Capability becomes useful in a particular job.
A model can produce fluent output while still missing the context a business needs. A useful system starts with a defined task, the right information, permitted actions, and a way to recognize when an expert should take over.
For Cloud Genius, I separate the customer experience, approved knowledge, workflow logic, and tools. That makes it possible to test each part and understand where a problem occurs.
Research into collaboration is evolving.
Dell’Acqua and colleagues’ published 2026 field experiment involved 791 randomized P&G professionals working on product development tasks. It examined how AI affected performance, expertise sharing, and collaboration, with results specific to that setting. Read the published field experiment ↗
The relevance to my work is a question about team design: which tasks benefit from machine support, and where does expert evaluation create value? The study does not establish an outcome for a Cloud Genius deployment.
Knowledge and training are different design choices.
Lewis and colleagues investigated combining information retrieval with generation on defined NLP tasks. Retrieval can connect a response to external material, but it still needs source quality and evaluation in an operating business. Read the original machine learning paper ↗
Curated sources, instructions, examples, model tuning, and tool permissions do different jobs. An implementation should make those choices explicit, maintain the information it uses, and test the behavior the business actually needs.
Define who acts when something is uncertain.
NIST’s voluntary generative AI risk profile provides guidance for identifying and managing risks throughout a system’s life cycle. It is a useful reference for planning evaluation and responsibility. Explore the NIST guidance ↗
In practice, I agree on action boundaries, review points, ownership, and measures with the business. Then I test realistic requests and difficult exceptions. The useful result is support the team can understand, question, and maintain.
SOURCES & CONTEXT
Preregistered field experiment. 791 randomized professionals at P&G in a product development workshop; 776 completed post task surveys.
Voluntary risk management guidance. Guidance for design and evaluation; not a certification of Crave Genius.
Machine learning research. Specific models and benchmark tasks; deployed systems still require evaluation.