TSA's AI Agent Handles 100,000 Conversations a Month

The Transportation Security Administration says its new AI agent handles 100,000 routine traveler conversations a month. The number matters, but the work underneath it is the more useful story for marketers building service agents.

Ace is the TSA’s customer-facing agent. Salesforce said on September 14 that it resolves 96% of routine inquiries without escalation and has cut the cost of consumer interactions by more than 90%. The agency says those interactions previously cost more than $4 each. Nextgov reported that Ace is available on TSA.gov and the myTSA app, where it answers questions from the federal site’s material and more than 550 knowledge articles.

That is a very constrained job. It is not an agent asked to improvise travel advice, resolve an obscure dispute, or carry a conversation beyond its source material. It answers a large volume of repeated questions that travelers already bring to a public service.

Niki French, the TSA’s Customer Experience Branch Manager, described the operating goal in the announcement:

“If we can help the traveling public easily access accurate, timely information, we can help alleviate stress in the field on travel days and increase the TSA’s overall effectiveness.”

The lesson for a marketing or customer team is less glamorous than “deploy an agent.” The useful unit of work is the question that is documented, consistently answered, and safe to escalate when it is not. A returns policy, eligibility rule, pricing question, implementation step, or product-availability check can fit that shape. A vague request to make a bot “more helpful” does not.

The data work came first

The TSA didn’t start with Ace. Salesforce says the agency first created two applications to centralize traveler interactions. One pulls together passenger inquiries from email, phone, and social media. The other connects TSA Cares requests with the agency’s contact center and regional teams. Only then did it add an agent that can respond with guidance and the relevant context.

That sequence is a better planning checklist than most agent demos. Before a service agent gets added to a website or a messaging channel, teams need to know where the approved answer lives, who owns it, when it changed, and what should happen when the agent lacks confidence. A polished conversational layer cannot repair a policy library that conflicts with the CRM, or a support process that relies on a person remembering the answer.

The TSA’s reported efficiency figures are also unusually specific. It projects 70,000 annual transactions will be automated and says that will free more than 11,660 staff hours. It estimates $11 million in labor-based value over three years. MeriTalk’s coverage carries the same reported results. These are Salesforce-published figures, not an independent audit, and the agency hasn’t disclosed the share of its total traveler inquiries that count as routine. They are still more operationally useful than the generic satisfaction claims that usually accompany agent launches.

Paul Tatum, Salesforce’s executive vice president of global public sector solutions, tied the result to that foundation:

“They are delivering a secure, data-driven, and AI-first foundation that scales operations, lowers costs, and protects public trust, even during the most stressful travel times.”

What marketers can copy

The TSA case should not be read as permission to drop an LLM onto every high-stakes conversation. Government travel information is unusually structured, and the public-service context is very different from a brand trying to persuade a buyer. The underlying deployment date is also not public, so the September 14 announcement is a results disclosure rather than a clean launch date.

But the pattern transfers. A customer agent works best when the first version does a small number of familiar jobs, has a clean source of truth, and passes the exceptions to a person who can act. The handoff is part of the experience.

We covered Salesforce’s new Koa CRM reasoning model this week. That announcement was about what model sits behind a CRM agent. TSA’s numbers show the less visible prerequisite: the work and data around the model determine whether an agent is useful to a customer. The underlying building blocks also resemble OpenAI’s work on agents, where the model’s access to real workflows decides what it can do.

For a marketing team, the starting point is simple. Pull the last month of chats, call transcripts, contact-form submissions, and sales questions. Find the 10 questions that recur, identify the approved source for each one, and define the escalation. That is the real first build.