Ask most businesses if they are using Artificial Intelligence (AI) and they will likely say, yes. Some businesses have advanced predictive models and Large Language Models (LLM) based customer service chatbots. Many others who are “adopting AI” are simply using ChatGPT to draft emails. While AI is a buzzword, we would argue that it is worth the hype. This article proposes a framework for navigating and adopting AI and LLMs.
Scope of LLMs
In the map of AI systems and their relationship to humans shown above, we have the LLM, also known as the brain, at the centre, which we typically interact with via chatbots, such as GPT or Claude. Increasingly, we also interact with AI in third-party apps. Some are easy to spot, like the Microsoft Copilot in our Outlook. Others are better hidden, such as the LLM in our TikTok algorithm. Since 2025, we have seen the rise of agentic workflow platforms where LLMs automate well-specified tasks – almost like an assembly line of “mini” LLMs. For example, a workflow might have multiple agents reading, sorting, and responding to your emails. In the future, many forecast the emergence of Artificial General Intelligence (AGI), which would be science-fiction-like AI systems. Think Jarvis from Iron Man or Skynet from The Terminator. LLMs are a growing landscape, which means that AI adoption frameworks must look beyond simple chatbots. As for humans, we see ourselves in the loop. We could build workflows and audit outputs, assign supervisory roles, setting model goals, and providing strategic oversight.
Speed of improvements
While we might not end up in Terminator-dystopia, we should not underestimate what AI is capable of. For most of the 20th century, scientists believed neural networks were incapable of playing chess or recognising images. In the 2010s, advancements in deep learning proved otherwise. Ten years ago, we believed deep learning would be incapable of problem-solving or understanding sarcasm – then came LLMs. In 2021, state-of-the-art LLMs could do Grade One maths. In 2023, they passed the Scholastic Assessment Test and in 2025, CFA Level III exam. Similar leaps have been made in writing, coding, and video generation. AI has a track record of doing more than we think, faster than we think.
For businesses, this means “adoption” must be about preparing for the AI of tomorrow.
The paradox
If LLMs are smart, accessible, and improving exponentially, we should have seen an explosion in productivity worldwide. Our big idea is that companies have not experienced the gains from AI because they treat it like another digital tool. AI is a general-purpose technology that requires imagination, behavioural changes, and continuous implementation. In the rest of the article, we explore what is different and what to do.
General purpose technology
The iPhone is not simply a better landline. It is our bank, grocery shop, taxi, camera and more. In the same way, electricity is more than just a better gas lamp. The internet is more than an encyclopaedia. And so, with AI, we expect it will be more than a “better google”. These widely disruptive technologies are called general purpose (GP).
From history, we know that GP technologies first take over as obvious replacements but later disrupt all industries. Early mobile phones simply replaced landlines but decades later they are making shopping malls obsolete. We call this “replacement theory” and it explains why AI has been more disruptive to certain industries now but will affect everyone later. For example, call centre agents or copywriters or junior developers are frequently cited as at-risk roles, because ChatGPT can answer queries, draft blogs and debug code. As LLMs improve at problem-solving and long-term tasks, more jobs become at risk. To adapt, businesses need imagination. The internet created new industries and jobs, from social media to cloud computing. Those with vision will survive in an AI-native economy.
Behavioural approach
Unlike normal digital transformations, e.g., moving from paper to digital records, AI adoption is exposing strong personality differences that must be handled to get human buy-in. We propose the following quadrant to navigate employees’ behaviour.
The innovator is enthusiastic about AI and uses it extensively but does not always check output and might waste time experimenting. By contrast, the pragmatist uses AI for clear and reliable wins. Unlike the pragmatist, the conserver is distrustful of AI’s reliability and safety and prefers to do things “the way they always have”. Lastly, the resistor does not use AI, but their reasons are more emotional, relating to job security or fear of change. These personality differences explain why simply issuing Microsoft Copilot licenses does not boost employee productivity.
To manage emotions, we think safety and guardrails are key. Clear rules around uploading data and using work-approved tools. Double-checking standards in teams ensures that output is reasonable. Encourage an innovative culture where it is safe-to-fail. Have clear messaging around job security and how people are expected to adapt. To increase usage, tools and super-users are not enough. We suggest a combination of upskilling workers, sharing use cases, offering explicit incentives, and investing in AI licenses. By understanding behavioural differences, the quadrant framework increases the likelihood of safe and wide adoption.
Continuous implementation
For normal digital tools, requirements are clear, and we can plan months in advance. AI is unpredictable, so detailed upfront planning often means planning for a version of the technology that is already outdated by the time we ship. We propose two complementary approaches to adopt AI dynamically: Vibe work for day-to-day tasks, and AI Build for engineering new AI tools. Vibe work means drafting with AI and then refining conversationally. It reduces the friction between brainstorming and implementation, boosting both creativity and productivity. AI Build, by contrast, is about engineering new AI products: dream big, then build what is possible. This differs from waterfall, where the plan is fixed upfront, and from agile, where the scope stays fixed, but delivery reiterates. With AI Build, even the goal itself can shift as we learn what is technically achievable – for example, we set out to build an AI portfolio manager assistant but ultimately delivered a position-tracking tool. AI Build favours fast, experimental, and flexible project goals, so that as the technology improves, we improve with it.