Navigating the Maze of Service Address Management in IT Network Reselling

For carriers in the IT network reselling business, managing service addresses effectively is more than a clerical task - it's a cornerstone for robust revenue and cost management. However, ensuring the accuracy of these addresses is often more complicated than it seems.

The Challenge with Service Addresses

While valid service addresses are essential for provisioning circuits, the reality is that accurate information isn't always readily available for inventory management. This gap arises from various operational practices: some orders are placed through informal email exchanges bypassing the standard ordering systems, and frequent company acquisitions often result in a jumbled mix of information scattered across different systems.

As a result, carriers frequently turn to their billing documents (carrier invoices) to extract service addresses. But here lies a new challenge: the data on these invoices is often inconsistent and disorganized. This disarray means that clean, usable addresses are not always available. In my experience, only about 20% of the service addresses extracted from invoices were immediately usable without cleansing step.

The Machine Learning Misstep

Initially, I was convinced that machine learning (ML) would be the silver bullet for these address issues. After experimenting with several ML techniques, the results were underwhelming, to say the least. This was a valuable lesson: ML isn't a panacea. Specifically, structured data like addresses don't always lend themselves well to ML solutions. So, the question arose: how could we transform the dirty invoice addresses into clean accurate service addresses?

Practical Solutions Over Tech Hype

The answer, it turns out, lies in simpler, more grounded approaches. For instance, service addresses on invoices often abbreviate city names, creating mismatches with geocoding data when trying to extract full service address data. A simple yet effective workaround is to use zip codes to derive the correct city names. Additionally, sophisticated regex (regular expression) functions can be applied to filter out irrelevant or erroneous data.

Implementing these methods led to significant improvement in the quality of our service address inventory. Initially, I could only match 20% of service addresses using the geocoding tool on uncleaned invoice data. But with more refined processing, this figure jumped to over 60% - a substantial leap forward. And we're not stopping there. By continuing to refine our methods, we aim to push this match rate even higher.

The Takeaway

The key lesson here is not to default to ML or other advanced technologies, hoping they will automatically solve all your problems. Instead, start by understanding the specific challenges and the types of data you're dealing with. Sometimes, the best solution might be a simpler, more direct approach that addresses the core of the problem.