Nearly one-third of enterprises admit they don't use AI-based solutions. Tackling the data, tech and deployment-related challenges in a large company can be overwhelming, but not impossible. Filipa Castro, Data Scientist at Continental shared some good practices with us on how to successfully implement ML projects in the enterprise.
According to the recent O’reilly’s AI Adoption in the Enterprise 2022 report, only slightly more than one in four (26%) of companies surveyed implement AI-powered solutions in production. A vast majority (43%) is currently evaluating different use cases and 31% openly admits not to use AI-powered solutions at all.
Challenges companies face when implementing ML based solutions
Continental is an international automotive parts manufacturer and the world’s fourth largest tire-making company. Headquartered in Hanover, Germany, it operates in the whole world, providing tires for all major automotive companies and brands, including Ford and Volvo among others. The company employs over 232,000 people and generates 37.72 billion euros (roughly 46.28 billion USD) in revenue (as of 2020) .
In companies of this size, there are already established processes and guidelines regarding building a traditional software. If you plan well, everything moves smoothly”, comments Filipa Castro, Data Scientist at Continental and one of keynote speakers during the recent DSS Virtual event focusing on AI and Machine Learning in the Enterprise. “For a machine learning software the cycle looks different” she adds.
Tech-related challenges
In contrast to typical software development, building machine learning-based solutions requires vast amounts of data to train the model and an established ecosystem of supporting tools. As Filipa Castro pointed out during her speech, this can be a huge challenge for companies, even for established and large ones. Among the most common she named:
- Data preparation and quality - the company may not have enough data to launch a desired solution, or the data can have quality issues like non-standardized format or noise.
- Research - developing an ML solution may require an R&D related to the problem to solve.
- Speed requirements - the delivered solution may be required to run on a particular hardware that can come with limited memory and computing power.
- Deployment - the solution may be required to run as an app, a dashboard or just an endpoint, depending on the needs of the team and company.
Tackling these challenges requires establishing good practices and delivering new workflows that will support the new paradigms of software development. This is usually a part of a larger AI governance plan - yet one is not always in place.
The graph below shows the share of companies having the AI governance plan in place. The majority of all organizations don’t have an AI governance plan at all, with 51% of those using the AI in production, and 78% of those evaluating the usage of AI claiming to not have one.
Best practices for implementing ML based solutions in the enterprise
“Sometimes we approach teams that don’t have any data at all and we need to guide them through the process of data collection. In contrast, there are teams with good data and clear expectations, who need guidance only” says Filipa Castro.
Using her experiences in Continental she delivered five tips that facilitate the implementation of AI-based solutions in the company.
Show what is unfeasible with the current dataset
“Managing expectations regarding the performance and the abilities of an ML model is crucial for the success of the whole project” says Filipa Castro and mentions the real-life example of a model able to detect defects on manufactured tires.
“Some defects are not represented in the dataset, so there is no way to detect them using ML. Such a situation needs to be communicated to the stakeholders to not set unrealistic expectations.”
Break the problem into sub-problems and focus on the most relevant and critical ones
According to Filipa, there were multiple types of defects to detect in the example mentioned above, with various levels of severity. Initially, the stakeholders expected the solution to detect all types of them.
“It was an extremely ambitious task and focusing on the most common defects could have been enough. On the other hand, if one particular type of defect is uncommon yet has the greatest impact on the business, we might need to focus on collecting data and automate the detection” she said.
Use human agreement as a guideline to set expectations
ML based solutions are expected to augment and sometimes replace human labor. “However, humans labeling data do not always agree when classifying images for example. If they don’t agree at least 60% of the time, we cannot expect the model to perform better than that”, says Filipa Castro.
Design standards and processes for data collection
Data science comes with several repeatable challenges to tackle, including storage, data collection, labeling and tracking experiments among others. Having a ready-to-use tech stack is a direct response for this issue.
Follow a data-centric approach
Last but not least, the nature of the project is usually determined by the data available to the company.
“There are a lot of questions to ask, such as: what data to collect, how much? What will be the source of it? Without the guidelines, the teams would start to collect data randomly, doing a lot of unnecessary work” says the expert.
Summary
According to Filipa Castro, the most effective way to enrich enterprises with ML based solutions is to enable data teams to forge solutions on their own while communicating actively with the management. Potential use cases need to be carefully analyzed and data scientists can support the management in setting clear expectations and guide them through the ML process without taking on the responsibility for the entire project.
