressing these five dimensions, AI literacy, sound engineering practices tied to champions and community, AI governance, use-case-driven experiments, and the AI playbook, organizations are likely to achieve measurable outcomes: improved developer experience through reduced waste and cognitive load across the value stream, reduced technology risk from fewer manual operations, faster value creation by leveraging proven use cases with defined business value, and improved agility through capabilities that evolve across the organization. This approach lets teams move beyond initial enthusiasm, fostering sustainable and impactful AI adoption patterns that fundamentally redefine how software is built and delivered in the age of AI. The question is no longer whether to adopt AI, but with what organizational discipline. Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect the positions of Adservio. ## FAQ ### Why does generative AI adoption often disappoint organizations? Because it has been approached purely through the lens of tools. After the initial hype driven by productivity claims (such as GitHub Copilot's 55%), most organizations find that a year later only 20% of their employees are actually using these tools in their role. ### What are the five dimensions of the AI adoption framework? AI literacy (skills and capabilities), sound engineering practices tied to champions and community, AI governance, use-case-driven experiments, and the AI playbook that captures what has been learned. ### How can an organization set up lightweight, effective AI governance? By combining three concrete levers: an innovation lab with work-in-progress limits, an AI technology radar to track adopted, trialed, or retired tools, and AI ADRs to document architecture decisions and make them traceable to stakeholders.