Customer experience leaders are used to hearing that the customer is changing. The shift Andreas Welsch, an AI leadership expert, described in a recent CX Network session is more literal: the customer may soon stop calling, emailing or chatting at all, and send a personal AI agent instead.
Welsch spoke with a CX Network host about agentic AI as part of the Mastering CX AI Agents event, and the conversation kept returning to one theme he called underrepresented: agentic commerce. In his view, personal agents that shop, reorder, cancel subscriptions and file complaints will change the volume and pace of work that CX teams handle.
This article follows that thread, from customer-side agents to security, discoverability and conversion, and then covers the autonomy questions that frame any agentic AI decision in customer experience.
Executive Summary
- Customers are starting to delegate shopping, reordering, cancellations and complaints to personal agents.
- CX teams should expect more tickets, calls and submissions, arriving at a faster pace.
- Security, authentication, payments and disputes sit at the center of agentic commerce.
- Being cited by AI engines is the first step; conversion is the harder second one.
- Start from the business outcome, then choose how much autonomy a task needs.
Key Takeaways
- Agentic commerce could become the first consumer use of AI that feels useful to the general population, according to Welsch.
- When customers send agents instead of calling, scripted deflection works less well.
- An agent on the customer side may call for an agent on the company side to keep pace.
- Hiring and procurement already show the same cat-and-mouse pattern, with AI on both sides.
- Authentication, traceability and the ability to dispute an agent’s purchase are central, not optional.
- Welsch reported roughly 4,000 AI citations of his own website over a three-month period after working on generative engine optimization (GEO).
- Citations are not conversions: schema, query research and measurement decide what happens next.
What is agentic AI?
Agentic AI describes systems that work toward a goal over time instead of answering one prompt at a time. Welsch contrasted it with earlier stages of automation. Rule-based systems follow fixed steps. Assistants respond when a user asks a question. Agents take a task or goal, run for minutes at a stretch, find sources, produce a draft and return with an analysis. For customer experience, the practical question is how much of that autonomy a given use case needs, not whether a product deserves the label.
When the customer sends an agent
Welsch called agentic commerce a possible watershed moment for consumer AI. Much of the public conversation about AI centers on security risks, job displacement and dystopia, he noted, and a positive vision of what AI can do for people has been missing.
Personal agents, marketed as assistants, are an early answer. Welsch pointed to social media posts from people who say an agent found savings, canceled subscriptions or booked a trip. Over the next two to three years, he expects this to move from novelty toward the new normal in how customers deal with brands.
Key Insight: Welsch explained that when a customer sends an agent to find the best product, reorder an item, cancel a subscription, file a complaint or get a replacement, the brand is no longer serving a person on the phone. It is serving a delegated task, and its CX organization has to handle that.
What it means for CX volume and pace
The operational consequence, according to Welsch, is higher volume at a faster pace. A human agent reading a script and trying to deflect may not work as well against a persistent counterpart that does not tire.
He advised CX leaders to prepare for more tickets, questions and calls, whether AI agents place them on a customer’s behalf or submit information directly. His answer to whether a company needs agents of its own was a qualified yes: it may be the new reality for keeping up.
The cat-and-mouse pattern
Welsch placed this inside a pattern he has watched for years. About eight or nine years ago at SAP, he worked on systems that matched resumes to help recruiters sift through hundreds of applications for high-volume roles. Today, candidates use ChatGPT, Claude and low-cost tools to optimize their resumes.
Procurement followed the same arc. Drafting a request for proposal and answering it once took a great deal of time, and AI now handles much of the work on both sides. Humans still deliver the service and sign the final agreement.
Key Insight: In Welsch’s words, it is a cat and mouse game. Each side adopts AI to keep pace with the other, and the lesson for CX is to expect the same dynamic in customer contact, then decide deliberately where the human stays accountable.
Security, authentication, payments and disputes
Asked about cybersecurity and privacy, Welsch said they are central to each step. A user has to authenticate and authorize an agent to act for them. The agent has to authenticate toward a website. Payments have to be made on someone’s behalf and remain fully trackable and traceable.
He also raised disputes: a customer should be able to say that their agent made a purchase in error, much as an executive assistant acting in a boss’s name might make a mistake. Data use is the other open point, because agents that act on personal data could also use it to improve products or sell more ads. In his words, security is “absolutely critical to all of this.”
Staying discoverable to agents
If agents shop and transact, brands need to be found by them. In the spring, Welsch set out to improve his own website so that both agents and humans could learn about him and his services. Over three months ending in late September, he reported roughly 4,000 AI citations that did not exist before, split about evenly between Bing and Copilot on one side and Claude, ChatGPT and others on the other.
His point was that the same work should help any brand show up. He described generative engine optimization (GEO) and answer engine optimization (AEO) as additions to search engine optimization, and noted that Google and Bing have both added generative AI features that show what people and agents ask.
From citation to conversion
Welsch cautioned that “showing up is only one part of the equation.” The next step is getting a customer, or the customer’s agent, to take action. That takes more work: adding schema to websites so agents can process information as structured data, understanding which phrases agents search for, and using services that surface the queries people and agents ask.
Key Insight: Welsch argued that being discoverable is the first step and conversion is the next. A brand that earns citations still has to make the sale, keep the customer returning, and solve the attribution challenges that CX and marketing leaders have faced for years.
Matching autonomy to the problem
Alongside agentic commerce, Welsch returned to a theme from his work with clients: ask how much autonomy a use case needs before asking whether a product is truly agentic. He described a ladder from deterministic rules to probabilistic patterns, to assistants, to agents working with more autonomy toward a goal.
Building a multi-agent system to run his own business over the summer taught him where agents fall short, especially when one agent passes information to another and the second misreads it. His conclusion was that a workflow with clear steps is often a sound foundation, with AI added where data is unstructured or decisions involve uncertainty, such as customer questions, complaints and disputes. Throwing AI at an entire business function adds complexity, a cost he said few vendors discuss.
For vendor conversations, he suggested keeping the discussion about operations and business benefit: how a chatbot knows the right answers, connects to a knowledge or ticket system, and stays current as products and markets change. He also described the watermelon problem, in which status turns from red on the ground to green at the top, so a contact center that looks AI-ready in a report may frustrate the first customer who reaches it.
Leadership Implications
- Plan CX capacity for agent-originated contact: more tickets, calls and submissions at a higher pace.
- Define how your organization authenticates customer agents, records who acted for whom, and handles disputes over an agent’s mistake.
- Run a GEO and AEO review of your brand’s presence in AI answers, and decide which new metrics to track.
- Treat citations as the first step and design the path to conversion, including structured data on your site.
- Choose the lowest autonomy that solves the problem, and call your own service line to see what customers experience.
Why this conversation matters
The CX Network session brought a CX audience a view of agentic AI that starts from the customer’s side. For leaders in customer experience and marketing, it connects Welsch’s work on AI leadership, agentic AI adoption and discoverability to a near-term operating question: what happens when the person on the other end is an agent.
Original source: Andreas Welsch at CX Network’s Mastering CX AI Agents
Conclusion
Welsch’s message to CX leaders was practical. Prepare for agentic AI on the customer side, take security and disputes seriously, measure whether customers actually get their answers, and treat discoverability as a start rather than a finish. His closing advice was to stay on top of personal assistants, review how your brand shows up in AI results, and agree on the metrics that matter next.
Related Reading
- Agentic Commerce Infrastructure: 4 Questions for CX Leaders
- Agentic AI Security: Five Layers and Agent Commerce
- How Leaders Adopt Agentic AI Responsibly
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is customers delegating shopping and service tasks to personal AI agents that act on their behalf. Welsch described agents that find products, reorder items, cancel subscriptions, file complaints and arrange replacements, so brands deal with a delegated task instead of a person calling or typing.
How will agentic commerce change CX volume?
CX teams should expect more tickets, questions and calls arriving at a faster pace, according to Welsch. Agents do not tire of waiting or repeating a request, so scripted deflection by a human agent may work less well. Capacity planning should assume agent-originated contact.
Do CX teams need AI agents of their own?
Welsch answered that you might need an agent, and called it the new reality for keeping up in a cat and mouse game. The decision should follow the business outcome, including where a human stays accountable for the final response and for complex or sensitive cases.
Why is security central to agentic commerce?
Security touches every step: how a user authenticates and authorizes an agent, how the agent authenticates to a website, how payments are made and traced, and how a purchase made in error can be disputed. Data use matters too, since agents act on personal data.
What is the difference between GEO and AEO?
Welsch described generative engine optimization (GEO) and answer engine optimization (AEO) as additions to search engine optimization. Both concern how a brand is found and cited when people and agents ask AI systems questions, rather than only how it ranks in classic search results.
Is being cited by AI engines enough?
No. Welsch said showing up is only one part of the equation. A citation does not guarantee a visit or a sale, and attribution remains hard. The next step is conversion: structured data on the site, understanding agent queries, and making the purchase path work.
How did Welsch measure his own GEO results?
He reported roughly 4,000 AI citations over a three-month period ending in late September, split about evenly between Bing and Copilot and other engines such as Claude and ChatGPT. These were citations of his website that did not exist before his spring work.
What should a CX director do in the next 90 days?
Welsch gave three actions. Learn about personal assistants and how they work. Review how your brand shows up in GEO and AEO results, and run an analysis inside your organization. Agree on the new metrics you need to track.
How much autonomy does a CX use case need?
Start with the business outcome. Welsch described a ladder from rules to assistants to agents and advised using traditional workflows where steps are clear, adding AI for unstructured data and decisions under uncertainty, because full automation adds complexity and cost.