The 7-stage roadmap for human-AI collaboration (2026 Edition)
My most comprehensive framework for understanding how AI enters knowledge work. Each stage builds on the previous—you can't skip steps.
2026 edition published June 10, 2026 — Original post. Supersedes the June 24, 2025 original.
What changed from the 2025 version
The 2025 original framed each stage around what the AI does. The 2026 version reframes around what the worker does and becomes—the impact is easier to see that way. Two other changes: the timelines were off by roughly 3x (stages predicted for 2027+ were live by January 2026), and Stage 3 (cognitive extension / second brain) is entirely new, since that way of working only became viable in January 2026. The roadmap diverges from the 2025 version from Stage 3 onward.
The stages
Stage 1: AI as faster search Simple one-and-done tasks: summarizing documents, writing emails, answering questions. The worker types a question, gets an answer, moves on. Most workers are still here. Unchanged from 2025.
Stage 2: AI as a thinking partner Workers go back-and-forth with AI instead of asking one-off questions. They start loading documents into conversations and using Projects/Notebooks to give the AI project-level context shared across multiple chats. Usually where voice and dictation start, and usually the first "aha" moment. Maybe 20% of workers are here today.
Stage 3: AI as a cognitive extension (new for 2026) The flip: instead of bringing documents to the AI, the worker brings their AI to all their docs—and everything else. The AI points at all of it: documents, emails, notes, meeting transcripts, half-finished ideas. This is the "context vault" / "second brain" / "personal Wikipedia with AI read/write access." AI stops being an occasional tool and becomes part of how you think. It holds your full context all the time, so you never start from a blank page, and it can connect something said in a meeting today to a document written two months ago. This stage only became viable in January 2026.
Stage 4: AI as a multi-tool agent Once the AI has persistent context, the next step is letting it reach out into the world and do things. CUAs, browser operation, MCP connections, specialized sub-agents—the "claws" of AI. The AI is "doing" more than "thinking": pulling data, filling forms, running analyses, driving apps and websites. Because the claws are powered by a brain (Stage 3), the AI uses its skills to know how to get things done rather than the worker dictating each step. Apps start to feel like compatibility layers—Word, Excel, Outlook, Teams are legacy interfaces AI uses when it has to, not where work happens.
Stage 5: AI as a fleet Workers use multiple AI systems. The primary AI fires sub-agents to fan out across parallel tasks. Many apps and systems have their own AI interfaces, and workers' AIs coordinate with other AIs the same way human workers coordinate with each other. Workers shift from doing to directing—everyone becomes a manager. The skills that matter are delegation, review, and deciding what's good enough. Genuinely uncomfortable for most, the way first-time managers struggle to stop doing the work themselves.
Stage 6: AI as a pod The agents don't stop working when the human does. They work overnight, coordinate with other agents, and queue up questions and decisions for whenever the human next checks in. The model flips: the AIs do the work, reaching out to the human for guidance as needed, rather than the human directing every step. Each employee has their own pod. The unit of work stops being "one human for one 8-hour day" and becomes a small team—one human plus a handful of agents—running more or less continuously.
Stage 7: The published self (optional fork) Not chronological—it can happen at any point after Stage 3. Once your second brain holds your context, judgment, and way of working, you can publish it for other people's AIs to connect to, letting them draw on your expertise directly without you in the room. Not just for public influencers; it will be used extensively inside companies as workers pull company-wide, department-level, or individual-expert feeds into their own AI context systems. I publish my own at brianmadden.ai.
How to use this framework
Each worker moves through the roadmap at their own pace, so different workers at the same company will be at different stages—it will be jagged for the next few years. The trajectory is clear even if the timing isn't: plan for stages to arrive sooner than feels reasonable, because the 2025 version's timeline was wrong by ~3x.
Keep in mind that "when AI can do a thing" and "when all workers actually work this way" are very different timelines. Even where a stage is technically possible, it will be years before every worker is there.
My self-reported position as of June 2026: deep in Stage 3, just entering Stage 4, and fully in the optional Stage 7. Predicted trajectory: deep in Stage 4 by end of 2026 and probably starting Stage 5, with Stage 6 real by end of 2027.
Relationship to other frameworks
This roadmap covers the mechanics of human-AI collaboration. The five levels of AI in knowledge work framework covers what the human-AI relationship is like at each level—they're complementary, not competing. Stage 3 here is what the cognitive stack calls the brain layer; Stage 4's claws are its agentic sub-process and interface layers.
Lineage: the self-correction that produced this version
In February 2026 I publicly corrected the 2025 roadmap as "faster horse" thinking—it described evolution within the existing model of work, while second brains showed the model itself dissolving. The 2026 edition folds that correction in: Stage 3 puts the cognitive extension (the invisible 80%) at the center, and Stage 7 becomes the published self rather than AI-orchestrated work. The 2025 roadmap wasn't wrong—it was incomplete. It described the visible 20%; the frontier work is in the invisible 80%.