Library: McKinsey – Operationalising Machine Learning in Processes
November 2021 featured report:
While machine learning shows great potential for increasing process efficiency, this report, Operationalising Machine Learning in Processes by McKinsey, says that to generate “real, lasting value requires more than just the best algorithms”.
Far more complex than it sounds
The difference between basic automation and machine learning is, well, the learning part. Over time, ML can adapt to handle more complex roles, too, contributing to greater accuracy and efficiency.
Many companies have piloted ML, McKinsey’s global survey found that only about 15% have successfully scaled automation across multiple parts of the business, while only 36% said that machine learning algorithms had been deployed beyond the pilot stage.
Some of this is down to the way that ML is being implemented, says the report. This is complex stuff. It’s not easily distilled into simple rule sets. And is highly technical. So only the IT nerds really understand what’s going on under the bonnet. This can leave business leaders somewhat clueless and unable to direct the project’s development.
As a result, a significant amount of value is being left on the table, say the report’s authors. Those who build machine learning into the processes have increased process efficiency by 30% or more. This will tend to drive up revenues by between 5% and 10%.
The report provides an example of a healthcare company that uses a predictive model classifying claims across different risk classes, which has increased the number of claims paid automatically by 30% while decreasing manual effort by a quarter. These innovations provide scalable, and resilient processes that will continue to unlock value in the future. So, this is not a one time deal.
Four steps to ML heaven
The report identifies a four step approach for the successful implementation and application of machine learning with insurance companies or insurance companies in general.
Step one: create economies of scale and skill
Part of the problem where machine learning is failing, is because processes often are bigger than business units. Teams tend to focus on the steps they control and use ML to automate them. This is a mistake, says the report, because having a number of different small teams operating machine learning dilutes the effort across the organisation as a whole and “siloed efforts are difficult to scale beyond a proof of concept and critical aspects of implementation, such as model Integration and data governance. are easily overlooked”. Deploying machine learning to steps in the process is counterproductive, says the report. Instead, the company should be designing processes that are automated from end to end.
Where there are similarities for use cases – for example, processing documents such as invoices, claims and contracts – these can be grouped together as “archetypes” to have the ML applied to them.
This is advantageous as it makes the return on investment far more attractive for the organisation. It also allows knowledge sharing and refining this can create faster, better development step to assess capability needs and development methods.
Step two: Assess capability needs and development methods
Companies that focus on improving controls will need to build capabilities for anomaly detection says the report. The companies are struggling to move to digital channels probably focusing more heavily on language processing and text extraction.
There are three things companies can do:
1) build fully tailored models;
2) take advantage of platform based solutions;
3) purchase point solutions for specific use cases, because this is easier and faster, but does come with some compromise.
Step three: Give models on the job training
The main challenge to automation is not identifying the steps to automate, but finding high quality data that the algorithms can analyse and learn from. The old adage of BS in, BS out is no less true today as in the past. Even if a company has high quality data, it may not be suitable to train the machine learning model.
There are three distinct environments that deployments inhabit, says the report. The first is the developer environment, where models built and easily modified.
The next is a test environment – or user acceptance testing (UAT) – where users can test functionalities, but modified modifications cannot be made.
The final one is a production environment, where the system is live and available at scale to end users.
The models can be trained in any of these, but the optimal place is within the production environment, because it contains real world data. This is important, because not all data can be used in these three environments.
Regulatory requirements may prevent developers from playing with data development yet it’s also true that models won’t operate properly if they haven’t had the right kind of data to be taught or learn on. Some deal with this by operating human review of machine learning model outputs.
Using humans can improve accuracy and the report highlights a healthcare company that raised the accuracy of its models over a three month period where straight-through processing doubled from 40% to 80%.
Step four: Standardise machine learning projects for deployment and scalability
Experimentation is essential, says the report, and an organisation should be accumulating knowledge even when their experiments fail. MLOps (machine learning ops) combines software development IT operations as applied to machine learning and artificial intelligence, and when used can shorten analytics development life cycles and increase stability of the model.
People remain important to the process and organisations must assemble dedicated cross functional teams in order to embed ML into daily operations. This is because Some of the obstacles to adoption and implementation are cultural. ML requires the right mindset to make application of machine learning a success. This is not rules based automation. And machine learning requires data, so the more data the better. Planning is essential. Break it down to manageable steps, Otherwise, you can get overexcited and plan too many things too quickly. What are the archetypal use cases which capabilities needed and how will you scale them? End to end is essential. Don’t ask managers with siloed functions to develop use cases. Reimagine the entire process from beginning to end and change what is being done now.
This will deliver a far better result tomorrow.
For more, see the full report.
Link to Full Article:: click here
Link to Source:: click here