Enterprise AI needs more than agents to reach the bottom line

McKinsey’s 2026 survey found that 80% of respondents reported improved personal productivity from AI, but only 37% reported a positive contribution to enterprise earnings before interest and taxes (EBIT).

A professor of innovation begins each AI session with a fragment of the first page of “Attention Is All You Need,” the 2017 paper written by eight Google researchers. The fragment avoids a full-paper reading while preserving the reference that made the session’s central point unavoidable.

The paper did more than improve an existing design. Its Transformer architecture reorganized computation around self-attention, removed recurrence and made parallel training easier. That architectural break later supported GPT and much of the generative-AI wave. “The rest, as I said, is history,” the professor writes.

Google understood the technology but underestimated the category it could create. It incorporated the architecture into Translate and Search, yet failed to build the consumer product that defined the next era.

OpenAI, founded in 2015, two years before the paper, saw the broader opportunity and built GPT, then ChatGPT. All eight paper authors eventually left Google: seven became founders, and one joined OpenAI.

This is the author’s definition of a “Transformer moment”: an organizing principle whose eventual importance greatly exceeds its first application. The analogy is not about another model release. It is about discovering a system-level category, then failing or succeeding to recognize its wider reach.

Why do productivity gains not translate into EBIT?

An individual, human or artificial, can complete a task. A company must also preserve state, coordinate specialized capabilities, enforce authority, remember outcomes and adapt over time. Transformers advanced model- and task-level performance, but those capabilities do not supply those company-wide mechanisms.

McKinsey reported no significant change from the previous year in the enterprise EBIT result. Its analysis does identify the high performers: companies that redesign workflows rather than insert AI into existing ones. The survey therefore separates output produced by an individual worker from performance produced by an operating system for the whole company.

Can coordinating agents create organizational intelligence?

The industry’s current starting point is the agent. Anthropic explicitly starts with the “augmented LLM,” a model supplied with retrieval, tools and memory. Microsoft presents sequential, concurrent, group-chat, handoff and other multi-agent patterns. These techniques connect workers, but they leave the larger organization implicit. That distinction matters: coordination is not the same as organizational intelligence.

“The agent should not be the organizing unit,” the professor argues. Instead, a persistent, goal-directed system should own the business representation, operational state, objectives, constraints, authority and accumulated experience. Models and agents can contribute reasoning and execution; they should not own those layers.

The growing catalog of patterns exposes the missing foundation. Each implementation reconstructs identity, permissions, state, policy, evidence, recovery and learning because no common foundation exists.

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 because of cost, unclear value or inadequate risk controls. Its advice is to measure enterprise productivity as a whole, not only augmented tasks.

What must sit between an agent and the business?

A workable architecture needs specific interfaces between reasoning and organizational action:

• Persistent organizational state that survives a session.
• Formal objects, relationships and permissible actions.
• Explicit objectives and trade-offs.
• Deterministic authorization outside the model.
• Ephemeral agents and credentials.
• Escalation mechanisms.
• An append-only record linking intent, action and outcome.

NIST guidance supports that separation between reasoning and policy, deterministic access decisions and immutable evidence. Microsoft’s reference architecture divides the agent runtime, policy layer, tools, systems of record and audit pipeline.

The interface between reasoning and action therefore enforces who may act, on which record, under which policy, and provides proof afterward.

What would prove a genuine architectural breakthrough?

The test should not be an agent beating another agent on a benchmark. One design should make coordination, governance, memory, replaceability, accountability and continuous learning easier to manage together. Success comes from removing repeated infrastructure and making behavior inspectable, not from adding another autonomous role.

Google also deployed a Transformer descendant, BERT, but OpenAI captured the defining consumer moment. Clayton Christensen’s “The Innovator’s Dilemma” helps explain the hesitation: LLMs were expensive to run for web search, often answered so completely that users would not click ads, and sometimes hallucinated, threatening trust Google had spent decades building.

Alphabet’s hurried Bard response produced errors and roughly $100 billion in lost market value in February 2023. Google has improved since then and outperformed rivals on several occasions, but the episode shows how research leadership can fail to become product or category leadership.

Earlier, Larry Page and Sergey Brin tried to license PageRank to leading internet companies. Yahoo! declined even to meet; Excite doubted its business value and declined to license it. Both continued with what they knew until it stopped working. Today’s AI leaders face the same choice: treat organizational intelligence as a category, or dismiss narrow implementations and leave another company free to define it. Centering the LLM or agent risks optimizing the previous paradigm instead.

The enterprise AI race may ultimately reward whoever supplies the organizing principle for intelligence at organizational scale, and that company need not be today’s model leader.