← Back to blog·Trends·4 min read

AT&T Makes Its 500 Agents Earn Their Autonomy, AWS Sells Cited Answers as a Connector, and New Jersey Moves to Treat Chatbots as Products

AT&T says nearly 500 agentic workflows run in production, each supervised by a person until it reaches mid-to-high-90s accuracy. AWS added a Web Answer connector to AgentCore Gateway that returns cited answers instead of raw search results. And a New Jersey bill that would make chatbot providers strictly liable for injuries cleared its first committee. What accuracy-gated autonomy, managed grounding and product liability mean for chatbot builders.

By Maya Brennan · Writer, Smillee AI
November 18, 2026

This week's stories are about trust: how a big company decides an agent can run without a person watching, how much of the work of citing sources a cloud vendor will now do for you, and what a state wants to happen when a chatbot hurts someone anyway. The figures below come from company talks, vendor documentation and a bill that is still moving, so treat specifics as provisional.

1. Autonomy as Something an Agent Earns

SiliconANGLE reported on November 14 that AT&T has nearly 500 agentic workflows in production and plans to roughly triple that in 2027. The numbers come from Luke Corbin, AT&T Business's assistant vice president for digital transformation and AI, speaking at an AWS event in Boston. He put the load at more than a million transactions a month and about 45 billion tokens a day.

The scale is less interesting than the rule behind it. Every workflow starts with a person reviewing its output. An agent only runs on its own once its accuracy reaches the mid-to-high 90s within an accepted risk profile, and monitoring then watches for drift. Corbin claimed production AI investments return about five times their cost within a year. AT&T hasn't published how it calculates that, or how it scores "acceptable risk".

Most teams treat human review as a launch phase they eventually remove. AT&T treats it as the default, and autonomy as a status each workflow earns and can lose. That is easy to encode:

type Mode = 'supervised' | 'autonomous';

// Promote only on enough reviewed runs; demote as soon as recent accuracy drifts.
export function nextMode(mode: Mode, reviewed: number, accuracy: number, recentAccuracy: number): Mode {
  if (mode === 'supervised' && reviewed >= 500 && accuracy >= 0.96) return 'autonomous';
  if (mode === 'autonomous' && recentAccuracy < 0.93) return 'supervised';
  return mode;
}

The thresholds are placeholders. The hard part is having labeled outcomes at all, so log every human correction from day one.

2. Grounding Becomes a Managed Connector

AWS documented a Web Answer connector for Amazon Bedrock AgentCore Gateway. AgentCore's release notes list it as generally available in us-east-1. You add it to a gateway as a target, and it is called under an AWS-owned service role. Instead of raw search results, it returns markdown with inline citation markers plus a structured list of the cited URLs. Version 1.1.0 adds domain include and exclude lists set on the target. The same release notes add an OverheadLatency CloudWatch metric that separates AgentCore Runtime's own share of a request's latency from your agent's.

Anyone who has anchored citations by hand knows the appeal: when a model writes its own reference numbers, they can point anywhere. A connector that returns the text and its sources together moves that work to the vendor. It also gives you the vendor's idea of a good source. Domain allowlists help. Still, check what comes back: confirm that every marker resolves to an entry in the source list, and decide what your UI shows when one doesn't. The latency metric matters for the same reason. When a grounded answer is slow, you need to know whether the time went to the platform, the search or the model.

3. New Jersey Wants Chatbots Treated as Products

New Jersey Assembly Bill A5607 was introduced October 5 and cleared the Public Safety and Preparedness Committee on October 19. It now sits in Appropriations and hasn't passed either house. As introduced, it would treat a chatbot as a product under product-liability law. A provider would be liable for injuries even if it used all reasonable care, and even if it never distributed the chatbot to the user directly. The state could seek $5,000 per violation, and a harmed user could recover actual damages or $5,000 per violation, whichever is greater. Per a summary of the committee vote posted by a state lawmaker, providers couldn't sell a user's chat logs or keep them longer than two years unless another law requires it. EPIC supports it and says it is based on a model bill EPIC wrote with Consumer Federation of America and Fairplay.

The bill may change before it passes, if it passes at all. The direction matters anyway: it would move chatbot harm from negligence, where you can argue you were careful, toward strict liability, where being careful isn't a defense. Builders should keep evidence that holds up either way: versioned system prompts, safety-filter decisions logged per turn, and retention you can actually enforce.

Conclusion

The common thread is accountability. AT&T measures each workflow before letting it run alone, AWS will take on the work of citing sources if you trust its choices, and New Jersey wants providers to answer for harm regardless of intent. For builders, that means recording human corrections as training data for promotion decisions, checking managed citations instead of trusting them, and keeping per-turn safety logs with enforceable retention.

Suggested visuals: a state diagram of a workflow moving between supervised and autonomous modes, with promotion and drift thresholds on the edges; a request trace for a grounded agent turn splitting latency into platform overhead, web lookup and model time; and a table comparing negligence and strict liability for a chatbot provider: what a plaintiff must prove and which defenses still apply.

— Maya

Frequently asked questions

How does AT&T decide when an AI agent can run without human review?

According to a November 14, 2026 SiliconANGLE report on remarks by AT&T’s Luke Corbin, every production workflow starts with human supervision and runs autonomously only after reaching mid-to-high-90s accuracy within an accepted risk profile, with ongoing monitoring for drift.

What does the AgentCore Gateway Web Answer connector return?

AWS documentation says it returns markdown text with inline citation markers and a structured list of the cited source URLs, rather than raw search results. Version 1.1.0 and later support domain include and exclude lists.

Is New Jersey’s chatbot liability bill A5607 law?

No. As of mid-November 2026 it had cleared the Assembly Public Safety and Preparedness Committee on October 19 and was in the Appropriations Committee. As introduced, it would treat chatbots as products with strict liability and set $5,000 penalties per violation.

Maya Brennan
Writer, Smillee AI

I'm Maya — I write most of what you'll read here. I spent years as a copywriter before I got a little obsessed with what these AI tools can actually do, so now I spend my days poking at chatbots, breaking them, and writing up what's worth your time. Everything here is something I've actually tried. If a prompt didn't work for me, it doesn't make the cut.

Want to try any of this?

Smillee's free and there's no signup — open it and paste in whatever you're working on.

Start chatting →

More from the blog