American Express scales gen AI for travel counselors with employee buy-in and multi-LLM flexibility
American Express applies AI for sentiment at scale and agent assist to cut handle time and lift NPS. Agents stay in control with sandboxed testing and a model-agnostic architecture.
How American Express Uses Gen AI to Speed Support, Boost Empathy, and Stay Flexible
Customer expectations are simple: fast answers, useful options, and a human who gets it. American Express is pushing in that direction with two AI initiatives that cut handle time, improve coaching, and keep agents firmly in control.
If you support customers, here's the playbook: start with listening at scale, embed AI where it removes friction, and build your tech to swap models as your needs change.
Two bets: sentiment at scale and agent assist
In 2022, American Express rolled out ML-based sentiment analysis across calls. Today, it scores conversations for about 12,900 support agents in the US, Canada, Australia, New Zealand, and Mexico.
The output feeds targeted coaching on empathy and call quality. Leadership reports a lift in Net Promoter Score tied to this capability, giving frontline teams direct feedback after every call. For a quick primer on NPS, see Bain & Company's overview here.
In late 2023, the company piloted Travel Counselor Assist, a gen AI tool that helps counselors prepare for calls, build itineraries, search deals, find hotels, and generate personalized recommendations for restaurants and activities. It also summarizes interactions for easy follow-up.
By October 2024, this assistive tool reached all 5,000 travel counselors across 19 markets.
Real-world complexity, handled in real time
Requests aren't simple. Counselors see everything from "Can you find a Taylor Swift or Harry Styles ticket?" to "Book me to the Maldives via Dubai, plus a vegan restaurant."
Before assistive AI, agents often put customers on hold or called back later. Now, they can assemble options during the conversation. Shorter calls, fewer holds, and more completed bookings are the reported outcomes.
Employee buy-in first
Change lands better when people shape it. American Express ran focus groups, encouraged agents to experiment, and surfaced ideas from the field.
About 100 automation concepts sit in "test kitchens." Some move forward. Others pause due to high hallucination rates or hard-to-audit code. Nothing gets shelved forever; ideas are revisited as models improve.
Guardrails: test, supervise, iterate
Every automated action is tested in a sandbox with redacted customer data. Human oversight remains in place for all bookings-whether it's a simple ferry ticket or a $1 million, six-month cruise.
The discipline is straightforward: validate accuracy, measure impact, and keep a human accountable for final decisions.
Architecture that avoids vendor lock-in
To keep options open as models change, the team built two layers: an enablement layer (starter code, frameworks, and common recipes for consistency) and an orchestration layer that connects to multiple large language models.
They can swap models per use case without heavy rewrites, staying adaptable as performance and risk profiles shift. Leadership also urges CIOs to be at the table from day one-co-creating use cases with product and support leaders, not reacting after plans are set.
What this means for customer support leaders
- Start with listening: Sentiment analysis gives every agent fast feedback on empathy and clarity.
- Embed AI into live workflows: Prep, search, summarize, and suggest in the flow of the call-no extra tabs, no context switching.
- Keep humans in charge: Use AI to draft; agents approve. This builds trust and protects outcomes.
- Measure the right metrics: NPS/CSAT, average handle time, hold time, first-contact resolution, and conversion (bookings or sales).
- Design for swap-ability: Multi-LLM orchestration avoids lock-in and lets you match models to tasks.
- Test in a sandbox: Redacted data, clear acceptance criteria, staged rollouts, and regression checks.
- Co-create with agents: Run focus groups, log feature requests, and prioritize by call impact.
- Be honest about limits: Track hallucinations and pause features that don't meet thresholds.
A 90-day starter plan
- Days 0-30: Pick one call type with high volume and long handle time. Stand up sentiment scoring and baseline your metrics.
- Days 30-60: Pilot an agent assist that drafts answers, options, and summaries for 20-50 agents. Build a simple review workflow.
- Days 60-90: Add an orchestration layer to test two LLMs on the same task. Expand the pilot, compare results, and set go/no-go criteria.
Analyst snapshot
Industry analysts note that natural-language agents and conversational analytics aren't new-but embedding them into mission-critical travel workflows at scale reduces friction for customers and makes insights more accessible for managers. See Gartner's coverage of AI trends here.
Tools and training
If you're building these skills for a support team, explore role-based learning paths and certifications that focus on automation and agent-assist patterns:
The takeaway: listen better, assist faster, keep humans responsible, and build for flexibility. That's how you improve service today and stay ready for what's next.
