Small-Wedge AI Adoption
Prove AI on one tangible problem, then expand through demonstrated success
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 96%
Begin with a small operational problem whose inputs, intervention, and outcome can all be measured. Build a focused AI application using data the organization already produces, such as customer behavior or transactions. Rather than leading with a company-wide transformation, demonstrate that the system can predict an important event and enable a useful response. The resulting evidence reduces the need to win an abstract argument about technology. Once the pilot shows tangible value, extend the approach into neighboring processes one use case at a time. Repeated successes create organizational buy-in, establish reusable data and engineering capabilities, and eventually justify a centralized AI function that can support larger changes.
Origin
Songyee Yoon described using churn prediction as a small, tangible entry point for AI adoption at NCSoft before the company developed a larger centralized AI lab.
Core principles
- 01Start with a narrow and measurable business problem
- 02Use existing behavioral data to identify an actionable prediction
- 03Demonstrate value before requesting broad organizational change
- 04Let visible results create internal support
- 05Centralize capabilities only after multiple use cases succeed
How to run it
- 1
Find a measurable pain point
Select a recurring business problem with a visible cost, such as customer churn. Define the outcome clearly enough that a pilot can be evaluated.
Pro tip Prefer a problem stakeholders already agree is important.
Watch out Do not begin with a vague mandate to become an AI company.
- 2
Inventory the available signals
Identify transaction, behavioral, or workflow data that could indicate the outcome before it occurs. Confirm that the data is sufficiently timely and reliable.
Pro tip Start with data already generated by normal operations.
Watch out A sophisticated model cannot compensate for irrelevant or inaccessible data.
- 3
Build a narrow prediction
Create a focused system that predicts the chosen event or improves one bounded task. Keep the initial scope small enough to explain and test.
Pro tip Use the simplest approach capable of producing actionable information.
Watch out Avoid turning the pilot into a general-purpose platform project.
- 4
Attach an intervention
Specify what the business will do when the system produces a prediction. A prediction creates value only when it changes an operational decision.
Pro tip Design the intervention alongside the model rather than afterward.
Watch out Do not measure model accuracy without measuring the business response.
- 5
Prove the result
Compare the pilot's outcome against an appropriate baseline and communicate the evidence. Use the result to answer objections with demonstrated value.
Pro tip Report a metric stakeholders already understand, such as retention or cost saved.
Watch out Do not overclaim from a small or poorly controlled result.
- 6
Expand one use case at a time
Apply the demonstrated capabilities to adjacent problems and integrate successful tools into established processes. Centralize the function when the portfolio of proven applications warrants it.
Pro tip Reuse data pipelines and lessons from each prior deployment.
Watch out Scaling faster than organizational learning can recreate the original resistance.
In the wild
NCSoft possessed transaction and behavioral data showing how players engaged with its games. Yoon proposed predicting when a player was likely to leave and using interventions to help that player remain engaged. Although developers and business staff initially resisted the idea, the narrow use case offered a tangible way to demonstrate data-driven value before expanding AI into other operations.
→ Successive practical applications helped AI become integrated into business processes and supported the creation of a centralized AI lab.
Common mistakes
Leading with transformation rhetoric
A sweeping vision can intensify resistance when stakeholders have not yet seen practical value. Begin with a bounded result they can inspect.
Stopping at prediction
A forecast without a defined operational response does not produce the intended business outcome.
Scaling before proving value
Large teams and platforms created before a successful use case can increase cost without increasing confidence.
Is it for you?
Best for
It is best for established companies with useful data, skeptical stakeholders, and a clear operational pain point.
Not ideal for
It is not ideal when the proposed technology has no measurable use case or the organization lacks usable data.
From the transcript
“a small, very tangible area that we can apply to technology and once you show the success, I think there is a gradually one by…”
“you can predict if that player is about to turn out or continue with a with a game, and some kind of interventions could help…”
From the episode
Investing for the AI Shift: Masters in Business with Songyee Yoon
Songyee Yoon