The promise of AI revolutionizing the modern workplace is a rather seductive one. You feed it your data, find patterns that might have been missed, and optimize your decisions based on said findings. You might even find the ability to automate those processes, freeing up valuable man-hours for other work. For the most part, this has been a rousing success, as boardrooms, C-suite executives, and startups are buying into it. This isn’t the key to AI success, however, as we’ve seen some organizations struggling to see a return on their investments.

The key to implementing any new technology, like AI, is process stability. Without it firmly in place, you’re not building stable processes or intelligence at your company. Instead, you’re simply industrializing and automating waste, with all the drawbacks that brings.

Unchanging Principles

One of the oldest principles in computing is the phrase, “garbage in, garbage out.” This isn’t an outright condemnation of organizations looking to optimize processes, smooth out workflows, and ultimately find more efficient means of capitalizing on the latest technology. Instead, it simply states the obvious: that your inputs matter. Without sanitized, vetted inputs, your outputs are going to amplify the underlying faults and flaws. Now, as we’ve stated, this isn’t the path forward when it comes to AI success. Despite the name, artificial intelligence isn’t able to think or reason like a human being, and you require a deft hand to make the most of the technology.

If anything, the previously stated principle is only further exacerbated by the use of machine learning and artificial intelligence. A traditional software system might give tell-tale signs when something fails. You can trace it through logs and other outputs. By comparison, an artificial intelligence model’s failure isn’t going to be readily evident. When you have an AI agent that is looking for patterns in data, it will find them, even if that data is garbage. Patterns will be constructed, and confident projections assembled based on those inputs.

Now, ideally, this isn’t a bad thing. If you’re giving it good inputs, you’ll get good outputs. When you’re just feeding it data without stopping to consider what’s happening, that’s where problems arise. You won’t see the issue right away. The patterns, predictions, and optimizations the model develops will bear the burden of your unstable processes. This isn’t a flaw in the algorithm, but calls to mind another acronym used in IT, PEBKAC, or Problem Exists Between Keyboard and Chair.

What Unstable Processes Look Like

Process instability isn’t inherently defined by sloppy data, although that will play a part in it. There is an entire ecosystem of inconsistency that accumulates in any organization over time. Workflows vary from department to department, definitions drift quarterly, and handoffs start making themselves known for the sake of compliance without any consideration as to what they should be doing.

Consider a manufacturer that is looking to build a predictive maintenance model. The overall goal is to curtail equipment failures before they occur, with the added benefit of reducing downtimes and repair costs. On the surface, this is an ambitious and legitimate use of machine learning, and possibly a true indicator of AI success. However, problems start arising when you look at how things are logged. In the aftermath of this implementation, a team starts seeing maintenance events logged across multiple shifts.

One shift might see the symptom, another recording the root cause, and a third implementing some sort of remediation. All three might occur over the same shift, multiple times. The timestamps are inconsistent, as there was no documentation developed. Quality begins to deviate, timekeeping isn’t standardized, and the rework needed is costing just as much as the maintenance. There was no plan in place, and the deliverables suffered because of it.

How AI Amplifies It

So, how does process instability hinder AI success? There are a few things we need to knock out before fully addressing that.

Firstly, AI scales decisions. A human being can make mistakes, and often does, even multiple times over a workday. You have factors that play into this, like throughput, mental exhaustion, and so forth. There are only so many of those mistakes made per day. What happens if you take those same mistakes, perhaps the fault of bad data, and perform them thousands or even millions of times? That level of volume isn’t just a manageable, accounted-for inconsistency, but a systemic failure on a massive scale.

Next, AI obscures causation. When a human being makes a mistake, there’s at least some reasoning behind it. When a model makes a bad decision, the answer is buried deep within hundreds of features spread across billions of parameters. An analyst saying, “The model said so,” isn’t a satisfactory answer when things start failing, and you start seeing damaged relationships between suppliers, customers, or even stakeholders.

Finally, AI creates feedback loops. A deployed model is going to influence its own training data based on the inputs fed to it. This quickly becomes a problem, as businesses confidently employ AI across the board. When you’re told to trust the machine, how well can you trust it if it’s being fed garbage data? Unstable process inputs compound this problem over time, while informing the structural bias of the model, which is going to be a major thorn in any organization’s side.

Stability Isn’t Perfection

Process stability isn’t about perfection, simply put. Perfect is the enemy of good, and that is going to apply to your processes. Stable processes will have errors. That’s just part and parcel of doing any sort of work these days. However, these errors are randomized, and they’re taken into consideration. A stable process is going to be consistent above all else, with its parameters and boundaries clearly defined. When failure occurs, and it will, then said failures are clearly documented and adjusted for.

As such, the practical question isn’t about the cleanliness of your inputs, but whether your processes are consistent enough. If you find yourself answering no to this. If a change in personnel, undocumented workarounds, or even the introduction of new technologies serves to wreak havoc on your workflows, then you’re not quite ready for AI success.

This is admittedly less exciting than embracing new technologies. You’ll be conducting workshops, documenting processes, and learning how to optimize things like handoffs and cross-functional functionality. That sort of organizational attention isn’t going to excite shareholders or your C-suite executives, but it is fundamental for the health of any organization. Make no mistake, this sort of work isn’t a detour from implementing AI en masse at your organization, but rather the foundation for any work done in the future.

Plenty of organizations have implemented AI and found some degree of success with it. Whether you’re looking at healthcare, manufacturing, logistics, or any number of fields, they share a common thread. That thread is going to be a heavy investment into operational consistency across the board. They didn’t start asking algorithms where to go next, but instead took the time to learn from their daily operations.

Conclusion

My hope for this piece isn’t that it’s leaving you discouraged about implementing AI. I think the technology is here to stay, and we’re only seeing the tip of the iceberg when it comes to successful implementations. As such, your organization’s success with AI implementation isn’t going to be a matter of affordability when it comes to daily token burn rates, learning the toolchains, or even the tooling needed. Instead, it’s something far more banal: determining whether your organization’s processes are legible.

A company with modest infrastructure and stable, documented processes is going to have more success quarter after quarter when implementing any sort of new technology. Contrast that to a startup that is built on the notion of state-of-the-art tooling, but operational chaos front and center. The machine learns from you. Fix the process, build the model, and enjoy the results of your AI success and everything it brings. Just don’t skip the legwork.