California Is Regulating Chatbots Just as Wall Street Stops Funding Bad Agents

Two numbers landed in the same week and neither one is about a model release. California's Legislature passed more than two dozen AI bills and sent them to the Governor's desk. Gartner said over 40% of agentic AI projects will be canceled by the end of 2027. Read separately, one is a policy story and the other is a market story. Read together, they describe the same maturing phase from opposite directions: the rules are catching up to what chatbots actually do, and the money is getting pickier about which ones are worth building. A third thread — how teams are quietly redesigning the chat interface itself — shows where the surviving pattern is heading.

1. California Sends Newsom Two Dozen Chatbot Bills

California's Democratic-controlled Legislature closed its session having passed 26 AI and social-media bills, several aimed squarely at conversational products, and Governor Newsom has until September 30 to sign or veto them. SB 903 would bar AI systems from providing or marketing themselves as "therapy," setting standards for how licensed professionals can use AI assistance without letting the AI stand in for the clinician. SB 1119 updates the state's existing companion-chatbot law, SB 243 — in effect since January 1, 2026 — which already requires covered chatbots to verify user age, periodically disclose their non-human status, and bars them from claiming to be a licensed therapist, lawyer, or physician. Together the bills push California from "disclose you're a bot" toward "prove you're not impersonating a professional and prove the user isn't a minor." Because few consumer chatbots can afford to geofence out California's user base, these requirements are becoming de facto national defaults the way CCPA did for privacy.

2. The Agentic AI Reality Check

At the same time, the market is tightening around which agentic projects survive a budget review. Gartner projects agentic AI spending will grow 141% in 2026, to $201.9 billion, and overtake chatbot spending outright by 2027 — the aggregate trend is still up and to the right. But Gartner also expects more than 40% of agentic AI projects to be scrapped by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls, while Forrester expects enterprises to defer roughly 25% of planned 2026 AI spend into next year as financial rigor kills proofs of concept that can't show results. The ROI numbers explain why: enterprises report expecting 171% ROI from agentic AI, but only 39% can point to any measurable EBIT impact so far. The money isn't leaving the category — it's leaving the thin wrappers. A single LLM call with a system prompt calling itself an "agent" is exactly what this correction is built to cut; something wired to real tools, real state, and a real audit trail is what keeps getting funded.

3. Conversation Becomes the Control Surface

Both pressures point the same direction architecturally. McKinsey now reports 78% of companies have integrated conversational AI into at least one core operational area, and the pattern that's surviving budget reviews isn't a bigger FAQ bot — it's a chat interface repositioned as a control surface that interprets intent, pulls the right context, and invokes real backend actions, with the audit and confirmation steps regulators like California are now writing into law. That means the chat turn itself needs to route to a typed action, not just a token stream:

interface ChatAction {
  intent: string;
  handler: (params: Record<string, unknown>) => Promise<ActionResult>;
  requiresConfirmation: boolean; // e.g. anything financial, medical, or destructive
  auditLog: (params: Record<string, unknown>, result: ActionResult) => void;
}

async function routeTurn(intent: string, params: Record<string, unknown>, actions: ChatAction[]) {
  const action = actions.find((a) => a.intent === intent);
  if (!action) return { reply: 'I cannot take that action yet.' };

  if (action.requiresConfirmation) {
    return { reply: 'Confirm before I proceed?', pending: { intent, params } };
  }

  const result = await action.handler(params);
  action.auditLog(params, result); // the record a regulator — or a postmortem — will ask for
  return result;
}

The confirmation gate and the audit log aren't defensive boilerplate; they're the two pieces that let a chatbot claim it isn't impersonating a professional and let a finance team point to a measurable outcome instead of a vibe. Model-Context-Protocol-style tool registries and multi-agent task boards are converging on the same shape for the same reason.

What This Means for Builders

None of these three stories is about a smarter model. They're about the chat interface graduating from novelty to infrastructure, and infrastructure gets audited, regulated, and cost-justified in ways a demo never was. If you're shipping a consumer-facing assistant, assume California's disclosure, age-verification, and no-professional-impersonation rules are your baseline regardless of where your users sit. If you're building for the enterprise, assume the next budget cycle wants a defensible action log and a measurable outcome, not a bigger context window. The projects that get canceled in the next eighteen months will mostly be the ones that never needed to be agents in the first place; the ones that survive will look less like a chatbot and more like a governed API with a conversational front end.

Suggested visuals: a split-timeline graphic showing SB 243 (effective Jan 2026) flowing into SB 903 and SB 1119 (pending signature, deadline Sept 30, 2026); a funnel diagram illustrating the agentic-AI project pipeline — 141% spending growth in, 40%+ cancellation by 2027 out; and an architecture diagram of the routeTurn pattern showing a chat turn branching into confirm/audit/execute paths.

— Maya