Matching the company information we receive from our clients, in whatever format we receive it, to validated information from credit bureaux or our own internal data, is the foundation of so many of the services we offer at Bondaval.
If we can ingest our client’s data and automatically match it, that allows us to combine datasets reliably and deliver instant value in the form of utilisation tracking or aggregation monitoring.
Similarly, identifying exactly which company we’re all talking about is also critical for validating that we’re insuring the right customer so our clients know that they are covered in the case of a claim.
Preparing to match
To start with, we need to clean our data to give us the best chance at finding a high-quality match. All data is run through a cleaning and preprocessing stage, where irregular information is automatically converted into the correct format.
For example, as we preprocess country information, we might remove superfluous special characters like parentheses or commas, correct spelling errors or idiosyncratic listings (e.g. “Korea, South” instead of “South Korea”), and then deliver a final list, converted to ISO 2-digit country codes.
Equally, we often receive data in the CSV file format, which, while uniformly ingestible, is often exported with some characters wrongly encoded, meaning we frequently receive company names with strange collections of punctuation embedded within them.
Embedding LLM support within this data cleansing process allows us to straightforwardly and quickly standardise our data, preparing it for the matching analysis and significantly reducing manual effort and overall costs.
Making the match
Now we have our prepared data, and we’re ready to start matching. But what makes a quality match? The most leading information is, of course, a company identifier number: where one exists.
However, frequently we are supplied with lists of names that do not contain that information. Similarly, when we receive accounts receivable data on different companies, customer names are also often stored as more informal labels, rather than precisely as the company’s legal entity. Making a match in these less-than-perfect situations is where things get interesting.
Working closely with our Risk Underwriters, we can encode further information that gives us clues to close in on our match. Can we infer the country of a company through its company legal suffix? Another example: if we know a business operates in Texas or in Tokyo, we might be able to deduce that it’s in the US or in Japan.
Turning this into a fully mechanised process is the key to delivering high-confidence matches in as many instances as possible. To do so, each possible contributor to the quality of the match is scored and arranged hierarchically from the most deterministic to the least.
Where confidence is high, we can auto-match the input to a validated output company, or where confidence just falls shy, we can provide a high-confidence match for our Risk team to approve. Already we’ve matched more than 82,000 individual companies, including from over 15 million records of accounts receivable data.
Continually improving our matching
To ensure people have trust in the matches our system makes, we need to deliver as close to perfect accuracy as possible. As a result, we’ve been painstakingly iterating on our matching algorithm to improve accuracy and speed.
In addition to helping us to make the match, designated LLMs also review each match to add an additional layer of checking. We’ve also been building in additional edge-case rules in collaboration with our Risk team with the goal of increasing our accuracy rate to over 99.999%.
Matching companies is a critical challenge for any underwriter. At Bondaval, we’ve found using technology and AI has helped us to make that challenge far more manageable, so we can give even more value to our clients.
