California Signed Its Strongest Chatbot Child-Safety Laws Yet, Salesforce Gave Its Agents Names and Job Titles, and the Numbers Show Why It Had To

Three things landed within a week of each other, and none of them are really about a model release. California turned a stack of chatbot bills into enforceable law with real penalties attached. Salesforce stopped shipping "an AI agent" and started shipping seven of them, each with a name, a job description, and a shared governance layer. And new contact-center data explained, almost too neatly, why that second move was necessary. Read together, they're the same story from the regulatory side and the product side: a single undifferentiated chatbot is no longer a shape that survives contact with either a state legislature or a real support queue.

1. California's Chatbot Bills Stopped Being Bills

On September 10, Governor Newsom signed 13 child-safety laws covering social media and companion chatbots, closing out a legislative package that had only been sitting on his desk since late August. SB 1119, from Senator Steve Padilla with Assemblymembers Buffy Wicks and Rebecca Bauer-Kahan, tightens the companion-chatbot children's-safety rules first established by SB 243 back in October 2025. AB 1856 adds a standardized age-verification signal that apps have to honor; AB 2246 restricts how children can access a covered online product or feature in the first place; AB 2 raises civil penalties for injuries to children, reportedly as high as $1 million per child where a company is found to have caused harm. Most of the package takes effect January 1, 2027, which sounds distant until you remember that "risk assessment before a new chatbot rollout" is not a checkbox anyone bolts on the week before a deadline — it's a review step that has to already exist in the release process by then.

The detail worth sitting with is what changed between the two stories. In late August this was a stack of bills with a signature still pending. Now it's law, with a bill number, an effective date, and a dollar figure attached to getting it wrong. For any team running a chatbot that a minor could plausibly reach — and "plausibly reach" is doing a lot of work in a country where geofencing California's population out of a consumer product usually isn't viable — the risk-assessment-before-rollout requirement is the one to build a process around now, not in December 2026.

2. Salesforce Renamed the Problem: Agentforce Got Seven Employees

On September 11, Salesforce rolled out Agentforce 360 with seven named agents, each scoped to one job: Casey takes customer service across voice, SMS, WhatsApp, and web chat; Paige handles internal IT and HR requests through Slack; Carter helps shoppers compare products and checks them out; Piper qualifies inbound B2B leads; Marshall runs back-office supply-chain work with deterministic execution and a full audit trail; Fin covers customer-experience workflows across channels. Six are generally available now. The seventh, Hunter, is still in pilot — an outbound sales agent built on what Salesforce calls a long-horizon runtime, meant to keep working a pipeline over weeks instead of resetting at the end of a session.

The naming isn't just branding. It's an answer to a real integration problem: a company that deploys "an AI agent" for support and a separate one for sales and a third for internal IT ends up with three black boxes that don't share context, permissions, or an audit log. Giving each agent a fixed scope, a job title, and a place on a shared Customer 360 data layer is Salesforce's bet that the fix for agent sprawl is org structure, not a smarter single model. Hunter's long-horizon runtime is the most interesting piece precisely because it's the one still in pilot — persisting goal state across weeks is a much harder reliability problem than routing a support ticket, and it's the part of this launch most likely to reveal where the approach breaks.

3. The Data Explains Why This Had to Happen

New contact-center survey data landed almost simultaneously with the Agentforce rollout, and it's a good explanation for why "seven named agents on one control plane" is the pitch right now rather than a year ago. 88% of contact centers report using AI in some form, but only 25% have it fully integrated into daily operations — the other 63% have deployed something without it actually running the floor. 64% of enterprise CX teams ran an agentic AI pilot this year, and only 27% got even one channel to full production. The average organization is running 3.9 different contact-center technologies, with only 3% on a single unified platform, and only 21% report having a mature agent-governance model despite three-quarters planning to expand agentic deployment within two years.

That's a fragmentation problem wearing an adoption-statistics costume. Companies aren't struggling to get a chatbot to answer a question — they're struggling to get five different chatbot-shaped things, bought at different times from different vendors, to share a customer record, a permission boundary, and an escalation path. A vendor consolidating seven scoped agents under one governance layer isn't solving a model-quality problem. It's solving the 3.9-platforms problem, and the fact that it's shipping now, not eighteen months ago when agentic pilots first got hyped, is the tell that the market finally noticed integration was the actual bottleneck.

The Same Shape, From Two Directions

A regulator writing "assess the risk before you roll out a new chatbot" and a vendor writing "here are seven agents with fixed scopes on one shared data and governance layer" are converging on the same requirement from opposite motivations: know what each conversational surface is allowed to do, to whom, before it talks to anyone. That's a property you can build into an architecture cheaply now, and it's expensive to retrofit once SB 1119-style risk assessments are mandatory and your five vendors' agents don't share an audit trail:

interface AgentScope {
  agentId: string;              // 'support-casey', 'sales-hunter', ...
  allowedActions: string[];     // capability allowlist, not a blocklist
  minorSafe: boolean;           // passed a COPPA/SB1119-style risk assessment
  auditLog: (event: AgentEvent) => void;
}

function beforeAgentTurn(scope: AgentScope, userAge: number | 'unknown') {
  if (userAge === 'unknown' && !scope.minorSafe) {
    throw new ComplianceError(scope.agentId + ' requires age verification before responding');
  }
  scope.auditLog({ agentId: scope.agentId, ts: Date.now(), event: 'turn_started' });
}

Cheap to add per-agent scoping and an audit hook today; expensive to bolt a risk assessment onto seven black-box vendor integrations after the fact.

Suggested visuals: a timeline from SB 243 (October 2025) through the late-August bill package to the September 10 signing, marking the January 1, 2027 effective date; an org-chart-style diagram of the seven named Agentforce agents grouped by function with Hunter marked "pilot"; and a simple bar chart contrasting the 88%-deployed / 25%-fully-integrated gap against the 64%-piloted / 27%-in-production gap, since the two numbers make the fragmentation argument better than a paragraph does.

— Maya