Could AI boost every team's output individually, yet still leave a company poorer overall because those improvements no longer fit together?

This explores whether AI's value inside firms is limited mainly by coordination rather than capability. The corpus mostly says yes, though for an unexpected reason. AI doesn't just fail to fix coordination problems. It creates new ones. A formal economic model in Does AI adoption require rethinking how organizations coordinate? shows that when AI improves individual tasks, decisions across the organization become more varied. Each team gets better locally while drifting apart overall. The gains from AI only add up when the organization communicates and restructures so that interdependent decisions still fit together. Bolting AI onto one task at a time can leave much of the value unclaimed.

That changes how to read the familiar complaint that AI 'isn't delivering ROI.' Does easier tool-building actually solve enterprise adoption problems? argues that making AI tools easier to build misses the actual bottleneck. Most workers don't see their own work as automatable, and real adoption requires decisions that cross departments and budget cycles. The same pattern appears in evaluation. Why do agent benchmarks not predict real economic value? finds that agents ace benchmark contests but stumble on long, real professional workflows, which are exactly the tasks that involve handoffs, context, and coordination. Benchmarks measure solo performance. Firms run on joint performance.

One idea worth taking from this is that 'communication' may be the wrong word for what AI needs. Does AI really communicate or just distribute information? argues that AI distributes information but doesn't communicate in the human sense: no speaker taking responsibility, no mutual understanding. Several notes suggest the fix is structure, not more conversation. In MetaGPT (Does structured artifact sharing outperform conversational coordination?), agents that exchange standardized documents, like specs and design files, coordinate better than agents that chat. That mirrors how effective workplaces run on shared artifacts, not meetings. In the other direction, Can AI agents communicate efficiently in joint decision problems? shows that today's LLMs are poor partners in joint decisions. They answer what they're asked but don't decide on their own what to share, what to ask, or what to infer.

As agents start buying, deploying, and transacting on a firm's behalf, coordination becomes a matter of infrastructure and accountability. Does agent capability matter more than coordination infrastructure? argues that the bottleneck shifts from smarter models to identity, delegation, and audit trails. Can traditional governance tools actually control AI agents? explains why: the classic tools for managing human agents, such as incentives, monitoring, and legal enforcement, break down when the agent acts fast and can't explain its choices. Survey data in Does agentic AI fit the tool-or-coworker framework? shows executives already treat agentic AI as a coworker. That leaves firms managing something that is both an asset and a member of staff, and neither the IT department's playbook nor HR's was built for it.

So coordination looks necessary for capturing AI's value but not sufficient on its own. The corpus suggests firms need to redesign three things together: how decisions connect across teams, what structured artifacts humans and agents share, and who is accountable when an agent acts. One limit: these notes argue the case mostly through models, frameworks, and surveys. The collection has little firm-level evidence measuring how much value better coordination actually unlocks.