HAIA Agents

NIST’s ten design tenets for agent identity management put most controls in infrastructure and send risky actions to human review. Tenet 05 warns that frequent escalation can turn approval into habit, and leaves who rules, and what record that ruling leaves, to each organization.

TRAI’s draft guidelines on responsible AI in telecom would require a designated officer to approve every AI system before use and make operators accountable for vendor-built AI. They say less about who rules on outputs once a system is running. Comments are due November 27.

O’Reilly found that only 24% of UK employees always verify AI-generated answers, and 39% of those using them have acted on one that proved wrong. Perceptyx and RGP point to the missing piece: deciding which AI outputs need a named person’s ruling.

Google launched an AI certificate for healthcare professionals that takes under five hours and carries up to 5.0 CME credits, and Scale and UMSL launched a 12-hour microcredential for teachers. Both teach evaluation of AI output. A European teacher framework asks for evidence of it, and employers can too.

OpenAI’s October 29 post describes up to four automated layers that watch its agents, then discloses that human paging runs at any hour only for highest-priority alerts on frontier training runs, during business hours for most other traffic, and is currently off for internal Codex. For enterprise deployments, alerts go to the customer’s own webhook, and on our reading so does the decision.

Letters dated October 27 ask five frontier AI companies for a full accounting of misalignment incidents by November 11 and notice within 48 hours of discovery. None of the fourteen requested items asks who inside a company decides a flagged event is an incident, and that decision is what starts the clock.

New Jersey’s A5607, reported out of committee on October 19, would make a chatbot a product and hold its provider liable for any injury to a user from its use, even after all reasonable care. The monthly harm assessment it requires still needs a named person to decide what the numbers mean.

The European Commission proposed on October 6 to make standards referenced in EU law free to read, speed up how they are requested and assessed, and write its own common specifications when standards bodies run late. With EN 18286 now the first AI Act standard, the question is what a faster standard can certify.

New York City’s consumer agency proposed rules on October 20 for AI chatbots and phone menus: disclose the bot up front, offer a one-step path to a human, and keep logs showing whether customers got help. The 72-hour repeat-contact count may matter most.

Barclays plans about 5,000 job cuts with fewer developers as AI lifts productivity, and American Express is cutting about 5% of positions to prepare for AI. Wider spans of control and fewer reviewers meet more AI output. Four questions to answer before an AI-justified cut.