AI at work, and the question of trust
OpenAI’s latest customer stories put everyday workflows in focus. A closer look at what they demonstrate—and the safety questions they leave open.
Two customer stories published by OpenAI today focus on the less glamorous parts of AI: answering email, testing software, and deciding how much work to hand over. They offer useful examples, with an important limit: these are accounts published by a vendor about its customers, not independent evaluations.
Email is a workflow
OpenAI’s Fyxer profile describes an assistant that separates email into smaller tasks, including deciding whether a reply is needed, retrieving context, and drafting a response. The company tests changes against its own email tasks and learns from the edits users make.
Our takeaway: for someone building a service or improving their own workflow, the interesting part is the sequence of decisions around the draft. A useful experiment would measure how often a draft needs correction, and which corrections matter, before granting the system more responsibility.
Testing the tools that do the work
OpenAI’s Perplexity story describes using GPT-6 Astra to generate testing programs and stand-ins for connected services. Perplexity’s cofounder also describes using the model for communications and production software, with less frequent checking than before.
That is a customer’s description of its experience. It is not a general reliability guarantee for other applications. Our view: tests, explicit permissions, and a way to reverse changes remain useful even when a system appears capable.
The safety context
For a broader reference, the International AI Safety Report 2026, published in February, describes uneven capabilities and growing difficulty with reliable pre-deployment evaluations. It also notes that most risk-management initiatives remain voluntary, while some jurisdictions are formalizing requirements.
This is background reading, not news published today. Its distinction between observed evidence and uncertainty is worth keeping in mind when reading any company’s success story. The report does not endorse a particular policy approach.
Our view: useful applications and meaningful oversight belong in the same conversation. We can appreciate an effective tool while asking who is accountable when it fails.