Research Roadmap | AI Wars of 2026

The reader-facing research agenda and evidence triggers for AI Wars of 2026.

Roadmap

This edition follows an unfolding contest. We publish what the evidence can support, make its limits visible, and identify the questions that could change the argument.

The argument so far

Part 0 begins with the human conditions of adoption, then defines the overlapping wars, strategic prizes, and evidence tests that make "winning" meaningful. Part I traces the move from chat to delegated action and examines how the system around a model shapes work. Part II compares the routes companies are testing. Its provisional chapters state what evidence would change their status. Chapter 12 adds research on Chinese consumer assistants; chapter 13 maps interdependent routes without ranking them.

The book's working hypotheses are that intelligence, a host, or a fragmented stack may capture durable value under different conditions. The evidence does not yet establish which position will capture the most value, or identify an overall winner.

What you can read now

  • The AI agent contest. Chapter 0 defines the strategic contests, prizes, positions, and tests of winning that frame the rest of the book.
  • The human-agency imperative. Chapter 00 examines the voluntary judgment, value, trust, and agency that connect AI availability to repeated use.
  • The framework. Chapters 01 through 04 in Part I define the questions that matter and the conditions under which intelligence, context, authority, execution, and distribution may create durable value.
  • The contenders. Part II examines core company routes and keeps provisional contenders visible as evidence develops. The chapters state what evidence could change each route's status; they do not rank companies or measure their relative strength.
  • The contest map. Chapter 13 in Part II maps the strategic positions being tested across the AI system. It distinguishes core contenders, provisional contenders, and the remaining watch list; it is not a ranking or a claim of ownership.
  • The systems of action. Chapters 01 through 03 in Part III are in research. They examine the runtimes, protocols and governed data access, and controlled action that let agents carry work forward.
  • Assistant products. Chapter 04 in Part III is an interactive research snapshot of AI assistants, with product profiles, filters, a comparison table, and source caveats. Its reported figures and autonomy labels are not comparable measures of performance.
  • Assistant architectures. Chapter 05 in Part III is the technical companion to the landscape. It explains architecture patterns, similar assistant modes, and their tradeoffs, with component evidence and proposed validation methods. It does not rank product performance.
  • Preparing the workforce. Chapters 01 through 04 in Part IV are in research. They ask whether schools, universities, and early working life build the judgment delegated work needs.
  • The battlegrounds. Chapter 01 in Part V is in research. It frames five questions for the next evidence increment: agent capability, safety and control, economics, talent, and government and regulation.
  • The current synthesis. Chapter 01 in Part VI reads the available cross-company evidence without assigning category leaders or an overall winner.
  • The Timeline. A source-linked record of selected historical and current events. Its historical DEFCON treatment is an editorial reconstruction, not a measured time series.

Where the research goes next

Build comparable evidence

The central gap is comparable evidence across contenders. We prioritize clearly defined, current measures of active use, retention, task success, incident and recovery performance, governed deployment, throughput, customer ROI, and workflow outcomes. Revenue, investment, capital spending, and expected ROI are context rather than proof of transformation. A company claim, demonstration, or installed-base figure is not enough on its own.

The Evidence Board is withheld from this edition until it can offer a more useful reader-facing view. A measure enters the record only when it has a defined unit, population, reporting period, comparison, source, and material caveat. Until then, an absence of comparable disclosure remains part of the research record.

A watch-list company becomes provisional when it has a source-backed route to a material strategic prize. A provisional company becomes core only with dated evidence that can assess its position, dependencies, and outcomes. A product launch, model release, or large company scale alone does not meet the core requirement.

Test the contender theses

Each contender chapter will continue to test its strategic argument against independent evidence and counterexamples. The questions are whether embedded agents create repeated use and measurable workflow value; whether context and permissions improve outcomes; which controls earn trust for consequential actions; and where switching costs are durable rather than temporary.

Develop the battlegrounds

The next major research tracks are:

  1. Agent capability: What is repeatable in real work, not just demonstrated?
  2. Safety and control: Which controls are effective in operation, and how do they affect adoption?
  3. Economics: Where do model, tool, orchestration, integration, and human-oversight costs accumulate?
  4. Talent: Which organizational capabilities translate into lasting advantage?
  5. Government and regulation: Which policy developments materially change the competitive conditions?

Examine the possible futures

The book will use evidence to assess five scenarios: platform consolidation, fragmentation, model-layer commoditization, the agent becoming the interface, and continuing human approval and accountability. These are structured possibilities, not predictions. The final assessment will name conditional leaders by category only when its evidence and assumptions can be inspected.

What could change the argument

The working hypothesis is designed to be falsifiable. It would weaken if independent evidence showed that model choice alone consistently determines workflow outcomes, that embedded distribution does not produce repeat use or retention, or that users can readily switch models without rebuilding context, integrations, evaluations, and governance.

It would strengthen if comparable evidence showed that the surrounding system reliably improves outcomes, trust, retention, or economics across meaningful tasks and organizations.

How this book changes

We update the book when new evidence changes the argument, corrects a source, clarifies a measure, or provides a compatible comparison.

Read the Methodology for the standard behind this work and DEFCON for the current competition assessment.

Return to AI Wars of 2026