Josh Elliot, head of operations at Modzy, discusses how the build versus buy decision is not mutually exclusive when it comes to artificial intelligence implementation. Organizations can use a combined approach to integrate AI that best fits their needs and drives tangible value and outcomes.
Today’s economic climate has put a microscope on how organizations are expending valuable budget and resources. With this backdrop, leaders charged with implementing Artificial Intelligence needs to demonstrate value, quickly and responsibly. Given the maturity of this technology and the talent shortage, leaders also need the flexibility to shift direction when introduced to a better approach. So how can organizations smartly make forward progress on their AI journey?Â
Many start with a capability maturity assessment against their organization’s AI goals to identify gaps or deficiencies in their existing capabilities. Then, a diverse group of stakeholders discusses the Tradespace for how to best fill those gaps. Inevitably, the group is asked to choose between building the capability in-house or buying it from a supplier.
What many organizations don’t realize is that the build versus buy decision is not mutually exclusive. Yes, there are edge cases where you might need to choose one over the other, but many organizations are doing both to expand their competitive differentiation and improve mission effectiveness. Organizations can use a combined approach to drive tangible value and outcomes.
Learn More:Â How is Artificial Intelligence Impacting Healthcare?
Try a Modular Approach
If you’re trying to establish an AutoML or machine learning operations (MLOps) capability for your organization, buying commercial-off-the-shelf (COTS) software solutions might be the right path for you. According to Gartner, getting AI from the lab to production can take up to nine months, and buying COTS could be a significant accelerant and create efficiencies for your organization.Â However, be wary of proprietary platforms that claim to offer everything from data preparation to model training, model deployment, and model operations. Yes, an end-to-end solution may look tempting on paper, but they rarely deliver. Not to mention, the technical debt associated with an end-to-end solution can be costly to unravel.
Instead, consider a modular approach that puts you in control of which capabilities you preserve for in-house development and which capabilities are best enabled by procuring and integrating software. Focus on finding open architecture software solutions with robust APIs and SDKs that complement your existing infrastructure and workflows and allow your teams to work how they like.Â
Your data scientists will thank you because they get to use their preferred ML training tool, language, or framework. Your software developers will embrace the minimal code required to embed AI in production applications via available APIs. And, your execs will appreciate the accelerated AI adoption timelines, the flexibility and extensibility of a future-proof tech stack, and a generally lower total cost of ownership.Â
On top of that, this approach could help put a stop to the ownership â€œfood fightâ€ between many IT departments and business units. The IT department can own the capabilities that service the enterprise while the deeply proprietary capabilities are built and owned by the business units.
When it comes to AI models, in-sourcing the design, development, and maintenance of a model might be a better approach if you have the talent, unique data sets, infrastructure, a domain-specific mission or business needs, time, and budget. Building your own bespoke AI models certainly has its advantages.Â
It is especially useful if you work in a highly-regulated or mission-critical space, where you have strict service uptime requirements and need to know who built the model and how it works, should anything suddenly break or require retraining. Those that choose to go this route should make sure they have a model repository, runtime, and monitoring capability to ensure intellectual property is maintained and governed effectively.
When To Explore the Marketplace?
If you are missing one or more of the ingredients for successfully building your own models, consider having your data science team explore existing model marketplaces for available open-source and commercial pre-trained models before venturing down the build path. You may find that an existing model gives you 70 percent of your solution. While it may not be perfect, it will be better than your current performance. A subscription to the model or API could provide a helpful temporary solution and cost a lot less than building it yourself.
If you think subscribing to pre-trained models makes sense for your organization, look for AI software vendors and solutions that enable you to retrain models on your data while preserving your intellectual property and offer key features like explainability and adversarial defense. These added capabilities could get you to your target performance objectives and advert the entire build process (and cost).
As you explore build-and-buy approaches, remember that success is about getting your stakeholders excited about both the process as well as the outcomes. At the end of the day, any AI solution Opens a new window must have a long-lasting impact on your organization.