On 27 September 2023 I delivered the opening keynote at bbv's "KI Revolution" event. The talk, titled "AI first how a digital native thinks about generative AI," framed the shift from classical AI to generative AI and introduced the agent paradigm as the bridge between a language model and real enterprise value.
Classical AI (pre-2022) solves narrowly defined problems. YOLO detects 90 object classes in an image; the business question is "does YOLO solve my problem?" Generative AI flips this: ChatGPT has broad capabilities but solves no specific problem out of the box. The business question becomes "where would a new worker help me, and can generative AI be that worker?" This reframing moves AI from the algorithm category into the workforce category.
An AI agent maps directly onto a human employee. A human has a job description; an agent has a task. A human has intelligence; an agent has ChatGPT for text understanding and reasoning. A human has a profile and skills; an agent has a character definition and tools (reading files, sending emails, querying databases). This analogy makes the technology accessible to non-technical decision-makers.
I presented a four-stage model for enterprise AI adoption. Stage one, instruction: correctly prompting and configuring the model via system prompts, user prompts, and parameter tuning (temperature, etc.). Stage two, context: connecting the agent to internal knowledge bases through retrieval-augmented generation, so answers draw on domain-specific data. Stage three, tools: letting the agent interact with enterprise systems through APIs, reflect on which tools to use, and iterate until the task is solved. Stage four, autonomy: the agent receives a high-level goal and independently plans, executes, adapts its strategy, and iterates, calling on other agents or humans as needed.
I closed with a set of predictions. ChatGPT would gain image understanding within a month (it did). Further modalities would follow within a year. Within two years, some companies would employ many virtual workers. Within three years, companies would divide into those using AI and those not. Within five years, only one group would remain.
Classical AI solves clearly defined, narrow problems (e.g. YOLO detecting 90 objects in an image). Generative AI has broad capabilities without solving a specific problem out of the box (e.g. ChatGPT generates text). The key question shifts from 'Does this AI solve my problem?' to 'Where would a new worker help me, and can generative AI be that worker?'
An AI agent is a virtual employee. Like a human worker, it has a job description (task/goal), intelligence (ChatGPT for text understanding and reasoning), and a profile (character traits and tools such as reading files or sending emails).
1. Instruction: correctly prompting and configuring AI. 2. Context: giving AI access to internal, domain-specific knowledge. 3. Tools: letting AI interact with enterprise systems via APIs. 4. Autonomy: deploying AI as an independently acting entity that plans, executes, adapts, and iterates toward a high-level goal.
In one month ChatGPT would gain image understanding. In 12 months further modalities would follow. In two years some companies would have many virtual employees. In three years companies would split into those using AI and those not. In five years only one of those groups would remain.
.
Copyright 2026 - Joel P. Barmettler