Machine learning is all about optimization

Operations in Year 2022

Gaurav Sharma

Operations in Year 2022

I am motivated by the vision of how manufacturing operations industry and procurement will evolve in the year 2022. To me, the new foundation for the role of an effective CPO and COO in 2022 has already been laid out. Having core domain knowledge of procurement and supply chain operations is a pre-requisite, but having a hands-on understanding of data science and automation in business scenarios is where the strategic direction lies for CXOs.


The diagram below shows a value stream map of a manufacturing organization’s operation process. It is an arrangement of 6 major verticals, starting from the Sales Forecast to Order Fulfillment.

Each vertical has its own individual business environment variables and multiple stakeholders. For example, the Sales Forecast is generally a consolidated demand summary statement of each individual sales office or unit in the distribution network.

Sales and Operation planning (also known as S&OP) is an optimization activity where SKUs production cost, lead time etc are the objectives to be optimized.

It is important to know that an efficient interaction between each of these verticals is extremely important for the bottom line of a company. An efficient interaction means a seamless transfer of clear, specific and accurate requirement. How I wish this was the case. In my view, demand forecasting and S&OP continue to be the initial point of “Leaky Buckets” phenomenon. “Leaky Buckets” is my metaphor of an inefficient operation where costs are leaked due to the poor quality of execution.

A COO must develop a set of performance indicators to measure the Leaky Buckets. On a prima facie, here is what I think the problems look like from a data science perspective. Although the diagram is self-explanatory, if you want to take the 3 key problems from a data science perspective they are data forecasting, data availability and data optimization.


Before I propose a solution I think it is important to explain the “As-is” scenario, i.e. The Present Practice.

Sales Forecasting

Let us take the Sales Forecast as an example. We all are aware of the bullwhip effect in the supply chain. How does it start in the first place? The answer is inaccurate sales forecasts.

Each sales unit in the distribution network does forecasting of each SKU it sells every month. The data from all the sales units are then consolidated into one Excel File. Then, an average scaling/normalization function is applied all across the SKUs in the direction proposed by management. Fair enough? No, it isn’t. I find this approach lazy.

The process has inherent problems in itself. Past data on a standalone basis is not sufficient to generate an accurate sales forecast. A good sales forecast should factor in the following:

1.) Correlation between different SKU sales to identify the cannibalization (this can be done for each region).

2.) Sales seasonality

3.) Sales boost due to promotions and discounts

My favourite parameter to factor in is sales lost due to non-availability of stock. This is where most ERP software and excel forecasts struggle. How do you model cases where the sale was nil because of low demand versus stock out?!

S&OP depends upon these forecasts to make production plans. A typical cycle from Sales Forecast to Production Plan is 3–4 weeks long. That means any correction in the forecast cannot be factored in before 3–4 weeks. Good luck with those unsold inventories!

If a poor sales forecast affects the top line of the company, a reactive approach to procurement affects the bottom line. I have studied the procurement methodologies of some leading manufacturers in great detail. Two words I would use to describe the present state of procurement are reactive and laborious.

Most CPOs converge to an understanding that online platforms are the answer to solve the laborious aspect. I disagree; most of the online platforms are mere data entry user interfaces only. Also, during the development phase of such online platforms ownership becomes an issue; either it an internal IT team or an external consulting agency. Outsourcing your problem to a different team doesn’t solve the issue; that is being lazy. ERP enthusiasts will present a counter-narrative of MRP/MRP modules of today but I am skeptical about garbage-in and garbage-out of the “One Size Fit All” ERP suites of today.

The world is quickly moving from statistics to calculus. The majority of the current ERP software implementation is statistics based and will become irrelevant in 2022 if a transition is not made. Most of the tools/analysis in the procurement function is still done in Excel with short-term measurement tools such as bar charts, weighted averages, cost indices, and shares, which is reactive.


The future of operations in 2022 looks radically different than that of today. Data science and machine learning will become a pre-requisite skill for purchasing and operation managers. And yes, everybody will need to learn the basics of computer programming.

Operations in 2022 will be shaped into autonomous, data-driven, and most importantly, reliable activity. Only the definition of objectives will be determined by managers; everything else will run autonomously. That’s OPS 2.0 for me!

How to Start the Journey?

In my view, the manufacturing operations industry can learn a lot form internet B2C companies, which have already started the journey toward ops 2.o.o.  The answer? Machine Learning.

In Ops 2.0 the “Leaky Buckets” structure will be replaced with 3 simple verticals: pre-production, production and post-production. Linking these 3 verticals will be the service layer of data engines.

Three principle areas required for the data management journey are:

1.) Data Organization

2.) Data Analytics

3.) Data Availability

Data Organization refers to the consolidation of all the different data sources into one database. With the diversification of teams, this is the most time-intensive operation.

Multiple Excel files, the same information in different formats, different assumption sets on the same set of problems, and multiple stakeholders are just a few factors which put high stress on taking up the OPS 2.0 journey.

  Step 1 is a collaboration. Online platforms are a good step forward, but collaboration is the step to begin with “One-time correct information accessible to every stakeholder” is the motto to adopt.

Step 2 is where the core domain knowledge of CPOs, business managers, procurement managers and operation managers is required. Business analytics will be powered by machine learning algorithms. No algorithm is perfect and even while using standard algorithm, one has to choose the hyper -parameters according to the business environment.

Machine learning is all about optimization. But, it has to be guided by an objective or as the term defined by data scientists: Cost function has either to be optimized or classified. If you are counting on your IT team to implement a one- size- fits- all machine learning algorithm prepare to be disappointed.

In each of the 3 verticals, there will be dedicated teams of data scientists needed for algorithm development + business managers who are hands-on with python and minimal coding + IT team for making the information accessible. This is where OPS2.0 will be driven by data layer as service deployment. I wish you all the best for OPS2.0 journey!

More than one perspective

Whether engaged in an innovative side project, or working on a complex challenge, Procurement League community members have a problem-solving frame of mind and lots of perspectives to share. Learn from your peers as we showcase practitioners in the top of their fields. Our members provide up-to date and valued insights.

Share your knowledge