4. The war over the agent layer
If whoever owns the agent owns the customer, then every platform with something to lose should be fighting for that layer right now. They are. In China.
Late last year ByteDance, the company behind TikTok, put its Doubao assistant inside a phone. The Nubia handset, sold by ZTE, handed the agent the run of the device, reaching into messages, payments and preferences across every app. Chinese consumers loved it. The first 30,000 units sold out and resold at a premium. Within days, Yu reports, Tencent's WeChat, Alibaba's Taobao and Alipay all moved to restrict ByteDance's agent. The official reason was security. The real reason was survival. If one agent could open any app on the phone, the customer no longer belonged to the app. The customer belonged to whoever owned the agent - that’s the nub of the argument.
For twenty years the platform wars were fought over attention: the search result, the scroll, the ad slot. This one is fought over execution, over who is allowed to act on the customer's behalf. Greeven makes the same point from the other direction. The real control point, he argues, is migrating away from the interface and downstream into orchestration, into who governs how decisions get made and executed. Most companies are still fighting over the screen. The screen is no longer where it is decided.
You might tell yourself this is a China story, that the West is years behind and the pipes are not laid. You would be partly right, but it would not save you. When OpenAI let people check out inside ChatGPT, Yu reports, Walmart found those purchases converted at a third of the rate of its own site, because the delivery estimates and shipping costs were wrong, so Walmart walked away. Read that as reassurance if you like. Or read it for what it is. The agent layer in the West is not failing because the idea is wrong. It is failing because the data underneath it is not clean enough for a machine to trust yet. That is a problem brands fix, and the ones that fix it first get onto the shelf first.
5. The bank becomes a set of services
Financial services shows this most clearly, though it applies to anyone reading. To serve a customer well, an agent has to be objective. It has to compare your mortgage against the one next door, your savings rate against the better one, your insurance premium against the cheaper equivalent cover. An agent that favours the company that built it is not serving the customer, and customers will work that out fast. So the useful agent, the one people actually delegate to, cannot belong to you. It belongs to the customer, or to whoever the customer trusts to hold it.
Follow that through. If the agent is objective and it sits between you and your customer, then you are no longer the relationship. You are a supplier the agent calls when your terms win. A credit card becomes a lending action chosen for that moment. An investment platform becomes a vendor picked on its rate and risk. The bank stops owning the end-to-end experience and starts competing as one reliable provider of tasks among many, on availability, performance and price. The customer may never log in again. Their agent will.
This is the part that should unsettle a loyalty director more than any traffic chart. Loyalty was never really stored in your CRM. It was stored in the customer's inertia, and you read the inertia as devotion. An agent has no inertia. It re-checks the market on every transaction and re-decides. So loyalty stops being a thing you own and accumulate. It becomes a verdict the agent reaches again and again, on the evidence, every time.
The MyCFO paper from Trusted Agents poses the question that follows, and it is the one worth taking into a leadership meeting.
If we weren't already doing it this way, is this how we would start?
6. Delegation, trust and context
So what do you actually do, beyond worry. At Trusted Agents we work three questions, and together they are the machinery that lets an agent act for a customer at all. Delegation, trust and context. Get them straight first, because most of the panic in the market comes from blurring them.
Delegation is authority. Has the customer actually told the agent it may act, and to what limit. A valid delegation is what makes the transaction the agent attempts a real one rather than a guess or a fraud. For you, the question is whether you are ready to transact with an agent that holds a customer's mandate instead of the customer themselves. Most B2C systems today assume a human is on the other end and break when they are not.
Trust is identity. Know Your Customer (KYC) has a sibling now, Know Your Agent (KYA). When an agent presents itself at your checkout or your API, do you know which agent it is, and who stands behind it. An agent with no verifiable identity is one you cannot safely let act, however good its intentions look. The brands that get this wrong will either block legitimate agents and lose the sale, or wave through ones they cannot account for and carry the liability.
Context is data, and it has moved. Your CRM holds context you assembled from the outside, purchases, clicks, the things a customer almost bought. The agent holds context the customer hands it directly, in plain language, including the things they would never have told you. That store now sits on the customer's side of the table, not yours. The question is what your relationship is worth once the context that defined it lives somewhere you do not control.
Settle those three and a fourth thing follows, which is what the agent actually judges you on once it knows who you are and is cleared to act. This is Greeven's list and Yu's four numbers, the same point reached from two directions.
- On-time delivery rate.
- Refund speed.
- Time to resolve an exception.
- Error rate in your product data.
Ask your team to pull them this week. If it takes longer than a day, you have learned something, which is that your reliability does not yet exist in a form a machine can read. Everything you spend on storytelling assumes a human is reading the story. The agent is not reading it. It is checking the numbers, and right now it may not be able to find yours.
Three questions for the next leadership meeting. Are our systems ready to accept a validly delegated agent acting for a customer, or do they assume a human every time. Can we verify which agent is acting and who stands behind it, before we let it transact. And if an objective agent compared us against our three closest competitors today, on the numbers alone, would we make its shortlist. The one from the MyCFO paper sits underneath all three, and it travels well beyond banking: if we weren't already doing it this way, is this how we would start.