Artificial Intelligence answering your complaints

Customers are, sometimes, unhappy. When they are, one in sixty write to the brand to say it, reveals a number of studies. The other fifty-nine spread bad reputation around and some even leave the brand to one of its competitors.

When they complain in writing, many useful information is given to the brand to help it improve its Net Promoter Score ! Therefore we believe it should never be left unattended. What we did was to obtain five years of complaints of the group, and we dumped all of it in a huge Elastic Search Big Data pool. And we explored it for months.

Out of this work, a great many things were unearthed: some agencies had to be relocated, some specific employees had to be trained in a new manner, scripts of employees answering custoemrs' phone calls had to be modified. But all of these quick-wins were only a side-product of our project.

At the core of our exploration, we determined a powerful framework: we determined three scales, and trained our birthing AI model to determine, for each complaint, where it stands:

1. The motives of the complaint

his is absolutely essential to understand the root causes of the customer's frustration. Not only to better meet his needs, but also to report to the Regulator, as the banking industry has specific legislations regarding Complaints Management.

So we defined a list of possible motives, and associated each with a set of usual keywords used by customers. Our AI was then trained in five cycles to analyse these motives and give ponderations to the most probable ones. The difficulty for the AI revealed to understand if the customer was complaining for one motive only, or several. After 5 cycles, some smotives were discovered by the AI with a precision of 96%. However, some motives, lacking sufficient volumes of data, were still lagging behing at around 31% of precision.

2. The concerned business line

Banks are complex bodies. Personal Loans, Car Loans, Revolving Loans, Mortgages, Home Insurance, P&C Insurance, Travel Insurance, Credit Cards, Insurance linked to the Credit Cards, Claims covered or not by an Insurance, Savings Accounts, Regulated Savings, Securities, Fees... And for each product, the process might have failed in different parts: on-boarding, deliverance, or during any moment of the product lifecycle.

Our AI was trained with several cycles again, and by associating the right meta-data and keywords, we could obtain a fairly high positive match in over 87% of cases. The most difficult cases remained in the details of some insurance claim procedures.

3. The emotion of the customer

Unintuitively, the severity of a problem is not proportional to the customer's emotion vehiculated by the customer. One can write about a very critical situation keeping a neutral and factual tone. Others can go wild on simple and easily solvable situations.

So why bother with emotion at all ? We wanted our AI to show empathy. It felt important to us that the customers' feelings be appreciated, and taken into account in the answering process. Unfortunately, the difficulty here lies in the fact that we express our anger by using bold caracters, or by underlying some words, or with some specific punctuation. This revealed to be very tricky for our AI to learn, especially for very angry customers.

Disrupting the process.

Our AI was excellent on one point: writing answers. When a human would need several weeks to analyse and obtain answers from other internal employees, the AI just wrote in a probabilistic manner what should be answered anyway to the unhappy customer. Our AI also produced a set of probable root causes that a human could later check.

So our mission was completed: the AI produced two or three word documents in a matter of minutes, with a checklist that a human could use to select and modify the pre-written letters. It also prepared reports, both for monitoring and regulatory requirements. From two to four months, the process was reduced to a couple of hours.