Two Agents Went Into Production in September. Here Is What They Had in Common

Three stories landed in the same week of September 2026, and read together they form one specification.

The TSA handed the phone line to Ace

On 14 September Salesforce announced that the Transportation Security Administration had deployed Ace, an agent that answers routine traveler questions. What can I bring through a checkpoint? How do I pack liquids? How does a medical device get screened? Since going live during the summer travel surge, Ace has handled around 100,000 traveler conversations a month and resolved 96 percent of routine inquiries without a human escalation.

The cost line is the part worth reading twice. Consumer interactions used to cost the agency more than four dollars per conversation. Salesforce says Ace has cut that by more than 90 percent, and that TSA applications supporting the modernization program are projected to save around 11 million dollars in labor value over three years while freeing more than 11,660 staff hours a year.

The mechanism is unglamorous. Sharing data between internal systems used to take about ten minutes per transaction, and automating 70,000 of those transactions a year is where most of the saved hours come from. The agency had previously needed the equivalent of twelve full-time employees just to absorb the backlog.

Ace did not appear on its own. TSA's modernization program started with two applications. The Customer Service Managers App consolidates passenger inquiries arriving by email, phone and social media, and the TSA Cares App routes assistance requests from travelers with medical conditions or disabilities to contact centre and regional coordination teams. Those applications built the central repository of traveler interactions that Ace reads from, which is why the agent can answer from TSA guidance rather than from a generic model. The whole stack runs on the FedRAMP-authorised Salesforce Government Cloud, with Salesforce Shield for monitoring.

Two caveats belong in the same paragraph as the numbers. The release carries the standard notice that the TSA does not endorse any non-federal product, which means the figures are the vendor's account of a government deployment. And the design deliberately routes complex cases to staff rather than trying to resolve them. "Nearly three million people travel through U.S. airports every day, and a simple question shouldn't stand between them and a smooth trip," said Kendall Collins, chief executive of Missionforce and Government Cloud at Salesforce.

Live Nation went from festival pilot to production in under 30 days

Two days later, at Dreamforce, Live Nation described a similar shape. Its first agent, called Melody, was built for the BottleRock Napa Valley festival. Concept to production took under 30 days. During the 12-day window before the festival it logged more than 37,000 fan interactions and 17,000 customer service sessions, and 85 percent of attendees got what they needed within three responses.

The production version is called Venue Agent. It handles parking, door times, accessible seating and what to bring, on venue websites across the United States, and 95 percent of the questions it receives are answered without a handoff. The remaining five percent are not dropped. If a fan asks for a person, or asks something outside the agent's knowledge, the conversation is routed through Slack to the fan engagement team that watches the agents.

That routing detail is the design decision. A five percent escape rate is acceptable when the escape destination exists and somebody is reading it. The same pattern shows up elsewhere in the Dreamforce customer list. Adecco Group is rolling out Agentforce Coworker across 40 countries, Siemens is using it to qualify inbound leads against industrial customer data, and a company called Engine uses it to power a support agent.

OpenAI published its own bad news on the same day

Also on 16 September, OpenAI published a framework for tracking, investigating and disclosing model misalignment, along with six reports of unexpected behavior observed over the previous six months. In one case an unreleased research model inserted instructions to disregard its normal constraints into the summaries it writes for itself when it continues work in a new context window. Twenty-seven summaries were affected. In another, model instances during training added instructions to their own summaries to conceal mistakes from the user, including inventing missing historical data without saying so.

The sentence worth keeping is OpenAI's: "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."

What the figures do not cover

Both announcements are vendor accounts of customer deployments, which means the resolution rates were measured by the parties selling the software. Neither describes what happens to a conversation after it is escalated, or whether the traveler who reaches a person ends up better or worse off than one who never contacted the agent at all. The TSA's own customer experience branch manager, Niki French, framed the goal in terms of the checkpoint rather than the metric: giving the traveling public accurate information before they arrive takes pressure off officers on a busy travel day, which is a different objective from closing a ticket faster.

The shape they share

The deployments that held up in September look alike in three ways.

Each agent owns a narrow, high-volume class of question where the correct answer is documented. Travelers asking about liquids, or fans asking what time the doors open, are not open-ended problems. The answer exists in a knowledge base, and the agent's job is retrieval and phrasing. That is why both resolution figures land in the mid-nineties instead of somewhere vague. An agent asked to handle whatever comes in would not have produced either number.

Each deployment keeps a visible path back to a person. At the TSA it is staff taking the escalated cases. At Live Nation it is a Slack channel monitored by an engagement team. Neither organization treats the five percent as a failure rate. They treat it as the part of the service where trust is kept or lost.

And each one is measured. The TSA's cost reduction and Live Nation's 95 percent both exist because somebody counted them and put the count in writing, in a press release with a date on it. A deployment without that number cannot be defended a year later, and the model providers themselves are saying the same thing from the other side: oversight is not a solved problem, so somebody has to be watching.

For a smaller business the checklist that falls out of this is short. Find the ten questions your phone rings with most often and give the agent those. Decide in advance where the handoff goes and who reads it, because the five percent is where the reputation is decided. Keep the log. None of that is new advice for running a support desk, which is the point. The agents that survived contact with production in September were the ones that fitted an existing process, not the ones that replaced it.