Companies don't just add AI tools — they pick a role for it: a tool, an assistant, or almost a coworker.

This explores the main patterns that shape how organizations actually put generative AI to work. The answer is not about text versus images. It's about what role the AI plays, what problem it's aimed at, and how much the surrounding work gets rebuilt around it. The corpus suggests three overlapping ways to sort this: the problem being solved, the role the AI is given, and whether the organization adds AI on top of existing work or redesigns the work.

Start with problems, because that's how companies themselves describe it. MIT's study of more than 20 firms found adoption grouping around three recurring needs: bottleneck problems (too much routine work), cafeteria problems (pulling in input from many experts), and learning-curve problems (helping novices get up to speed) What problems do companies solve with generative AI?. Each maps onto field evidence. The learning-curve case shows up in the BCG experiment, where weaker consultants gained the most from GPT-4. There's a catch, though: that held when AI use was mandatory. When people could choose whether to use it, the gap-closing effect could reverse Does AI help lower-skilled workers catch up to top performers?. The cafeteria case shows up at Procter & Gamble, where individuals using AI matched two-person teams without it and produced solutions that drew more evenly across professional backgrounds Can generative AI replace the benefits of having a human teammate?. So how a tool is rolled out, required or optional, is itself a design choice that changes who benefits.

The second dimension is role, and here organizations are in the middle of a shift. Most real deployments are still conversational and experimental. At Argonne National Laboratory, staff used an internal chatbot mainly for structured writing, and few had built it into regular work How are national lab staff actually using generative AI?. Agentic AI breaks that pattern. In an MIT Sloan/BCG survey, 76% of executives saw it as a coworker rather than a tool, which creates a real management problem: it has to be governed as both a workforce and an asset at once Does agentic AI fit the tool-or-coworker framework?. Deloitte goes further. It argues agentic deployments fail when agents are layered onto old workflows, and succeed only with end-to-end process redesign, new data architecture, and new management models Does agentic AI succeed through process redesign or workflow layering?. Taken together, the chatbot-as-helper mode and the agent-as-coworker mode look like two different organizational projects, not two steps on one path.

The less obvious finding is that deployment patterns reshape the work itself, not just its speed. Heavy AI users increased their actions in productivity apps by 21.2% but their communication actions by only 7.1%, a shift toward solo documentation and away from coordination Does generative AI shift knowledge workers away from communication?. Larsen-Ledet adds a usability angle. Today's tools make workers restate context the system should already know, which she calls mundane articulation, and real workplace value depends on removing that overhead Can generative AI reduce overhead work that adds no value?. On the interface side, a proposed design framework for AI-generated interfaces sorts them by audience, constancy (how stable the interface stays), control, and timeframe How should generative UI be organized as a design space?. A benchmark is a reality check here: generative UI tools left about a quarter of stated design rationales unimplemented Do generative UI tools actually implement their stated design rationales?.

For a wider frame, Farrell and colleagues suggest treating large models less like employees and more like markets or bureaucracies: cultural technologies that compress and reorganize what people collectively know Are large AI models actually intelligent agents or social technologies?. That lens explains why the 'tool or coworker?' question keeps feeling slightly off, and why the biggest effects show up in how teams coordinate rather than in any single task. One gap: the corpus is strong on frameworks and single-company studies but thin on comparing deployment models across industries head to head.