In 2000, the music industry had a problem it misdiagnosed. Napster was stealing albums. The fix was obvious: stop the theft, protect the bundle. The album — twelve songs, a cover, a price — had structured the industry for fifty years. Labels, distributors, retailers, and artists all organized themselves around it.
By 2003, iTunes had unbundled the album into individual tracks at $0.99 each. By 2010, Spotify had unbundled ownership into access. Piracy was the symptom; the bundle had become the fiction — a transaction artifact that survived long past the technological conditions that made it necessary. The industry didn't adapt to streaming. It was restructured by it.
The same unbundling is coming for companies. Not just for their products. For their internal architecture.
The Bundle We Call a Department
Organizations are bundles. A marketing department bundles brand strategy, copywriting, campaign management, media buying, and analytics — not because those activities are naturally related, but because coordinating them across separate entities used to be prohibitively expensive. A finance department bundles bookkeeping, forecasting, compliance, and treasury for the same reason. Communication costs made consolidation rational.
This is the insight that a paper titled "The Agentic Economy", captures for consumer markets — and almost entirely fails to apply to the interior of companies. The authors argue convincingly that AI agents will reduce communication frictions between consumers and businesses, eroding the role of two-sided platform intermediaries like Amazon and Expedia. What they don't say, but what follows directly: the same logic applies inside organizations. Departments are two-sided platforms. They exist to intermediate between a need and a capability. When that intermediation becomes cheap, the bundle becomes optional.
An assistant agent can pull real-time performance data, draft a campaign brief, generate copy variants, place a media buy, and measure attribution — crossing the boundaries of four traditional departments without a single handoff. The question isn't whether AI will automate tasks within those departments. The question is whether those departments survive as organizing units at all.
To be precise about what "optional" means here: departments won't simply vanish. They exist for reasons that outlast coordination costs — they provide stable accountability for audits and liability, they're how headcount and budgets get allocated, they enforce data access boundaries, and they house career ladders and performance evaluation. Those forces are durable.
What changes is the execution layer. The more plausible near-term outcome isn't that departments disappear; it's that they persist as governance and accountability shells while the actual work becomes cross-functional, agent-mediated, and composable across those shells. The org chart stays. The work stops respecting it — unless governance, audit, and access controls are rebuilt around agent workflows.
Unbundling Has Already Begun — Informally
MIT's Project NANDA, in its "State of AI in Business 2025" report, named what's already happening: a "shadow AI economy." While only 40% of companies have purchased official LLM subscriptions, workers at more than 90% of companies surveyed report regular use of personal AI tools for work tasks — ChatGPT, Claude, Copilot — most of it without IT approval. A separate 2025 survey by WalkMe found that 78% of employees use AI tools not provided by their employer. Two studies, conducted independently, pointing to the same structural reality.
The pattern is telling. A corporate lawyer at a mid-sized firm, whose organization had invested $50,000 in a specialized contract analysis tool, consistently defaulted to a personal ChatGPT subscription — because it let her guide the conversation and iterate across the full scope of her work. A $20-per-month consumer tool outperforming a purpose-built enterprise system, not because the enterprise tool is technically worse, but because the lawyer's actual work crosses the boundaries the enterprise tool wasn't designed to handle. The unofficial tool traverses legal, procurement, and operations in a single session. The official one stays in its lane.
This is where McDonald's drive-through failure becomes instructive in a different way than it's usually told. The documented facts: McDonald's discontinued its IBM-powered AI ordering pilot in late July 2024, after roughly three years and tests across more than 100 restaurants. BTIG analyst coverage reported accuracy remaining in the low-to-mid 80% range, well below the approximately 95% threshold needed to be competitive with human workers. What actually killed it is a matter of interpretation — but the structural pattern that emerges is consistent: the AI system sat at the intersection of operations, HR, IT, legal, and franchisee relations, and no single function owned that intersection. When it broke, nobody had clear authority to fix it. The org chart assumed those boundaries would never need to overlap. The AI fell into the gap.
As agents become more capable, more gaps will appear — not because companies will deploy AI badly, but because capable AI makes existing organizational seams visible in ways that human workers, moving more slowly and deferring across those seams, never did.
Rebundling: The New Organizational Logic — And Who Controls It
Unbundling is not the destination. It's the transition. What follows is rebundling — and the rebundling logic will not resemble the organizational chart companies inherited from the industrial era.
The Microsoft Research authors describe a vivid consumer example: an assistant agent that knows a user's reading history can collaborate with a New York Times service agent to generate a customized article containing only what the reader doesn't already know. The bundle (the article) gets reconstructed around the individual's actual need rather than the producer's production convenience.
The same logic applies to internal work products. A quarterly business review isn't a natural unit of information. It's a coordination artifact — a bundle of data, analysis, narrative, and decision-making that exists because getting the right people in the right room at the right time was expensive. An agentic infrastructure can assemble relevant synthesis continuously, surfacing the right information to the right decision-maker at the moment of decision.
In practice, this will be resisted — and it's worth being clear-eyed about why. Executives like rituals. The QBR isn't just an information-delivery mechanism; it's a performance, a political event, a moment where accountability gets assigned and credit gets distributed. Continuous synthesis maps poorly onto those incentives unless accountability structures are deliberately redesigned around it. The organizations that rebundle successfully won't just deploy agents against the coordination cost problem. They'll redesign the political and governance structures that made the old bundle necessary in the first place.
The rebundled company organizes around outcomes rather than functions. Consider a revenue pod: a product manager, a marketer, an account executive, and a customer success lead — supported by agents handling competitive research, collateral generation, outreach sequencing, CRM updates, and billing exception handling. The humans set intent and make judgment calls. The agents handle the coordination overhead that previously required those functions to maintain separate staffs and separate fiefdoms. The structurally coherent early adopters building this aren't making a technology bet. They're making an organizational bet — that accountability structures can be redesigned before the coordination layer collapses under its own weight.
But here's the part that gets underplayed in most of this analysis: rebundling around outcomes assumes companies own their agentic infrastructure. Many won't. If the agent layer is delivered by Microsoft (Copilot + Azure + Entra ID), Google (Workspace + Gemini + Cloud), or Apple (Intelligence + device + enterprise MDM), then "rebundling around outcomes" happens inside a vendor's control plane. Identity sits with the vendor. That's not incidental — it means authorization policies, tool-routing decisions, audit log ownership, and eventually marketplace economics all flow from the platform provider, not from the company running the pods. The same walled-garden-versus-web-of-agents question the Microsoft Research paper poses for consumer markets applies directly to enterprise org design. Companies that treat agentic infrastructure as something they simply rent, without understanding what that vendor relationship authorizes and constrains, will find that their "outcome-oriented pods" are executing inside someone else's governance framework — subject to someone else's fees, someone else's model choices, and someone else's definitions of what agents are permitted to do.
Org design is downstream of platform control.
New Divisions — And the Power Shifts That Come With Them
The structural disruption of the agentic economy falls on specific internal functions — and not the ones that typically dominate AI transformation conversations.
Procurement becomes the runtime governor of enterprise spend. In an agent-to-agent vendor world, procurement stops being the RFP department and becomes something closer to product management for constraints. The new mechanics are: policy-as-code defining what agents may purchase, when, and from whom; data-sharing contracts governing what context an agent can reveal to a vendor's service agent; and evaluation harnesses for benchmarking vendor agents before authorizing them. Underlying all of it is continuous renegotiation — usage-based billing, dynamic pricing, and agent transaction volumes that fluctuate in ways annual contracts were never designed to handle. The teams that understand how to govern those relationships, not just negotiate initial terms, will hold organizational power that procurement has never previously had.
Legal and compliance become architectural functions. The EU AI Act requires human-in-the-loop accountability for consequential AI decisions in high-risk systems. As agents assume more execution responsibility, someone has to design the oversight topology in advance — defining which agent actions require human sign-off, what audit trails must be maintained, where liability attaches — not just respond to failures after the fact. Legal teams that can think structurally about AI governance will shape what agents are permitted to do. Those that can't will spend their time cleaning up what agents already did.
The turf war is coming for a new function that doesn't have a clean name yet. Call it Agent Operations, or Agentic Infrastructure. The mandate: design and govern the network of agents that execute the company's work. What makes it distinct is the specific combination of powers required to do it effectively — identity and permission management (what each agent is authorized to access and do), audit log ownership (the authoritative record of what agents actually did), tool access governance (which external agents and APIs internal agents can call), and sandboxing and escalation rules (what triggers human review). That combination currently exists in fragments across Platform Engineering, Security, MLOps, and FinOps — each of which will claim jurisdiction. The function that wins it will have significant leverage over how the company actually operates, because it will govern the execution layer that all the accountability shells above it depend on.
The Right Question — Again
The question that follows from McDonald's, from the shadow AI economy data, from the EU AI Act requirements, is always the same one: not what will AI do to jobs, but what does your organizational structure assume that is no longer true?
The agentic economy makes one assumption obsolete above all others: that communication costs make bundling rational. For fifty years, you assembled departments, hierarchies, and coordination rituals because the alternative — dynamically assembling the right capabilities for each task — was too expensive to manage.
It isn't anymore.
The music industry learned this the hard way. It spent a decade protecting the album. The companies that built durable positions in the streaming era didn't protect the old bundle. They designed new ones around what users actually wanted — and around what the underlying infrastructure actually made possible.
Your departments are albums. The question is whether you redesign the bundle — or wait for someone else to do it around you.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.
