Success with Agentic AI: The Different Requirements

There is a lot of excitement around Agentic AI at the moment. While agentic sounds a bit intimidating, it is rather straightforward—a program that can autonomously make decisions and execute actions. For example, an agentic system might schedule meetings by understanding calendar conflicts, negotiate appointment times with other participants, and handle rescheduling without human intervention. These capabilities represent a significant AI evolution.

Technology teams have jumped on the hype train with gusto, both due to the perceived potential—and the fact that it represents the latest shiny tool. However, more than ever before, implementing agentic AI successfully means taking a fundamentally different approach to requirements gathering and system design.

Traditionally, data, digital, and technology teams start by documenting requirements, identifying individual metrics, features, and specifications. Clarity of documentation here is key, given the rigorous nature of development processes. Requirements are typically broken down into discrete user stories, each addressing a specific capability, with measureable success criteria.

However, in this new paradigm, given both the breadth of agentic capabilities and their probabilistic nature—read: there is more than one way to skin a cat—such a restrictive approach can leave a lot of value untapped. Traditional requirements gathering faces several challenges when applied to agentic AI:

Feature Fragmentation: Breaking down capabilities into isolated user stories fails to capture the integrated, end-to-end nature of agentic systems.

Prescriptive Limitations: Specifying exact behaviours constrains the AI's ability to discover novel, potentially superior approaches to solving problems.

Outcome Blindness: Focusing on features rather than outcomes risks building technically compliant systems that fail to address underlying business needs.

Missed Integration Opportunities: Siloed requirements miss the synergies between different aspects of the business that agentic AI could bridge.

Instead, taking a "day in the life" approach can be far more rewarding. Walking a mile in the customer's shoes to understand the holistic journey can yield far richer insights and give a much more comprehensive view into requirements.

For example, whereas capturing specific requirements can help solve the proximate issue—e.g. "as an onboarding advisor I need to see a credit score"—this misses the opportunity to solve the ultimate issue—e.g. "as an onboarding advisor I want to filter out fraudulent applications." In practice, this might involve:

• Shadowing users in their natural work environment to observe their actions, challenges, and workarounds (what they do, not what they say they do)
• Focusing on mapping end-to-end workflows rather than isolated tasks
• Understanding context and constraints that might not surface in traditional requirements sessions—including implicit/unconscious decisions

Taking too narrow a lens also misses the opportunity to use agentic AI's powerful multi-agent capabilities. In the previous example, this might involve spinning up an agent to trawl social media for any relevant signals and feeding that into the master risk agent, further enhancing the latter's predictive power. These systems are particularly useful in automating complex workflows that span multiple business domains, given the need for agents with different skills.

All in all, agentic AI provides a rich suite of capabilities and business opportunities. However, these can only be realised by taking a more comprehensive, empathetic, and system-level approach to capturing customer needs. Given the maturity of the average business partnering and relationship management function, there is little doubt that those organisations that manage to crack this will gain a commanding edge over their competition. The key is to resist the temptation to apply traditional requirements-gathering approaches.

As we continue to explore the frontier of what is possible with agentic AI, our methodologies must evolve alongside the technology itself. The organisations that succeed will be those that reimagine not just what technology can do, but how we define what we want it to accomplish in the first place.

— Ryan

Cover image by DALL·E