Many organizations want to "do something with AI" but struggle to decide where to begin. The most useful first project is usually modest: a well-defined task, with clear success criteria, where people stay in control.
Start from the work, not the model
List the tasks your teams repeat every day that involve reading, sorting, summarizing or drafting. Common candidates include:
- Classifying and routing incoming requests
- Extracting fields from documents such as forms or invoices
- Summarizing long threads or case histories
- Drafting first-pass responses grounded in approved knowledge
Score each candidate
For each idea, ask four questions:
- Value: How much time or friction would this remove, and for whom?
- Data: Is the information the AI needs available, accessible and appropriate to use?
- Risk: What happens if the output is wrong? Can a person review it before it matters?
- Measurability: Can we judge quality on real examples before launch?
Prioritize ideas that score well on value and measurability, with low risk. Drafting for human review is often a better first step than fully automated decisions.
Design for oversight from day one
- Ground answers in sources you control and show those sources where possible.
- Set confidence thresholds and route uncertain cases to people.
- Log inputs and outputs so quality can be reviewed.
- Agree who owns the system after launch.
Evaluate before you scale
Build a small evaluation set from real, representative examples. Compare AI output against what a skilled person would produce, refine, and only then widen the rollout.
A successful first project creates something more valuable than the feature itself: confidence that your organization can adopt AI safely.
This article is a draft for review.