AI is not the problem

The AI craze is misunderstood.

I’ve spent the last year talking to enterprise stakeholders as an AI product lead. It goes without saying that when AI burst onto the scene, it took many executives by surprise. And despite investing heavily in machine learning, they've hit major roadblocks. The resounding message? Companies are not held back by AI, but by missing infrastructure. Here are four key barriers companies need to overcome to unlock the future value of AI:

1. The Hidden Data

Enterprise players are realizing the true value goes beyond surface-level generative apps. They need models to understand their use case, which is unique and complex for every company and customer. To provide the promised value, AI needs the proper data infrastructure. But up to this point, data capture and processing systems weren't built for training models—they were designed to give value in specific analytic and business use cases.

So while executives demand ROI on resources spent, engineers struggle to build critical data streams that currently lie out of reach. Enterprise workflows and legacy systems are lagging behind, and the old "if it’s not broke, don’t fix it" mentality is no longer an excuse. Logs, dumps, metadata, and system statistics need to be funneled, processed, and even simulated to properly train models. In order to give AI the context it needs, companies need to compile comprehensive data that their use case requires.

Lacking the proper data infrastructure will continue to be a pain for those building AI solutions. Accelerating that process is key to unlocking the massive value that was promised.

1. Building Context

Companies are heavily investing in data pipelines to funnel insights into machine learning. However, having the right data and a good model isn’t enough. Often, the raw data stream isn't in a digestible format and needs contextualization. This “context infrastructure” might break down a block of code, define a schema, or include additional context through Retrieval-Augmented Generation (RAG). Let me share a story that shows how these issues run much deeper than previously thought.

A couple of months ago, Braydon Jones and I tried to build Harmony: an integrated design tool for existing SaaS apps that enables designers to make UI edits and ship the code to Github. To start, we loaded a test React app into a simple design suite and tried to make edits to the CSS. However, as we instructed AI to make the updates, it always missed the mark. There wasn't enough context for the language model to decipher where to change the code. With very little progress, we quickly realized we weren't facing an AI problem; it was a contextualization problem.

Putting AI aside, we built Harmony’s revolutionary tech—an engine that could deconstruct and contextualize code, building a dynamic map that understands an existing codebase. When a user selected a UI element, Harmony could identify the correct place to make an update. With this, it was surprisingly easy to build our product. But it got us thinking: what other AI solutions are missing the mark simply because they lack the technology to provide enough context?

1. Defining Success

To train a model, there must be an ideal outcome. However, whenever human intelligence is involved, there's some level of subjectivity. Decisions that come from making a judgment call can't be easily defined by throwing an error. How do you define and teach an AI what a successful outcome is? This could be different for every use case, customer, and product. Here’s an example:

I was working with an enterprise product that was bringing in all sorts of metadata. However, there were no “error” states that we could teach the model. Instead, it was the irregularities within that data set. When a data point began to increase at an irregular pace, that was when we wanted to sound the alarm. So, we set statistical thresholds manually that helped us understand the state of the data stream. That way, we could train the AI on what to watch for.

Defining these success metrics and building the infrastructure to measure them will take time. Companies are currently in the process of deciding what success will be the easiest to define, as resources and time are still scarce in a fast-growing competitive scene.

1. Empowering decision-making with automation

For AI to have more power in decision-making, it needs to perform tasks in reaction to those decisions. However, with the majority of decisions still being implemented by human intelligence, AI needs access to these actions through automation. Whether built out through APIs or internal tools, these automation frameworks were not optimized for models to plug into and are critical for accelerating the value of AI.

One way to understand what is needed for automation to have AI complete a certain use case is to ask, “If a human was given these endpoints, could they complete the task?” This way, an model making decisions has access to everything it needs.

Again, these automation frameworks will be highly unique and specialized for every company. Learning how to properly give access to performing these actions through automation will be critical in giving AI the power to act alongside its decision-making.

AI is not the problem. The challenge lies in building the right infrastructure to support and contextualize AI's capabilities. Once we overcome this hurdle, the true potential of AI will be unlocked, providing immense value to businesses and, ultimately, transforming the way we work.