Walk into any supermarket or grocery store, and chances are that you will be greeted with an “irresistible offer” on everything from snacks to mobile recharges. So much so, that you may completely ignore these items, no matter how lucrative the offer. Driving consumer preference through promos is an industry norm. Consider this – $1 trillion is spent every year globally by brands on trade promotions. Given such massive stakes, organizations are increasingly leveraging data science to maximise RoIs on trade promotions. 

But all that glitters is not gold.

Research shows that more than two-thirds of trade promotions every year in the USA do not break even, despite big bucks spent on Big Data and AI sciences to drive them. AI holds tremendous potential to solve real-life problems, but as deployments of machine learning models increase, so do the possibilities of incorrect results arising, owing to excessive data, failed algorithms or management inefficiencies. 

Additionally, professionals can exacerbate the situation by not focusing on regular model upgrades and failing to understand business requirements fully, thereby resulting in an ineffective last mile – the crucial link required to bring the full force of data science to deliver on-ground business performance. The result is highly inefficient utilization of available resources – like the massive $1 trillion spends on trade promotions.

This is where Tredence identified a massive opportunity.

According to Shashank Dubey, Chief Revenue Officer & Co-founder of Tredence, "Our vision is to put meaningful analytics into the hands of every decision-maker through the last mile adoption of data science. Conquering the last mile can be a gargantuan task for many, but we go after it for our customers. Our clients in Retail, CPG, TMT, Healthcare and Industrials want us to operationalize their data science models and move them to production. The answer to operationalization lies in the three Ops - MLOps, AIOps, and DataOps. Our clients profess that our approach has been key to powering decisions that uncover growth and drive competitive differentiation.”

Named a "Dark Horse" among the Top 13 AI Consultancies in the world by Forrester Research, Tredence focuses on enterprise-grade, scalable AI for retail, consumer packaged goods, healthcare, telecommunications, media, and technology industries. Its industrialized machine learning operations platform called ML Works intelligently automates and scales IT incident management, replaces manually created rules. 

To understand what this means and how Tredence delivers value, let's consider the case of a Fortune 500 consumer packaged goods company in the United States that Tredence worked with. The objective was to improve the efficacy of their Global Trade Promotion program using Autonomous Machine Learning Operations to increase revenue. The client spends $300 million yearly on its Global Trade Promotions (TP) campaigns, spanning 13 countries. This initiative, which is driven by 66K ML models that run monthly, relies on inputs from 56K promotion managers from across geographies. The enormous trade marketing effort of the CPG leader is based on 1.5 million SKU-Distributor-Geography combinations.

So what was preventing the client to turn this massive effort into profits?

• The client's model error had accumulated over time, thereby reducing ROI on TPs
• The program's massive scope, effect, and complexities were difficult to manage effectively, leading to revenue loss
• The previous TP system had several flaws, preventing business managers from using machine learning to enhance business KPIs

To add to this were more operational challenges…

• Massive data volumes spread globally, with some geographies having 500K+ rows in monthly forecasting
• High data processing time with the read/write time from DB constantly exceeding 60 sec … therefore time/cost-inefficient
• Time inefficient drift analysis where algorithm took about 7hrs to generate insights for 500K production data
• Inconsistent data quality led to a lot of production failures

Tredence had noticed similar challenges in a number of other companies as well. Based on these experiences, their solution set was based on three key pillars of Tredence's ML Works platform, which automates and scales IT incident management to deliver intelligent automation to lower operational costs, improve service availability, and lower IT risk:

• Drift Detector – Identifies errors/deviations and ensures delivery of reliable insights to business users.
• Provenance Graph – Monitors and identifies the ML model and data discrepancy responsible for errors
• Explainability – Identifies which external factors adversely affect data inputs and ML model, thus ensuring those model insights are not a black box.

ML Works effectively monitored for deviations in expected outcomes (accuracy, cadence, etc), identified what was causing them, and presented potential remedies – all without requiring any user interaction. As a result, the system has become more dependable, accurate, and intuitive, allowing for improved trade promotion planning and execution.

The MLOps pipeline in this trade promotion effort would not be intuitive or autonomous if it relied just on analytics. ML Works takes responsibility for recognizing pipeline problems, issuing near-real-time alarms, and kicking off corrective steps by adding layers of iterative processing and machine intelligence to ensure robust model health and high prediction accuracy. 

The impact and complications of the program was so huge that it required an immediate intervention to stop revenue leakage.

The impact was massive. Sample this …

• 66K machine learning models every month managed effortlessly in production.
• Reduction in data processing time from >60 sec to < 3 sec 
• Reduction in time taken for insights analysis from 7hrs to 1.5hrs 

Improvement in data quality through a dashboard enabled RCA of data entry issues, data anomalies, data bloat & hierarchy mapping issues.

• 20% increase in application usage, thanks to reliable and explainable insights 
• Reduction in production downtime from 4-6 days to ~6 hours
• 3% increase in quarterly sales since go-live

What truly made a difference though was the sustainable ecosystem built by Tredence that complements their data science capabilities. From conducting listening tours to strategy workshops to co-create solutions, they bring analytical interventions throughout the customer lifecycle to drive scale and faster-value realization.

This Is Just The Beginning

The Tredence team believes that only the surface has been scratched so far. The focus is now on extending the explainability algorithm to include auto regressors in Time Series data for better model explanations. In addition, they are looking to deploy a testing framework to streamline ML model testing processes. 

Tredence is adding intelligent virtual workers to teams of revenue managers who are continually monitoring performance and informing decision-makers of potential value-enhancing interventions. So, every time a customer decides to buy into that “irresistible offer”, a Tredence virtual worker will get a little smarter and help brands improve performance.