AI agent requirements: why you can't prompt the room

AI agent requirements are the structured business specifications that define what an AI agent should do, for whom, and why - before any code is written. In a controlled study of 21 AI agent prototypes, 80% were abandoned not for technical reasons but because requirements failed to establish business value or data access.

In an era where large language models can generate thousands of lines of functional code in seconds, the technical act of building software has ceased to be the primary constraint for modern organizations. Instead, a new bottleneck has emerged: AI agent requirements. As the cost of generation falls toward zero, the cost of misdirected effort rises. Recent internal research highlights the severity of this shift - in a controlled environment where 21 distinct AI agent ideas were prototyped, 17 were ultimately abandoned before completion. The reason for this 80% failure rate was not technical incapacity, but a fundamental lack of business value or inaccessible data. This reveals a critical truth for operations leaders - you can prompt your AI, but you cannot prompt the room.

While Shadow AI continues to sprawl across departments, with individual employees experimenting with fragmented tools, the organizations that successfully cross the chasm from experiment to ROI are those focusing on the human element of requirements elicitation. The real work is no longer found in the syntax of the code, but in the nuance of the boardroom. To navigate this, leaders must adopt a structured approach to defining what is worth building before a single line of a prompt is written.

The expensive part of AI agent requirements

For the past decade, the software development life cycle (SDLC) was gated by engineering velocity. Companies hired the smartest people to sit behind screens and solve complex architectural puzzles. Today, the economics have inverted. Building has become cheap, but defining what to build has become the most expensive and risky phase of any project. When organizations treat AI implementation as a technical exercise rather than an operational one, they fall into the trap of the "faster horse."

As the famous Henry Ford analogy suggests, if you ask customers what they want, they will ask for a faster version of what they already have. In the context of AI, this leads to "vibe coding" - creating digital versions of existing, inefficient manual processes without questioning the underlying value. Because AI is trained on the average of human knowledge, it defaults to common, generic solutions. To move beyond the average and achieve a magnitude shift in performance, humans must bridge the gap between business intent and technical execution.

This shift requires moving your most strategic thinkers "upstream." Instead of focusing on the output, they must focus on the elicitation of requirements from stakeholders, decision-makers, and end-users. The new moat for a company is not its access to any single AI model - everyone has that - but its ability to deeply understand and map its own internal business logic.

The VAD framework for AI agent requirements: value, architecture, design

To move from fragmented Shadow AI to governed, sovereign systems, organizations need a repeatable methodology. We advocate for the VAD framework - Value, Architecture, and Design. This three-stage process ensures that every agentic system is grounded in operational reality.

Value: identifying the outcome

Before discussing models or integrations, the first step is to quantify the business outcome. This involves answering four fundamental questions for every proposed AI agent:

• Whose problem is this? You must be able to name a specific persona or department lead who is currently feeling the friction.
• What does winning look like? Define the specific outcome - is it a faster response, a safer process, or a more consistent output?
• What would make them refuse to use it? Identifying friction points - such as data security concerns or a cumbersome UI - early prevents the development of "shelfware."
• Does it change a decision? The highest-value agents don't just move data; they tilt the user toward making a better, faster decision.

Architecture: building the sovereign foundation

Once the value is clear, the focus shifts to the underlying architecture. This is where organizations must decide between a flimsy SaaS integration and a sovereign AI agent system. A sovereign approach ensures that the data, the memory, and the reasoning remain within the company's control. The architecture phase defines how the agent interacts with systems of record (like CRM or ERP) and where the persistent state is stored. This stage is critical for passing procurement and security audits, as it moves the agent from a "cool demo" to production-grade infrastructure.

Design: eliciting the user experience

Only after the value and architecture are locked do we move to design. This isn't just about the visual interface; it's about the interaction model. How does the agent communicate? How does it handle errors? By following this sequence, companies avoid the common mistake of designing a solution for a problem that doesn't exist or isn't worth solving.

Why user story mapping is the new AI agent requirements moat

One of the most effective tools for AI agent requirements elicitation is the traditional user story map. While it may seem like a relic of old-school product management, it is actually the most efficient way to package context for an AI. Large language models are highly optimized for pattern recognition, and they have been trained on millions of examples of structured user stories (Persona, Need, Why).

By creating a story map that breaks down a process into its backbone - for example, contacting, triaging, resolving, and closing a support case - you create a coherent roadmap for development. You can then identify which stories belong in a fixed-scope Starter Project (the MVP) and which belong in the long-term backlog.

When these stories are documented in a standard format and stored in a shared repository, they serve as the "source of truth" for the AI. This allows for "daisy chaining" agents together into a coherent system. If you give an AI a generic prompt like "build a support agent," the results will be mediocre. If you provide a markdown file containing a detailed user story map with specific acceptance criteria, the AI can generate a specification that is significantly more accurate and operationally relevant. See how this structured approach helped one team build an autonomous content system that consistently passes quality gates.

Avoiding the demo trap in AI implementation

As companies rush to prove they are "doing AI," many fall into the demo-as-deliverable trap. According to Gartner, 85% of AI projects never reach production - and the root cause is almost always requirements, not technology. It is remarkably easy to build a prototype that looks impressive in a controlled meeting but fails immediately in the messy reality of production. Identifying these anti-patterns is essential for any operations leader:

• The Adoption Gap: If your team is shipping features at a high velocity but the actual usage frequency is low, you are building the wrong thing.
• The Vibe Coding Crisis: This occurs when systems are built based on intuition rather than tested requirements. If a Product Requirement Document (PRD) hasn't been tested against a real user's workflow, it is likely to fail in production.
• Lack of Governance: Shadow AI often results in multiple agents performing the same task in different ways, leading to inconsistent data and security risks.

To combat this, the goal should be to move toward a professional middle ground. This means choosing a Solution-First model rather than getting caught in endless consulting cycles or unmanaged tool sprawl. By starting with a focused Starter Project, organizations can prove value within weeks, ensuring they are building one of the 4 successful agents rather than wasting resources on the 17 that don't make sense.

The strategic shift: moving talent upstream

The long-term implication of the AI shift is a change in how we value human expertise. We no longer need our smartest people to focus solely on the mechanics of code. We need them to become the bridge between business complexity and AI capability. This requires a shift in key performance indicators (KPIs). According to McKinsey, organizations that invest in cross-functional requirements teams see 2.5x faster time-to-value on AI projects. Instead of tracking the number of features shipped, leaders should track the number of features used more than twice.

Involving subject matter experts (SMEs) in the decision-making process is no longer optional. These individuals have the historical context of what has worked and what hasn't. When their expertise is combined with a production-grade operations automation infrastructure, the organization gains an autonomous system that doesn't just automate tasks but creates a durable competitive advantage.

Conclusion

The fundamental challenge of the AI era is not a lack of power, but a lack of direction. The ability to write code is now a commodity, but the ability to read the room - to understand the hidden needs, the political constraints, and the true value drivers of an organization - remains a uniquely human skill. By adopting frameworks like VAD and tools like user story mapping, operations leaders can stop the cycle of failed experiments. The path to a successful sovereign AI agent system begins with the recognition that while you can't prompt the room, you can lead it toward the right requirements.