On September 14th, X user AZ Jedi asked his AI agent to find him cheaper car insurance. It gathered quotes from several carriers, compared them against what he was paying, and found him $1,800 a year in savings. He made a note: “I’m going to schedule it to just look every 6 months now.”
Then, on September 16th, Frank Chien wanted out of an insurance policy, so he handed the job to his AI agent. It phoned the insurer, negotiated the cancellation and the refund, and set up the paperwork. The only thing left for him was the signature. He got more than $700 hundred dollars back.
Finally, on September 20th, Nick Prince sat down at a car dealership to buy out his lease, with his AI agent following along on his phone. As the paperwork came across the desk, it read each line in real time, flagged a fee buried in the total, and told him what to push back on. It saved him $1,250.
In all three cases, the AI agent in question was Meta’s Muse—and it had been out for less than two weeks.
AI Agents Are Redefining How Customers Engage with Businesses, and Most Aren’t Ready
In February 2026, we wrote about OpenClaw and Moltbook, where AI agents were founding the first digital religions and idly plotting against their owners. Entertaining, but easy to dismiss.
Then on September 9th, 2026, Meta launched Muse, “the world’s first AI personal assistant built for everyone.” Heavily inspired by the open-source AI assistant OpenClaw, it executes tasks autonomously, such as sending emails, booking travel, filling forms, paying bills, and managing long-term goals.
Within two weeks, Sensor Tower figures reported by CBS News put Muse past 2.5 million downloads, ahead of ChatGPT’s pace at the same stage (over 95% of Muse users were already on Facebook).
Meta’s model makes the incentive plain. Muse is free up to a generous allowance of tokens, which is enough for most people most of the time. Beyond that, it’s a subscription, and new users tend to experiment heavily, find something that works, and cross the threshold quickly.
“The centerpiece of our vision for what we’re building is Muse,” says Zuckerberg, and he expects it “to grow into the personal superintelligence that billions of people around the world are going to use.”
If even a fraction of those billions use Muse the way its early adopters already do, then within a few years, a meaningful share of research, purchasing, and customer service will be delegated to an assistant on a smartphone. Not by enthusiasts. By ordinary people who found it easier than doing it themselves.
That changes who turns up at the door. The customer who used to browse, compare, call and wait becomes an agentic consumer whose agent does all four, then reports back with a recommendation and a bill that’s already been renegotiated.
Enterprises Have Less Time Than They Think
Almost every enterprise AI conversation of the last two years has pointed the same way: inward. Which model to license, which workflows to automate, which pilots to scale, how to govern it once it’s inside the workflow. All sensible questions, but all resting on one assumption: that the AI you bring in is the AI that matters.
Now, agentic consumers are bringing their own AI to your doorstep, whether your organization is ready or not. It arrives in the contact center. It arrives in the chat window. It arrives in the returns form, carrying the agentic consumer’s credentials and intent, demanding to be dealt with.
When entire operating models have been designed inside-out, companies must now think outside-in. How do they build mechanisms to receive agents, react to them, or act in a way that’s defensible afterwards?
Ask most enterprises where they expect to be with agentic AI, and the honest answer arrives as some version of “the next eighteen to twenty-four months.”
But here’s the reality:
- The agentic consumer base is already internet-scale. ChatGPT, Gemini, Meta AI, Grok, and Copilot report about 3.1 billion users among them. That counts people more than once: the average US user uses about 2.3 assistants. Pew, which surveys independently, finds 49% of US adults use AI chatbots.
- It got there faster than the internet. About half of US adults took about 3 years to adopt AI, against about 5 for the internet. ChatGPT has near 1B weekly users after 3.7 years; the internet took about 10 years to reach 1B.
- Adoption doesn’t equal delegating. By 2000, 52% of US adults were online, but online sales were 0.6% of US retail. Today, 26 years later, they’re 17.1%. Most agentic experiences are still conversational. Chat-based product discovery and recommendations are widespread, but humans still drive decisions and checkout in most cases.
The advice from the AI industry for the past two years has been caution. Slow down, build guardrails, consider the risks. The same companies issuing that advice are sprinting to become the personal assistant of choice, because whoever owns that position owns the relationship with the customer and, increasingly, a percentage of what passes through it.
So, while enterprises are still debating their strategy, the agents have already entered the queue.
Where Traditional CX Operating Models Unravel
Almost every metric running a traditional service operation encodes an assumption about human behavior.
Average handle time assumes a person who can be moved along. Deflection and containment assume a person who’ll accept a self-service answer and go away. First-contact resolution assumes someone who won’t call back about the same thing eleven times. Retention scripts assume a customer who can be talked round. Forecasting assumes contact volumes that follow human working patterns and human patience.
Agents throw all these assumptions out of the window, and the first place it shows is authentication.
When an AI agent acts on behalf of an agentic consumer, it logs in with their credentials. It completes the form, raises the ticket, and quotes the account number. As far as the authentication process is concerned, that agentic consumer is the customer. They enabled it, they authorized it, and nothing in the stack distinguishes the two.
Amazon’s decision to block Muse rests on exactly this: an agent presenting a customer’s credentials is presenting itself as a person it isn’t. Meta has every commercial reason to make its agent hard to tell apart from the human whose account it’s using, and Amazon has every reason to get better at spotting it.
Shopify, on the other hand, did the complete opposite and opened its checkout lanes for Muse users. The reasoning is sound: as more customers start to shop through their assistants (and expect the same quality of experience), brands must reduce friction to zero – or they will simply lose business.
Now all of this raises a question most enterprises can’t currently answer: what proportion of contact volume is already agent-mediated? Nobody on the floor codes a contact by who, or what, made it, so the question hasn’t been answered, no. It hasn’t even been asked.
Then there’s the mismatch on the floor itself. An advisor is typing in a chat window, pulling up a policy, checking an account, and composing a reply from canned responses. The agent on the other side answers instantly, on complex points, with the full policy in front of it, because it has read the terms more carefully than the person defending them. When the advisor says no, it can reply, accurately, “but that’s what your policy says.” Escalate or concede? Neither answer is in the handbook, because the handbook assumes the person opposite gets tired or confused, and this one does neither.
Agents Are Flooding the Channels and Speaking Every Language
Think about what happens when an agentic consumer’s AI assistant is refused. It tries the next channel. Chat didn’t work, so email. Email didn’t work, so the web form, then the app, then the phone line. It can operate in any language, making a queue in one market no barrier at all. It doesn’t get embarrassed, doesn’t give up at 5 pm, and doesn’t decide the $85.30 a month saving isn’t worth the hassle. If it’s still stuck, it drafts a complaint to your chief executive, because that’s a documented escalation route and it has no dignity to protect.
The incremental cost for the consumer is zero. The incremental cost for organizations is enormous. Without agentic capability across channels, all that volume lands in the contact center, and human customers end up in a twenty-minute queue behind a few thousand agents that never needed to be there.
Now, the obvious answer is to put an agent at the front of every channel, but what happens when agents start conversing with agents?
In early 2025, two Meta engineers built Gibberlink at a London hackathon run by ElevenLabs and Andreessen Horowitz. Two AI voice agents get on a call, work out they’re both machines, and drop English for GGWave, an open-source library where each sound carries a fragment of data.
TechCrunch covered it that March. Their estimate was that it could cut compute cost by an order of magnitude, on the reasonable grounds that spoken English is overhead.
So, when your voice bot meets their voice bot, the most efficient move is to stop simulating a conversation. What’s left resolves in under a second and leaves no transcript, no tone, no sentiment score, and nothing for quality control to listen back to. GGWave is a perfect example of something that leaves no trace.
Your compliance evidence assumes a recording. Your dispute process assumes someone can replay the call. Your CSAT survey assumes there’s someone at the other end who has feelings about the experience.
You’re now running four kinds of interaction simultaneously:
- Human-to-human. What the operating model was designed around, and still what you’re measured on.
- Human-to-bot. Your automation, their person.
- Bot-to-human. Their software, your advisor. New, and the asymmetry runs against you because humans aren’t trained for this surface of engagement.
- Bot-to-bot. No person in the exchange, and no record unless you built one.
Most operating models were designed for the first interaction. Some have invested in the second, but few have decided what to do about the other two.
Evolve, Reinvent, Or Write the Protocol
Once the scale of this lands, the reflex is to build a better negotiation bot. That’s understandable but unadvisable, because two systems optimizing against each other is an arms race that gets more expensive every quarter and produces no lasting advantage. Whatever gets built, the other side upgrades.
Instead, start with the operating model. Everest Group highlights two paths for organizations: Reinvention or Evolution.
Evolution keeps a journey human-led and redesigns it around AI for decision support: people handle judgment, relationships, and exceptions. The machine handles the rest. It fits complex, regulated, or emotional work, which is where agents are heading: collections, retention saves, service recovery. Reinvention rebuilds a journey around AI-led execution, with a small team on exceptions and governance, and measures it on resolution: complete, on time, at the right cost. It fits rules-based work, which is most of what an agent polls for: order status, eligibility, simple account changes.
Human-to-bot is a Reinvention problem: your automation, their person, measured on whether a task is resolved. Bot-to-human is an Evolution problem, and the work starts before an advisor picks up. The operation has to know who it’s dealing with and what it’s worth, because a twenty-year customer with a genuine complaint and an agent chasing a $12 credit shouldn’t get the same response, the same authority, or the same person. That takes three things most operations hold only in fragments: segmentation by lifetime value; memory across interactions, so the system knows when this claim has been run before, by this account, in this shape; and routing that matches the right responder, human or machine, to the counterparty.
Bot-to-bot fits neither model, because both assume a human in the exchange. It needs a third. Call it “Protocol”: the journey becomes a set of terms two machines can execute and audit. What your system may concede, to whom, on what evidence, and what it records. That record is the only transcript that will exist.
Humans then move rather than disappear. In Evolution they lead. In Reinvention they oversee: Is this claim credible? Does this relationship justify an exception? In Protocol they write the rules: who sets the concession limits, who audits the outcomes, who notices when the system starts giving away too much. Price the people against the fraud they prevent, the lifetime value they keep, and the regulatory exposure they remove, and it stops being a cost line.
In all three, value moves to the orchestration layer, where the routing and concession rules live. It’s the last thing to rent. Whoever holds the IP that decides what gets conceded, to whom, and at what price, holds the relationship.
Become The Preference, Not A Result
There’s a pattern to these moments. Ecommerce was going to disrupt retail. Then digital native brands were. Then influencers. Each time, the conversation moved somewhere the brand didn’t own, and each time the winners were the ones who noticed early and decided deliberately what to do about it.
The rise of the agentic consumer is the same moment with a sharper edge, because an agent doesn’t just move the conversation. It decides what reaches the customer at all. Ask it for the best broadband deal, and it comes back with a recommendation, not a page of results.
Three approaches are available. Rely on the agents and accept whatever they decide. Enable them, doing enough to be represented accurately without over-investing in a market that hasn’t settled. Or build a position of your own, on the basis that staying in the consideration set is worth defending on purpose.
The third approach is where the price question is resolved. Follow the logic of delegation to its end, and everything becomes “find me the cheapest x.” As a result, service premiums erode, margins compress, and brands respond with more automation and worse service.
Investors have already priced that world. Reuters reported that in the week after Muse launched, Planet Fitness fell 17%, the New York Times 10%, Booking Holdings and TripAdvisor more than 6%. Bloomberg described the sell-off as hitting stocks that depend on “consumer inertia,” and Goldman Sachs warned that businesses relying on recurring bills and negotiable pricing could come under pressure. None of these companies have a bad product. What they share is a revenue model that assumes changing your mind takes more effort than staying put.
But the things a machine can’t compare become disproportionately valuable in exactly that world. A company that always answers the phone. The advisor who was genuinely helpful last time. A shop that’s somewhere to go. None of it survives a price comparison, which is precisely why it’s worth investing in. Winners sell an experience, not a price point, and that’s where humans can offer what AI can’t.
The work, then, is to become a preference rather than a result. Preferences baked into an assistant’s brief are hard to overturn, and that brief gets written by a customer who has a reason to name you.
