Billing is often a critical pain point of the customer journey in telcos and utilities. The problem is typically bigger than it seems at first glance: besides visible costs such as inbound calls that flood the customer support lines, there are many unhappy clients that do not complain. 

Heavy investments in the billing system are usually prescribed, to make interfaces error-proof and transactional processes reliable. Still, in most cases, incorrect invoices continue to be issued every month.

Faced with this challenge, our client, a convergent telco, set up a “last resort” bill review team, to manually validate invoices and send them for amendment. Since the team’s capacity is limited, how to select invoices to check,  to clear as many issues as possible?


Capitalizing on rich data from billing issues detected in the recent past, a model was developed to find potentially incorrect invoices. A state-of-the-art machine learning algorithm was combined with tailored business-driven system validations, for maximum effectiveness. Several relevant sources of billing issues were identified along the development process, such as expiry of promotional offers and product portfolio changes pushed by commercial agents.

The developed model was then embedded in a simple and friendly app, used by the bill review team to select the invoices to be scanned in each billing cycle. Close assistance was provided to the team in an initial phase, ensuring the full adoption of the solution.


In practice, the error rate in the invoices picked by the developed model is three times larger than the error rate in the invoices selected through the former approach. Additionally, by pinpointing the expected mistake, the model enables a targeted validation, making the bill review team twice as productive. The combination of both effects results in a sixfold increment in the invoices flagged by the team.

On top of that, the model also identifies invoices that are surely incorrect. Hence, no validation is needed, and such invoices are directly amended. Overall, considering both manual and automatic markings, 12 times more issues are now solved in the same time span, significantly improving the perceived reliability of the billing process.

Related Case Studies



To have impact, numbers need stories and vice versa. For business to form a complete picture, they need both Big Data (quantitative information) and Thick Data (qualitative information). Each of them produces different types of insights at varying scales and depths. They may yield interesting synergies and complement each other.   In particular, the tactical decision-making level has a large decision-making gap that can leverage this ‘dual’ approach. On the one hand, analytics-based methods are able to quantify and generalize insights (Big Data); on the other hand, there are empirical approaches that can improve the business perspective and the depth of analysis (Thick data). We’ve devised four main methods to integrate Thick into Big Data: Raise hypothesis: use Thick Data as a source of inspiration to raise hypothesis that are then tested over the population with Big Data Confirm correlations: check Thick Data insights to confirm correlations that were found through exploratory Big Data analysis Warm-up: plug Thick Data numbers into the warm-up phase of a Big Data project Full connector: start using Thick Data similarly to the Warm-up method, but then continue to use it to keep calibrating the Big Data model   Big & Thick Data in the Telco industry with a full connector approach A great example of a blend of Big & Thick Data is a program launched by a leading Telco provider, with the goal of empowering staff to interact with customers on a personalized level. The first approach was to combine 79k data variables around customer service and marketing into a single Net Promotor Score (NPS). This model helped move from around 30k NPS attributions per year to scoring the whole customer base with an accuracy of around 80%. The next step was a sentiment analysis over call center records to move into contextual customer insights, thus refining the NPS score. Leverage NPS data allowed to differentiate marketing strategy and interaction guidelines between customer groups (detractor vs. passives vs. advocates). For example, to make carefully crafted outbound calls for detractors and use electronic direct marketing (EDM) for advocates. The insights obtained were also essential to ensure there was a clear follow-up point to use with each customer. To engage staff was built a simple internal portal, with NPS, interaction scripts and key client info. At this point was critical to create strong engagement links with IT. During the first six months, 50k customer issues were identified and solved. As a result, the customer perception (NPS scores) improved.   The path towards transformational projects As we can learn from the previous case study, transformational projects involve cross-functional skills in Thick Data, Data Science and Management. Effective transformational projects that leverage Big and Thick Data require: a holistic perspective of the problem, covering multiple integrated processes, stakeholders and KPIs a solid methodological approach, to grasp the synergies of both data streams  a sustainable cultural change towards the execution and adoption of such projects to support decision making