Free tools Windows power users keep installed
One-click scans. No signup required.
American Express’s often-cited AI results describe two different outcomes, not a single company-wide efficiency gain. Amex reported that its generative-AI IT chatbot became 40% more capable of resolving queries without transferring users to a live engineer; separately, more than 85% of travel counselors said their AI tool saved time and improved recommendation quality. Neither figure means that total IT escalations fell by exactly 40% or that travel productivity or revenue rose by 85%.
The results, shared by Amex’s chief technology officer in an April 2025 VentureBeat interview, come from targeted employee-assistance workflows. A 2026 shareholder letter describes a broader and expanding AI program, but does not independently verify those earlier metrics.
What Amex’s headline AI metrics measure
| Use case | AI’s role | Reported result | What the figure does not establish |
|---|---|---|---|
| Internal IT support | Interactive troubleshooting in a chatbot | Amex reported a 40% increase in its ability to resolve queries without transferring users to a live engineer after the generative-AI chatbot launched in October 2023. | It is not necessarily a 40% reduction in all company IT escalations; the absolute baseline and case volume were not disclosed. |
| Travel counseling | Research and synthesis to help counselors prepare recommendations | More than 85% of surveyed travel counselors said the tool saved time and improved recommendation quality. | It is not an 85% increase in bookings, revenue, speed, or counselor output. |
These are company-reported measures from distinct workflows, so they should not be added together or treated as comparable productivity statistics. The available reporting does not give an independently audited result, financial return, headcount reduction, or customer-retention impact.
How the IT chatbot changed support
Amex said its earlier support approach used traditional natural-language-processing systems, including BERT-based technology. The later generative-AI workflow was described as moving beyond a page of knowledge-base links: it asks clarifying questions, offers step-by-step troubleshooting, checks whether a proposed fix worked, and continues with another remedy or transfers the unresolved issue to a live engineer.
#1 Best Overall
- Used Book in Good Condition
That loop matters because IT problems are often underspecified. A user might report that a device cannot connect; a useful support interaction needs to establish what is failing, guide the user through an approved check, and learn whether the step changed the outcome. The reported result supports guided resolution of routine queries, not a claim that the AI autonomously repairs every device, network, or infrastructure fault.
Human escalation remains part of the workflow. A lower transfer rate is valuable only if users can still reach an engineer when troubleshooting fails or the issue requires access or expertise the bot does not have.
How Travel Counselor Assist supports recommendations
Amex’s Travel Counselor Assist is described as a research aid for about 5,000 counselors serving premium customers, including Platinum and Centurion members. The 2025 interview described coverage across 19 markets; the 2026 shareholder letter says travel counselors in 19 countries continue using AI for faster recommendations and insights. Those formulations are related, but a market and a country are not necessarily the same unit.
Rank #2
The tool combines public or web-accessible details—such as venue hours, busy periods, and nearby restaurants—with Amex proprietary information and customer context, including spending-related signals. The counselor interprets that material and personalizes the final recommendation. For a customer seeking a restaurant or itinerary, current facts are only part of the answer: preferences, occasion, constraints, and service expectations matter too.
This is a strong assistance use case because counselors may handle different destinations and requests in succession. AI can reduce the time spent gathering and organizing information; it does not make travel expertise unnecessary. Amex’s reported 85%-plus result is a counselor-reported view of time savings and recommendation quality, not a measured booking or revenue lift.
What Amex has disclosed about its AI controls
The 2025 interview describes a central AI enablement layer with reusable “common recipes” or starter code, orchestration connecting applications to models, and the ability to select models for different use cases. Amex also cited an AI firewall, model-risk management and validation, retrieval-augmented generation (RAG), and prompt-engineering techniques. It said thousands of documents needed ongoing maintenance, validation, and reformatting.
RAG can help an AI answer from approved reference material rather than relying only on a model’s general learned patterns. It does not guarantee that an answer is true. The source documents must be current and correctly permissioned; retrieval must find the right material; and the model must interpret it accurately. Stale procedures, contradictory venue details, or a confident but incorrect troubleshooting step remain possible failure modes.
The public descriptions do not identify the specific generative models, cloud providers, databases, evaluation sets, confidence thresholds, or detailed security architecture. They also do not explain exactly how customer data is passed to the travel tool, how conflicting web details are resolved, or what technical controls separate one customer’s information from another’s. Those implementation details should not be assumed.
Recommended Free Tools
Other reported uses and newer scale
Amex’s 2025 interview described a wider employee-facing portfolio: a colleague help center with a reported 96% accuracy, intent-based search with a reported 26% improvement in responses, and coding assistance. These figures relate to particular tools and measures, not to the accuracy or performance of Amex’s entire AI program.
For engineering, the 2025 account said about 9,000 engineers used GitHub Copilot, mainly for testing and code completion, and reported a 10% developer-productivity increase and more than 85% coder satisfaction. The 2026 shareholder letter gives a later, different measure: AI-assisted development tools had expanded to more than 11,000 engineering professionals, with coding cycle time reduced by more than 30%. Productivity and cycle time are not interchangeable measures, and the later figure should not be read as a restatement of the earlier one.
The 2026 letter also describes AI in customer service, mobile-app search, fraud, marketing, sales, commercial products, and agentic commerce. It says U.S. Card Members make about one million mobile-app search inquiries per month, and outlines work involving conversational agents for legacy interactive voice response as well as travel, dining, offers, payments, and partner platforms. These are broader program descriptions, not proof that every use is fully deployed or autonomous.
On program breadth, Amex told VentureBeat it had initially identified roughly 500 potential use cases and narrowed focus to about 70 at various implementation stages. In 2026, the company described having explored hundreds of use cases and providing leading AI tools to nearly all colleagues globally. Neither statement should be converted into an exact count of production AI systems.
What the results do—and do not—tell other companies
Amex’s example is most useful as an operating pattern: put AI inside a defined workflow, connect it to relevant information, preserve a path to human judgment, and measure whether the task is completed better. It does not prove that the same results will transfer to another bank or enterprise, where data, processes, employee adoption, and risk controls may differ.
- Start with a bounded, repeatable task. Routine IT questions offer recognizable diagnostic paths; high-stakes or unusual cases need a clear route out of automation.
- Ground answers in maintained sources. A polished answer based on obsolete documentation can cause more harm than a slower search result.
- Design escalation as a success path. Track whether people get a correct resolution, not merely whether the system keeps them from opening a ticket.
- Measure distinct outcomes separately. Resolution without transfer, accuracy, time saved, cycle time, rework, satisfaction, and customer outcomes answer different questions.
- Keep sensitive decisions supervised. Personalization from customer context can be useful, but access, purpose, and human approval boundaries require explicit governance.
- Account for generated-code risks. Faster completion does not remove the need for code review, security testing, and maintenance standards.
- Treat knowledge operations as infrastructure. Document ownership, freshness, permissions, and validation are prerequisites for dependable retrieval—not cleanup to defer until after launch.
For financial-services readers, the central lesson is not that generative AI automatically produces a fixed efficiency gain. It is that narrowly designed colleague-assistance tools can improve specific work when they combine useful data, an interactive process, measurable outcomes, and human oversight. Amex’s published figures are promising company claims, but the disclosed evidence is not enough to calculate an enterprise-wide return on investment.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
