Everyone wants AI, but nobody wants to clean up their data. Why the path to a "smart" company isn't about buying a chatbot, but about the boring, essential work of modular APIs and finding a single source of truth.
In boardrooms everywhere, the mandate is clear: "We need AI." It’s seen as a magic switch, a software update that will suddenly make the business intelligent, automated, and efficient.
But as an engineer on the ground, I have to be the bearer of bad news. You can't just "add AI" to a mess. AI isn't a magic wand; it’s a consumer of data and a user of tools. If your data is messy and your tools are broken, adding AI will just automate your confusion.
If you want to be AI-ready in six months, stop looking for a "Chief AI Officer" and start looking at your infrastructure. Here are the three non-negotiable prerequisites we need to build first.
1. The Single Source of Truth (Or: Why Your Bot Will Lie)
The first thing people want is a chatbot that "knows everything about the company." But where does that knowledge live right now?
In my experience, it’s a fragmented mess. It’s scattered across outdated Wiki pages, PDF handbooks from 2019, Slack threads, and worst of all, logic that only exists inside the head of a senior developer or deeply buried in legacy code.
If we build an AI on top of this, it will hallucinate. It will confidently tell you the policy from three years ago because that was the only document it could find. Before we can have an intelligent bot, we need a Single Source of Truth. We need to do the unglamorous work of centralising documentation, cleaning data, and ensuring that "what the code does" matches "what the docs say."
2. Modularity: Breaking the Monolith for Agents
The next dream is "AI Agents", bots that can do things, like "refund this user" or "deploy this environment."
But an AI Agent is just a software program that calls functions. It needs a clean, discrete "tool" to pick up and use. We have struggled with "Big Tools", monolithic systems where business logic is trapped inside massive user interfaces or tangled backend spaghetti.
If I can't trigger a refund via a clean, documented API call POST /refund, then an AI Agent can't do it either. If you want AI to take action, you first need to refactor your monoliths into modular, API-first services. You need to build the handles for the AI to grab onto.
The Goal: The "CEO on the Plane" Experience
I was recently told a story that perfectly illustrates why this modularity matters. An observer watched a startup CEO on a flight manage an entire project from his phone. He received a lead via email, and his AI agent analyzed the need, checked his team's skills database, cross-referenced calendars to book a meeting, generated a briefing slide deck, and emailed the invites.
To the observer, it looked like magic. But to an engineer, it looked like APIs. That workflow wasn't possible because the AI was a genius; it was possible because the email, calendar, HR database, and document generation tools all had clean, accessible APIs that the agent could chain together. That is the level of modularity we need to aim for.
3. It’s Not Magic, It’s Math (and Data)
Finally, we need to reset the culture around what AI actually is. I’ve tried to explain at work: we can't just "do AI."
AI is, fundamentally, machine learning. It is a set of mathematical formulas that require structured data to run. It needs inputs to generate outputs. If we want an AI to predict churn, we need clean historical data. If we want it to write code, it needs context.
We need to stop treating AI as a creative entity and start treating it as a statistical engine. This shift in mindset forces us to focus on the inputs. Do we have the data? Is it accessible via an API? Is it structured?
Conclusion: Eat Your Vegetables
The path to a futuristic, AI-driven company is paved with "boring" engineering best practices. It’s about documentation, API design, modularity, and data hygiene.
You can't skip this step. If you try to layer a super-intelligent model on top of a fragile, undocumented infrastructure, you won't get innovation. You'll just get a very expensive, very fast way to break things.