Moving Expenses to Procurement with Procurable Insights

At Coupa, one of our three core values is “Ensuring Customer Success.” To us, this means that we have to go above and beyond to find value for our customers even when they aren’t looking for it. Procurable Insights is a key capability of Coupa’s Community Intelligence feature of the Coupa Expense offering in that direction. The solution uses a state-of-the-art and patent-pending artificial intelligence (AI) algorithm to identify expenses that could have been procured.

For those new to the terminology, expenses are submitted after the item or service has already been purchased. This is often called “post-approved.” A procured item or service goes through a procurement process where suppliers are identified, negotiated with, and possibly even subject to an RFP. This is often called “pre-approved.”

Our analysis shows that ~10–5% of the post-approved spend is something that our customers or the community had procured in the past. Identifying that spend helps our customers reduce costs by nudging users to submit orders instead of expenses and get the item at pre-negotiated, contracted, and discounted prices.

How we solved the problem

Coupa’s unique solution to this problem is typical of the AI work we do. Identifying procurable expenses is not something being done by other firms. The solution requires us to combine advanced AI algorithms along with common sense rules, and our vast store of community data allows us to learn patterns across many companies.

Since the problem is new, we had to start with first principles and imagine how a person would frame this problem. This is what we think a person might think: “If the item expensed has also been historically invoiced, then it might be procurable, although some categories of expenses (like meals, taxis, etc.) might not be realistically procurable even if they have been invoiced in the past.”

We had to translate this logic into a process that a machine can handle:

1. Assign every expense a category leveraging a classification approach (airfare, meals, lodging, etc.).
2. Select categories that would be realistically procurable (office supplies, gifts, etc.).
3. Compare the expense description to all invoice descriptions in the past. To make it more robust, use a semantic similarity measure. (For more discussion on semantic similarity, see our blog on duplicate invoices where we talk of creating procurement related natural language processing.)
4. If we find an expense description with high similarity in the same customer’s invoices, then it’s “procurable by you.” If we find an expense description with high similarity in invoices with the rest of the community, then it’s “procurable by community.”

Even with this approach, a technical challenge we faced was to compare every expense line (with more than 10 million for the community) to every invoice line (with more than 100 million for the community), which led to the number of comparisons on the order of 1,015. To address this, we used a combination of clustering and sampling to reduce the number of invoice lines for comparison by 90%. This allowed us to find a practical, implementable solution.

Here is a flow chart of the algorithm. You can see that the approach stitches together many different steps to solve this problem. And, we use a mix of machine learning (ML) and common sense rules.

Because of these innovations, we have a patent pending for this work. Take advantage of Coupa’s Procurable Insights capability and move more of your expenses to procurement.