GenAI Best Practices and Considerations: A Perspective from an ML Product Lead
Expert advice on mitigating risk and maximizing the benefits of generative AI
Generative AI (GenAI) and consequently large language models (LLMs) have gained significant traction and attention in the enterprise space this year. The hype promises transformative applications that are flexible to various industries and highly capable. As someone who has supported “conventional” ML product discovery, design, tuning, deployment, and monitoring in production, I am excited by the prospect of highly flexible functionality like GenAI. However, I believe many of the challenges and hurdles that conventional ML has historically faced in enterprise deployments remain and intersect with all GenAI use cases. By incorporating best practices and asking key questions at each stage of the development and deployment process, we can mitigate risks and maximize the benefits of GenAI.
I will mention that this article does not focus on GenAI and LLM capabilities, as there is already a wealth of credible material on that topic. However, as an organizational leader, we encourage you to build a foundational understanding of GenAI capabilities, such as fine-tuned quality assurance, documentation search, concept parsing, code development support, code debugging, text summarization, and more. Additionally, it is crucial to understand the nature of ML and GenAI outputs as a “prediction” or a best guess, which can range from being perfectly right to somewhat right or even not close at all. This spectrum of performance and its correlation to expectations of success and risk are crucial considerations. GenAI have traction because they get “close to right” often enough, with little effort, to justify the excitement.
Best practices and key questions
Consider the hype cycle and let outcomes drive the solution not the tool.
The hype surrounding GenAI can be both valuable and detrimental. On one hand, hype brings attention to new tools and fosters a willingness across stakeholders to adopt them, leading to exploration, experimentation, and refinement for new use cases. On the other hand, hype also invites critique, sometimes to the point of aversion, and necessitates a thorough examination of the landscape of evidence and opinions regarding a tool like GenAI. As adopters of GenAI, it is crucial that we navigate the hype cycle carefully and leverage it to drive meaningful impact while managing risks in our organizations.
GenAI can sometimes feel like a big hammer where everything is a nail. However, in this analogy, it is not the nails themselves that matter, but rather what they create when placed where needed and hit thoughtfully. The true measure of any GenAI deployment is the impact it has on the business, measured and traced back to key metrics such as revenue, time to delivery, customer acquisition cost, and operational costs.
Before even thinking about the tools, you should ask yourself a question to which you may already have an array of answers: What are the biggest bottlenecks to my business succeeding? From there, a competent ML product team can help design a solution that improves the metrics affected by those bottlenecks.
Define success criteria early and refine with new information.
- Why this use case?
- What business outcome will this use case achieve if successful?
- What is the measurable definition of success for this GenAI/LLM use case?
- How will we measure business success metrics and solution model success metrics?
- How often will we measure?
- How does this metric align to how we measure the success of our org?
You don’t need to have all the exact answers up front, but often organizations initiating ML-based solutions struggle to provide clearly defined criteria for even a few of these questions. The criteria will evolve as you make new discoveries from proof of concept (PoC) to pilot to production, but you need to start somewhere and consider how they may need to change based on new information. Report the impact of these changes. Otherwise, the cost of any ML or GenAI implementation can quickly become a sore spot for those tracking top-line metrics.
Consider risk and impact trade-offs of your use case candidates.
ML applications have always required considering the trade-off between the risk and impact of their predicted outcomes (fig. 1), often referred to as model risk management (MRM). The same now holds for any GenAI use case. For instance, automated medical diagnosis carries high impact but very high risk, whereas low-priority internal IT ticket triaging may have low impact and low risk. However, there may be rare but opportune cases where a seemingly simple task can have high impact without high risk, provided that incorrect predictions can be managed within your organization. This low-hanging fruit of high impact and low risk should be the mantra of your use case discovery team. This is the golden goose scenario, with other use cases spreading elsewhere across the risk and impact quadrant.
As you build that candidate set of GenAI cases, thoroughly discuss and role-play risk scenarios. This may include security risks, injection risks, performance threshold risks, skills and feasibility, development timeline risks, model performance drift, unplanned edge case risks. This best practice, which has been effective in ML deployments, should be considered a fundamental aspect of GenAI implementation as well.
For example, with performance risks, in a previous ML implementation, our team developed a solution for an insurance company to automatically adjudicate claims. We engaged in extensive discussions about the impact of reducing adjudication times versus the risk of incorrectly paying out claims. To address this, we helped the organization establish dollar/risk thresholds and claim volumes for potential claims that posed lower risk if adjudicated incorrectly but still benefit the org due to intensive cost savings in time and throughput. Even the process of establishing an acceptable threshold of success post-PoC can become a risk to development timelines. We have always needed to consider the risk of overinvesting in performance that may not be obtained and how “good enough” is defined. This constant search for “perfection” is the enemy of “impact” and can result in constant back-and-forth between fine-tuning, searches for new training data, and so forth, all of which can take a model from a six-week initial development timeline to quarter, half, or more of a fiscal year. Keep all these risks in mind, not just the obvious risks.
This process and approach to risk management should be extended to GenAI implementations, whether they are for internal or external purposes. Also, this process should yield directly measurable metrics that foster a clear understanding of ROI inclusive of costs associated with discovery, design, build, deployment, and monitoring. Additionally, this measured consideration of risk should be a component of your success criteria.
TL;DR
While GenAI presents immense opportunities in the ML product ecosystem, it is crucial to approach its adoption with a critical mindset and awareness of how a GenAI solution aligns to organizational measures of success and risk. By understanding the hype cycle, encouraging clarity in desired outcomes, and addressing challenges through best practices, we can harness the full potential of GenAI to create impactful enterprise solutions. In future series we hope to cover GenAI as a data product, workforce and operating model items, and other key cross-practice best practices and considerations related to GenAI and LLMs.
Reach out for a deep dive on any of these topics. We can share perspectives, ask the right questions, and tell stories about our experiences.
Acknowledgments: Chris Kercher, Nabor Reyna, Eugene Goei
Slalom is a global consulting firm that helps people and organizations dream bigger, move faster, and build better tomorrows for all. Learn more and reach out today.
