CNBC reported on September 15, 2026 that OpenAI had been talking with Anthropic and Google about collaboration on AI safety. Google DeepMind’s Demis Hassabis had proposed a U.S.-led standards body in July. TechXplore reported that the proposed body would be modeled in part on FINRA, the self-regulatory organization overseeing U.S. brokers and investment firms.
A shared body could address testing and reporting across AI developers. A business deploying those developers’ systems still has to specify what its own workflows may do, who approves consequential actions, and how errors are reported.
Translate external standards into company rules
The FINRA comparison brings supervision, records, audits, and enforcement into the discussion. Those mechanisms have costs and limitations. Their relevance to an individual business depends on the work its AI performs.
A system drafting a customer response needs a review rule. One updating a CRM record needs scoped write access and a record of the change. Hiring, financial reporting, and other sensitive work need explicit approval thresholds before AI outputs influence decisions.
The reporting also describes disagreement over who should set the rules. TechXplore quoted Cohere CEO Aidan Gomez on who writes the guardrails, who participates, and whose interests they protect. Senator Bernie Sanders argued that voluntary industry standards were insufficient. Whatever external model emerges, a company will need instructions its staff can apply to a particular workflow.
Write the standard against one live workflow
Name the accountable owner and separate drafting, recommending, acting after approval, and acting within pre-approved limits. Record the agent’s permissions alongside those operating modes. Customers, money, legal exposure, confidential data, hiring, firing, publishing, production systems, and regulated decisions require deliberate approval thresholds.
Keep a trail of the documents accessed, tools used, records changed, and decisions escalated. Reviewers need enough information to reconstruct a failure, as well as the final output. Decide how long the records are retained and who may inspect them.
OpenAI CEO Sam Altman told the Salesforce conference that accidents with new technology are unavoidable and called for transparent reporting. Internally, give staff a way to report AI errors candidly. An incident review should identify the failed control and assign the repair; hiding errors leaves the team unable to learn from them.
Before expanding the workflow, inspect its record with the owner. Check whether approvals happened as intended, whether access stayed within scope, and whether someone could pause the system when needed. This provides evidence for the next deployment decision while the wider standards debate continues.
Research and structure: Mai. Direction and voice: John Lipe. Field experience: Strategy Ninjas client engagements.