· 4 min
Choosing an AI seed wedge
An AI seed should answer one costly uncertainty before the company invests in a product team or a broad platform. Choose a workflow that has a visible user, a repeatable input and an outcome someone can judge. If the value is unclear without a model, start by testing the workflow manually; the experiment should prove the user need before it proves a model choice.
Write the wedge as a testable claim
State who has the problem, what arrives, what changes for that person and what they do today. Then name the smallest assisted result that could help. A hypothetical example: a founder may test whether an operations lead wants a first-pass summary of a recurring packet, with the human checking every extracted item. This does not prove a durable product, but it can expose whether the input is available, the output is useful and the review burden is acceptable.
List the assumptions that could break the idea: access to representative data, permission to process it, variation in input format, factual error cost, human review time, model cost and the buyer’s willingness to pay. Rank them by how quickly a false assumption would make the product uneconomic or unsafe. Test the highest-risk assumption first.
Keep the experiment small and honest
- Recruit people who currently do the job, not only people excited by AI.
- Use representative examples with unnecessary personal details removed.
- Compare assisted output with the current workflow and a simpler non-AI option.
- Record errors, corrections and operator time, not only whether the demo looked plausible.
- Agree in advance what evidence means continue, change the wedge or stop.
Set a cap on time, data access and model spend. Do not claim retention or product-market fit from a demo. NIST’s voluntary AI Risk Management Framework can help structure questions about system context and evaluation, while the seed decision remains specific to the business and experiment.
For the decision rules themselves, read kill criteria for AI experiments; if the result supports a larger product effort, see moving from seed to build sprint. We can help define the experiment through services or contact.