Few Consumers Trust AI Recommendations, ServiceNow Issues Agents a Passport, and Gemini 4 Argon Prices Double in December

Three stories this week share a theme: whatever your chatbot or agent does, someone downstream has to trust it, authorize it, or pay for it. We could not open most primary pages from our environment, so details come from search summaries and secondary coverage, and we flag where sources disagree or figures are vendor claims.

1. Consumers Use AI Assistants but Mostly Do Not Trust Their Recommendations

Coverage of Kantar's Media Reactions 2026 study, which reportedly surveyed 23,452 consumers in 33 markets between July and September, says only about a third use AI assistants to research brands and products, and under a quarter trust the recommendations they get. Sources disagree on the exact figures: one gives 32% and 23%, another 33% and 24%. We could not find the primary report, so treat both as approximate.

Separate Kantar data from earlier this year, focused on the UK, reportedly found that around 41% of consumers think brands recommended by generative AI paid to be there. Whatever the precise numbers, the gap between use and trust is the point for builders. If your chatbot recommends products, sources and uncertainty are product features, not footnotes. Show where a claim came from, say when you could not verify something, and keep sponsored content visibly separate from answers.

2. ServiceNow Gives Agents a Passport

On October 15, ServiceNow announced AI Passport, described as a runtime authority layer that controls what agents and MCP servers may do at each step of a multi-agent workflow. Per the press release and trade coverage, each agent receives a cryptographically signed credential, checked at runtime, that is tied to policy and workflow approvals. Scoped delegation narrows at each hop, permissions can be revoked, and out-of-scope behavior can trigger redirection, restriction or a kill switch, with everything audited. NVIDIA is named as a launch partner, and ServiceNow says the approach is built on open specifications. We found no independent testing and no confirmed pricing or availability, so this is a vendor announcement, not a verified result.

The design idea transfers to any chatbot with tools: check authority at the moment of the call, not once at login, and never let a model's own claim about its permissions count as evidence. A minimal version:

type Grant = { tool: string; maxAmount?: number; expiresAt: number };

function authorize(grants: Grant[], tool: string, amount = 0): boolean {
  const g = grants.find((x) => x.tool === tool && x.expiresAt > Date.now());
  return !!g && (g.maxAmount === undefined || amount <= g.maxAmount);
}

// every tool call, every time:
if (!authorize(session.grants, call.name, call.args.amount)) {
  return { error: 'not permitted' }; // log it; do not let the model argue
}

Grants live server-side, expire, and narrow when work is delegated to a sub-agent.

3. Gemini 4 Argon's Introductory Pricing Has a Deadline

Reports on Gemini 4 Argon, Google's model for long-running tasks, put API pricing at $2 per million input tokens and $10 per million output tokens through November 30, with both rates doubling in December. Sources also disagree on launch dates (some say September 30 for a restricted rollout, others October 13 for broad availability) and on whether the roughly 1 million token ceiling applies to output alone or to input and output combined. Benchmarks such as 77.9% on DeepSWE v1.1 are company-reported.

For anyone prototyping on it, the budgeting lesson is to model costs at the post-promotion price now. Long-running agents multiply output tokens, so a plan that works at $10 per million may not at $20:

const PRICE = { in: 4 / 1e6, out: 20 / 1e6 }; // assume the December rate
const cost = (u: { in: number; out: number }) =>
  u.in * PRICE.in + u.out * PRICE.out;

Pair that with a per-task token ceiling so one runaway job cannot spend a month's budget overnight.

Conclusion

Trust, authorization and cost are the unglamorous constraints that decide whether a chatbot survives contact with production. Cite your sources and admit uncertainty, check permissions on every tool call with expiring server-side grants, and budget against the price you will pay, not the promotional one. As always, verify vendor claims against primary sources before planning around them.

A simple chart comparing introductory and post-promotion monthly cost for a long-running agent workload would suit this post well.

— Maya