The customer just hired an intern who never sleeps

Tony Stark never waited on hold. He said "Jarvis, sort it out," and Jarvis did, presumably after a long negotiation with the airline's Jarvis, who had been haggling with the hotel's Jarvis since breakfast.

That is the future of CX in one sentence: the customer stops being the one who talks to your brand. Their agent does. Your carefully tuned chatbot, your IVR tree, your "we are experiencing higher than usual call volumes" are about to be interviewed by a counterpart with infinite patience, perfect recall and not much interest in your brand voice.

For a decade, CX has been the craft of designing for a human on the other end. The next decade is designing for a human's delegate. And the delegate reads the fine print.

Chatbots were the warm-up act

Until now, "AI in CX" meant your agent talking to their human. That is flipping. Meta launched Muse on 8 September 2026, and it hit the top of Apple's US free-app chart within a week. OpenAI's dots and Instinct arrived in the same month, and these agents now book, buy, cancel and negotiate for their owners.

Most still use your front door the way a human would. They load pages, click through forms, and when that fails they ring your support line or open your web chat. They are already landing on brands' AI agents. Two bots, one phone line, no adult supervision.

And your policies were not written for this. A refund rule assumes/hopes the customer eventually gives up or doesn't realise that he/she is eligible for a refund. A personal agent can try five phrasings of "please" to squeeze out a bigger concession, and it will not get tired, embarrassed or hungry.  

The friction is already visible. Amazon cut Muse off from Amazon.com on 20 September, saying the agent did not identify itself while browsing.  

What this does to CX: two customers per customer

Every customer now has a principal (the human who decides) and an agent (the one who executes). Five things break when both show up.

1. The front door splits. Humans and agents need different doors. Send detected agents to a separate channel with its own workflows, so the same request can follow a different path depending on who is asking.  

2. Effort stops rationing demand. Hold music and phone trees quietly capped contact volume. Personal agents remove that friction and will likely grow service demand, not shrink it. People will ask more, about more things.

3. Deflection becomes a vanity metric. Focus on resolution. Deflecting a bot is not a win; it is two systems failing to solve a human's problem. Measure resolution for the person behind the agent.

4. Brand voice loses its audience. Your tone guidelines will be read by a model that has already summarised them. What survives is clear policy, reliable APIs and a graceful handoff. Emotion lives at the escalation, and escalation now runs both ways: the business's agent may need a human, and the personal agent may need its owner.

5. Trust becomes a product feature. Permission to check a balance is not permission to move money. Read-only versus write access, scoped permissions and signed receipts are the new CX hygiene.

The new brief for CX leaders: be easy to talk to, for someone who is not in the room.

Who is building the Jarvis stack

Personal-agent CX is barely a month old, and its standards are only days old. On 6 October 2026, Sierra announced the Personal Agent Protocol (PAP) with Meta, with Genesys, NiCE, Instinct, Shopify, Stripe and Walmart among its launch partners, and Decagon open-sourced PACT. PAP still has no published spec, licence or governing body, with v0.1 due later this month, and OpenAI and Anthropic are not on the partner list.  

Payments are a separate contest between PAP's planned extension and protocols from Visa, Google and OpenAI. We expect the space to converge: PAP, PACT, A2A and MCP overlap enough that identity and consent will likely settle into one or two shared standards, vendors will compete on agent detection, policy control and orchestration rather than on the protocol itself, and incumbents such as Salesforce, Zendesk and Microsoft will follow once a spec and a governing body exist. This is a fast evolving space.  

What CX leaders should do right away?

1. Find the agents already calling. Detection is the first control. Useful signals include device fingerprints, account history, conversational patterns and request cadence, and they make a decent audit checklist for any vendor.

2. Pick a stance on purpose. Blocking agents could risk losing customers to competitors, while waving them in unchecked gives up control and can drive up the cost to serve customers. Amazon chose to block Muse. Most brands will want a middle path.

3. Open a front door built for agents. First-wave agent traffic will arrive through channels you already run, so expose A2A, MCP or plain APIs behind the same identity checks your customers face.

4. Write your scope menu. Decide which actions an agent can take read-only and which need the human's explicit grant, such as orders:read versus orders:cancel.

5. Rewrite policies for a tireless counterpart. Add limits on retries, concession ceilings and a stop rule for loops. Two agents can loop forever if neither is built to stop.

6. Re-cut the scorecard. Track resolution per principal, scope-grant completion and escalation rates in both directions. Retire deflection as a headline number.

7. Build behind an abstraction. PAP v0.1 is due this month, PACT is live, and A2A and MCP already exist. They overlap more than they collide, so keep your integration layer swappable.

The punchline

Every era of CX has had its interface: the counter, the call centre, the app, the chat window. The next one has almost none, just an agreement between two machines about what a human is allowed to want.

Jarvis never needed a brand to be charming. He needed it to be reachable, honest and consistent. The winners of this era will be the companies whose agents can say yes safely, say no consistently, and hand over to a human when it matters.

Somewhere right now, two agents are politely stuck in a loop. Make sure yours knows when to hang up.