Think Small

We are living through a strange moment in the development of artificial intelligence. On one hand, the demos are dazzling: agents that seem to plan and reason with ease, models that compose original songs from a text prompt, and research tools that produce detailed reports in minutes. On the other hand, many AI teams are stuck in what is often called “prototype purgatory”, the stage where projects that look impressive on paper fail to become reliable systems that create real value for the organization.

The data supports this. Most enterprise GenAI initiatives fail to deliver measurable business impact. The problem is not the power of the models but a learning gap: generic tools that cannot adapt to complex organizational work environments. It is similar to what I have seen in industrial analytics, where the core problem is not the algorithm but the underlying complexity of the environment it has to operate in.

The difficulty grows when you build agent-based AI systems. These systems are often seen as a “black box”. They are hard to debug, and their performance degrades unpredictably when they are integrated with specialized enterprise tools. They lack memory, struggle to generalize, and fail not because the AI “isn’t smart” but because the system around it is fragile. The challenge has shifted from writing prompts to building robust, verifiable systems.

What makes this especially frustrating is the existence of a “shadow AI economy”: employees using personal ChatGPT accounts to get their work done. It shows that demand for AI is real and tangible, but the large, ambitious solutions organizations are building simply do not meet it.

The strategic power of starting small

Given these difficulties, the most effective path forward may be the counterintuitive one. Instead of building broad, complex systems, AI teams should narrow their focus dramatically. In other words: think small. Much smaller.

This is where the idea of a wedge comes in: a deliberately narrow starting point that gives you something solid to build on.

A wedge is a highly focused initial solution that solves one painful, important problem for a single user or a small team, and does it exceptionally well. The idea is to build a standalone tool so useful that one person adopts it immediately, without needing organization-wide approval.

The critical ingredient is not just finding a small problem but finding the right person. Some call them change agents, the “tent-raisers”: influential employees who are empowered to step outside standard procedures to solve problems themselves.

For example, a sales operations manager who spends half a day cleaning lead data, or a customer support manager who manually triages every request. They are the hidden AI economy, already using consumer tools because the official solutions are not good enough. Build for them first.

This approach suits AI especially well because it addresses the root of the trust problem. A wedge solution creates a tight feedback loop with a small group of users, making it possible to build reliability and improve the system in a controlled environment. It solves not only the cold-start problem of a user network but also the problem of building trust in AI systems inside the organization.

From a system that records to a system that acts

AI teams also need to understand the deep shift enterprise software is going through. For decades, the goal was to be the “System of Record”: the central database where critical information lives, such as Salesforce or SAP.

But AI has changed the battlefield. The real prize today is to become a “System of Action”: an intelligence layer that actually does the work, not only storing data but automating the workflow itself.

The most powerful way to build this is a “Data Trojan Horse” strategy. You create an application that delivers immediate value and, in the process, captures a unique stream of proprietary data. This creates a positive growth loop: the tool delivers value, using it generates data, the data trains the AI, and the improved product becomes indispensable. Your moat is built not through a stronger model but through workflow-specific intelligence that accumulates over time.

A concrete example: the “messy inbox” problem.

Every organization has processes that start with a chaotic stream of unstructured information: emails, PDF files, voice messages. An AI tool that automatically decodes, sorts and routes this information creates immediate value.

Once you own this critical upstream step, you earn the right to manage everything that happens after it. You are not competing with an information system; you are tapping into its data stream and becoming the new operational center.