Building Enterprise AI That Actually Works

AI is often compared to the discovery of fire, yet most enterprise AI efforts stall before delivering real value. A common mistake is treating today’s AI as if we have already reached Artificial General Intelligence (AGI), meaning systems capable of autonomously running end-to-end workflows. In narrowly verifiable domains like coding, this is already close. In most others, it is not.

The pace of progress makes this hard to reason about. Models improve quickly, capabilities shift, and it can feel like building on moving ground. As Jeff Bezos has argued in other technological shifts, the more useful question is not what will change, but what will not.

Applied to AI today, that stability lies in its limits. In 2026, and for the medium term, AI can reason, explore, and execute, but it still lacks intent, judgment, and accountability due to technical, organizational, and legal limits. As a result, AI can significantly augment business processes but cannot reliably own outcomes.

This post proposes a model for the AI-augmented organization built around those limits: AI-native workflows combining embedded, task-specific agents with explicit human gates, creating infrastructure that strengthens as models improve rather than requiring constant redesign.

I. The Limits of AI

The prevailing mental model assumes an AI agent can be given a goal and autonomously execute it end to end. We may frame goals broadly, such as “launch a new product”, “design and run a marketing campaign”, “increase revenue in a specific segment”, and sometimes more narrowly, like “improve retention”.

At current levels of AI readiness, this agentic model breaks down in real business contexts, even for seemingly simple goals. The limitation is not intelligence, execution speed or even output quality. It is judgment. A goal is not a single instruction. It contains multiple, often conflicting paths.

Take “reduce shopping cart dropouts.” Possible approaches include:

• Remove upsell steps — hurts short-term revenue, improves conversion
• Redesign UX — medium effort, uncertain impact
• Add trust signals — low effort, modest gains
• Rebuild checkout architecture — 6 months of engineering, high risk

Choosing among these is not a data problem. It is a tradeoff decision involving cost versus impact, short term versus long term outcomes, brand philosophy versus revenue, and opportunity cost.

AI can enumerate options and simulate outcomes, often more comprehensively and far faster than humans. But it cannot legitimately decide which tradeoff is acceptable. Judgment requires context, experience, and accountability.

The source of confusion

The reason behind the inflation/confusion of what an AI agent can do today is that people collapse three distinct activities into one and call it an agent:

• Intent — deciding what we are trying to do
• Judgment — deciding which tradeoffs we accept
• Execution — making it real

Each of these requires different capabilities. AI can only support parts of each. Here is what AI can and cannot do at each layer:

• Intent
  • AI can: generate ideas, reframe problems, expand the option space
  • AI cannot: decide what matters, set direction, define success
• Judgment
  • AI can: compare options, surface risks, simulate consequences
  • AI cannot: prioritize, resolve value conflicts, make ethical calls
• Execution
  • AI can: generate plans, content, analysis, and assets at scale
  • AI cannot: own quality thresholds, brand meaning, or liability

II. Resulting Org Design

To respect these constraints, an AI-augmented organization must be built on a small set of core principles.

• Humans retain ownership of intent, judgment, and accountability
• AI is embedded where it is strongest: exploration, generation, and execution
• Transitions between stages are explicitly human-gated

In practice, this leads to AI-augmented workflows that are human-driven and selectively enhanced by AI. Now, any business problem follows the same core workflow: identify, explore, decide, execute, and learn. Assigning each stage to humans or AI yields a simple operating model:

• Identify & prioritize problems — human-led, AI-informed
• Explore solution space — AI-led, human-filtered
• Decide & commit — human-only authority
• Execute — AI-led, human-reviewed
• Measure & learn — AI-analyzed, human-interpreted

Beyond the core workflow, organizations rely on auxiliary functions such as legal, finance, brand, PR, security, hiring, and ethics. These follow the same principle: AI prepares, drafts, analyzes, and monitors; humans decide, sign off, and accept accountability.

III. Why This Improves Quality

Instead of general-purpose AI agents, this model uses embedded agents that are narrow and task-specific. This matters because general agents produce generic output, and in the worst cases, the "AI slop" we're becoming too familiar with, while specialized agents produce usable work.

Three points for effective agent design

The reason behind this is that AI agents are similar to humans in certain regards. The more specialized they are, the more depth and quality they will give. They need to know where to look, how to think, and what to communicate. Regardless of domain, an effective AI agent shares the same design principles:

• Scoped, relevant knowledge and data sources: Like a human, an AI that knows where to look, is likely to find more accurate results
• Task-appropriate LLM models: Some models/settings excel at creativity, others at specificity; use the correct model for each node
• Structured, editable outputs: structured outputs where possible (e.g. reporting) ensure stability, and easy editability is critical to control quality

The result: fewer hallucinations, higher consistency, and outputs closer to production-ready.

Example: AI augmented Marketing

Consider a marketing campaign workflow. It is a single end-to-end process, but each stage requires a different type of cognition:

• Exploration. A reasoning-heavy agent evaluates problems and generates strategic options. A marketing lead reviews them and selects directions aligned with brand and business goals
• Concept creation. A creative agent develops positioning, messaging angles, and campaign concepts. A marketing lead selects one direction to move forward
• Content creation. A generative agent produces channel-specific assets using brand voice and best practices. Marketing lead reviews, edits, and approves for publication
• Measurement. A reasoning-heavy agent analyzes performance and surfaces insights. Leadership interprets the results and decides on optimizations or pivots

The capability lives in the workflow, not the agent. Marketing is not automated end to end; it is augmented through specialized agents paired with explicit human judgment.

IV. Implementation

No single central team can (or should) build every AI-assisted workflow and agent across an organization. That doesn't scale, and it recreates bottlenecks. Instead, implementation must follow a federated model with a central owner.

The Central Owner

The central owner is responsible for structure, quality, and consistency, not for building agents. Its role is to define core workflows and human decision gates, establish shared design principles, provide reusable templates and patterns, set quality, safety, and governance standards, and train teams to build agents correctly.

The central owner also acts as a bar-raiser. This may slow things down initially, but in early stages consistency matters more than speed. All locally built agents are reviewed against a clear standard: correct scope, appropriate model choice, proper data grounding and safety, structured and editable outputs, and explicitly defined human judgment gates. Critically, the central team enforces how agents are built, not what they do.

The Individual Teams

Individual teams such as marketing, product, legal, and finance build their own agents. They are best positioned to do so because they understand the domain context, own the outcomes, and know the real failure modes.

To avoid generic or low-quality agents, teams build within the shared structure rather than ad hoc, defining the agent’s scope, required human inputs, knowledge and data sources, model type, output format, and review points.

Agent Number Matters

The goal is not to maximize the number of agents, but to define the right number, each covering a stable, meaningful unit of work. Agents that are too small create fragmentation and overhead. Agents that are too large recreate the general-agent problem. Strong agents map to real business artifacts, repeatable decisions, and known constraints.

V. Conclusion

In this model, AI is treated as a variable rather than a pillar. The structure does not depend on any single model being capable enough. As models improve, the workflows benefit naturally: options become stronger, reasoning improves, and output quality rises, without the organization needing to constantly redesign how it works.

The open question is not whether models will improve. They will. But how will legal and organizational frameworks for AI intent, judgment, and accountability evolve? How much agency/autonomy are we comfortable giving AI? Which constraints gradually loosen, and which remain fundamental? The road to AGI runs through these questions, and how we answer them will define the next decade of organizational design.

If this was useful, sign up here for the free newsletter to stay updated on new posts.