Introduction
The current state of generative AI might make decision makers to carefully weigh the adoption decisions. There are just too many options and the information overload coming from the hype, the attempts for clarity and the counter-hypes can overwhelm our usual methods.
This piece offers a way to start a generative AI adoption journey or to rethink a struggling one for an iterative buildup that the decision makers can follow, no matter how widely and deeply complex it might evolve into.
Figuring it out
Before buying-in to a GenAI proposal, we must make sense of our adoption stance. This table offers a good starting point for sensemaking.1
Topic | Decision maker’s dilemma | Seller's Opening Approach to Sensemaking |
Early Adoption | Is it too early, or is there a risk of being left behind? | "Would you rather play catch-up by observing others first, or do you know enough to lead the race?" |
Model Size | Is bigger better? | "Does your use case prioritize speed and latency, or a broader context? How might this change over time?" |
Hallucination | Should I wait for a model that is hallucination-proof? | "Can we design your process to catch and override potential hallucinations?" |
When the decision to adopt GenAI makes good sense, the following steps can be taken:
1. Technical project team proposal
2. Change management counterproposal
3. Reconciliation to goals and direction
4. Repeat the process until a definitive decision is arrived at
The initial steps to figuring out our adoption decisions are quite simple. If we can keep the discipline of reconciling the biases to the goals whenever it seems too complex, the chance to figure it out well fast enough must be high.
The triangular configuration
Among many options for GenAI configuration, taking a good look into the triangular setup of retrieval augmented generation, prompt engineering and grounding should lead your team to find the sustainable best fit for your organization.
Determining a good RAG setup must motivate your team to have a reasonable data classification.
Grounding requires a good dynamics of context buildup, sharing, reconciliation and refinement.
The prompting system relies on the definitive prompting cycles that renders verifiable results. It starts with reasonable prompting skills that serve your context well.
Grounding feeds prompting that supports RAG to sustain the cyclical iterations.
It does not cost much to experiment on this triangular configuration, and it will let your team see and consider what are doable and can be sustained for your adoption.
Lateral breadth buildup
Adding retrieval sources, expanding your grounding participation to supply the setup with more context and prompting activities is the natural way for lateral breadth buildup.
Thereafter, a systematic lateral breadth buildup can be done by connecting one setup to another, e.g., your HR team setup to other function teams or to operational teams, fostering broader organizational integration.
Vertical layering and sophistication
Primary consideration for your vertical layering is to wrap it well with security and safety system. Each layer must be covered well with a relatively rationalized approach.
The output of the bottom layer feeds the retrievable and usable context of the upper layer. There’s considerable dependency of the upper layers to the bottom ones for enrichment and accuracy of generated output.
Conclusion
Adopting GenAI requires a balanced, structured approach that integrates pragmatic agility principles. By addressing early adoption dilemmas, configuring a triangular setup to initiate experimentation, iterative improvements, expanding laterally, and layering vertically, organizations can develop sustainable GenAI use cases. This iterative experimenting methodology not only mitigates risks but also aligns AI capabilities with organizational goals, ensuring long-term success.
References
1 - Excerpted from a reflection paper for Brent Adamson's article "Sensemaking for Sales," published in February 2022 in the Harvard Business Review.
댓글