THE AI DESK
Artificial intelligence through the lens of products, economics and accountability.
131 stories to explore
Public results and machine-checkable proofs make research easier to inspect. For AI product teams, the lesson is to build evidence that customers can examine beyond a polished answer.
Embedding skills into an institution’s existing workflow is a distribution choice. Its value depends on delivery quality and evidence that learners benefit.
Google has widened access to its media verification tool. For startups, the useful question is how evidence of origin changes the product workflow.
AI products can have variable delivery costs and uneven customer usage. A workable price needs to make both sides of the transaction understandable.
A model can look impressive in a demonstration while failing the tasks that determine whether a customer trusts the product.
Retrieval can give an AI product access to company information. It also creates an operational responsibility to update, remove and permission that information.
An agent’s authority should follow the task it has been asked to perform. A capable model does not need unrestricted access to be useful.
Documents and webpages can contain instructions aimed at an AI system. Products need to keep source material separate from the authority to act.
Using different models for different jobs can improve economics, but retries and escalations must be included in the calculation.
A review step adds value only when the reviewer has enough context, time and authority to intervene.
Generating an attractive fashion image is one challenge. Producing a dependable catalogue is another. AdBanana’s product proposition puts that distinction in focus.
