AI Automation with MLOps
Getting real value from AI is a struggle, but MLOps (a compound between Machine Learning and Operations) can lend a hand. Datatonic’s Director of Machine Learning, Jamie Curtis, explains why and how MLOps enables AI automation in this blog.
Automating AI is the most profitable way of ensuring it has impact.
Companies are running full-steam-ahead in the quest for extracting value from data. Many now have fully capable in-house data science (DS) teams; some companies are built around them entirely. The successful ones may have adopted AI/ML use cases, and are seeing the benefits. The most successful though, have not just adopted AI with proof of concepts; instead, they have automated AI to productionize it rapidly.
A recent report  found that companies automating AI (full adopters) have 7pp higher profit margins than those who have built only proof of concepts (partial adopters). So if there’s so much value in automating AI, why isn’t everyone doing it?
Exhibit 1: Profit margin against respective industry average, for Full AI Adopters and Partial AI Adopters.
When it comes to automating AI, companies are struggling.
About 1 in 10 corporate AI initiatives actually make it into production. Even with the experimental and uncertain nature of AI applications, this is low and inefficient. This is a consequence of two natural occurrences, that have led to a host of issues:
- AI applications in business are novel. The pace of technological advancement in AI is far greater than the pace of validated AI use cases. In the race to deliver value, stopping to design automated solutions isn’t a high priority for a resource-limited enterprise. To put this pace of technological change in perspective, the number of academic machine learning publications on ArXiv is currently doubling every 18 months.
- There is a clear gap in automation capabilities within AI. Data Science roles are extremely sought after still, but with the handover of every data science project, comes a further requirement for embedding the output in the business. As a result, skilled ML engineers and data engineers are scarce. ‘Machine Learning Engineer’ was the highest growth rate role on Indeed’s site from 2015-2018 at over 60% YoY , and demand for ‘Data Engineering’ roles grew 50% last year .
The commercial viability is low for most data science projects.
Projects fall flat, confidence dries up and budgets dwindle when not approached correctly. With up to 80% of ML project overhead being spent on manual tasks – that could be automated – it’s easy to see why.
When businesses think about machine learning, they think typically about building a model with a data science team. In reality, building a model might get you 5% of the way to production. A successful machine learning project requires capability across a much wider spectrum. For companies without the end-to-end capability required to deliver these tasks, this is a major roadblock to progress.
Exhibit 2: There is more than just a model build required when productionizing ML.
The solution: MLOps, Machine Learning’s tailored approach to productionising models.
MLOps focuses on taking ML models into production by the most efficient means possible. MLOps has formed naturally over the past 3-5 years, in response to the challenges posed by enterprise ML, and is still evolving at speed.
Without MLOps, these tasks present a host of issues, driven by the factors mentioned earlier of 1.) AI’s novel usage in business applications, and 2.) the resourcing gap within automation & engineering of AI.
Issues presented come from:
- Manual Processes: Deploying and maintaining models becomes cumbersome for Data Scientists, and takes time away from tasks within their core skill set: e.g. analysing data, building models, researching methods, constructing features.
- Code Development & Management: Code isn’t easily shared or reused. Health checks, code updates, version control, testing and monitoring are time costly and out of a Data Science’s remit, and add to the overhead of projects that could be automated
- Reproducibility and Transparency: Deployment processes aren’t standardized, and decisioning from models can’t be tracked, or managed, which can pose enormous problems in a regulated environment.
- Tooling Selection: The choice of tooling and development practices is irrespective of deployment and automation needs, and functionality becomes restrictive towards adding real value.
With MLOps, the process is standardised, steps are automated, goals are aligned and tooling meets the business requirements. An ML Ops solution provides:
- Process Automation: Model deployment is automated for all models. Model iterations and updates become easy and quick.
- Process Standardisation: Hacking and short-cuts are eliminated. E2E process architecture becomes embedded in the business with full transparency and traceability.
- Flexible Ways of Working: Dynamic, adaptable pipelines combine with common goals amongst DS, Eng & Ops.
- Sufficient Tooling: Tools used support the process and create a robust architecture, which supports the business need. Functionality enables compatibility, parallel working, speed of decision making and cost-efficiency.
The application of MLOps to Automate AI is different from business to business. The certainty though, is that there’s no better time to start than now. Every model that a company puts into production outside of a standard process, is another model that must be rewired back through the standard process later on.
At Datatonic, we have world-class ML Engineers dedicated to automating the ML process. Our ML Engineers can productionise your AI initiatives, using an MLOps approach, and have done this for many leading companies. Check out our latest case studies with RealEyes and Lush to find out more.
 McKinsey, Artificial Intelligence: The Next Digital Frontier?, June 2017.
 Indeed, The Best Jobs in the US, 2019.
 Dice, Tech Job Report, 2020.