The Dual CX framework
CX has two designers now.
You are only in control of one. Your customers are the other designer now, arriving with their own AI, on their own terms. We are a leading resource for leaders getting their operation ready for customer AI agents.
This isn't a forecast
Your customers arrived with AI. They didn't wait for permission.
Real behaviour, pasting your emails into ChatGPT, drafting complaints, comparing your tariffs, and increasingly letting the AI place the call. The examples below are things your customers have already done to you. Most of it never surfaces in your CX analytics.
I pasted the brand's email into ChatGPT and asked it to write a stronger reply.
I had Claude summarise the 14-page Ts&Cs before I bought the policy.
I used Gemini to compare three energy tariffs and tell me which to switch to.
I let ChatGPT draft my complaint, then sent it word for word.
I asked an AI to find the right return policy for my situation. It did.
I told my AI to call the company. It passed verification. They didn't notice.
The pace
The earliest sign was two years ago. Most of it happened in the last twelve months.
Your customer service operations were designed for a world before any of this existed. Here's how we got from a chatbot lawsuit to AI agents calling your contact centre, and where it's heading next.
MCP
2M → 97M
monthly SDK downloads, launch to March 2026
Digital Applied
A2A
150+
backing organisations in year one
Linux Foundation
OpenClaw
~388K
GitHub stars in nine months
GitHub
Meta Muse
2.5M
downloads in 13 days
Sensor Tower via CNBC
- February 2024
The Air Canada chatbot ruling
A Canadian tribunal rules Air Canada bound by misleading information its chatbot gave a grieving customer. The first major case law on AI liability in customer service.
- October 2024
Claude can use a computer
Anthropic releases the first AI that can click, type, and navigate websites, like a human user, but faster. The capability layer for customer-side automation goes live.
- November 2024
Anthropic launches MCP
The first open standard for AI to read your files, query your databases, and use your tools. Until now, every AI needed bespoke integrations.
- January 2025
OpenAI launches Operator
The first mainstream consumer AI agent. It books, buys, fills forms, and completes tasks on the web on a user's behalf.
- April 2025
Google launches A2A
AI agents from different companies can now talk to each other directly. One AI can hand work to another AI without a human in between.
- September 2025
Google launches AP2
AI agents can now make payments on a human's behalf, with cryptographic proof of permission. Backed by 60+ partners including Mastercard, Visa, and PayPal.
- Late 2025
Neos Wave field test, AI calls a contact centre
We send a consumer AI agent to call a major UK contact centre on a customer's behalf. It passes verification. Completes the interaction. The team on the other end doesn't realise. The first documented evidence of Mode 3 in the wild.
- Early 2026
OpenClaw: customers start building their own agents
A weekend open-source project goes global. People run a personal AI agent on their own device and connect it to the messaging apps they already use. The first major step on the adoption curve: customers self-building AI agents. Powerful, but they set it up, run it, and set its limits.
- March 2026
Sequoia and Y Combinator back services-as-software
Sequoia's “Services: The New Software” essay lands and goes viral (3M views on X). Y Combinator's summer 2025 batch is roughly half AI-agent companies. The stated thesis: for every $1 spent on software, $6 is spent on services, and the next generation of category leaders will be AI services built for consumers, not tools sold to enterprises. The capital is now pointed at Mode 3 and Mode 4.
- March 2026
The Service Handshake published
The first open standard for managing AI-to-AI service interactions. Six declaration elements. Independent of any vendor. Free to use under CC BY 4.0.
- September 2026
Meta launches Muse, a personal AI agent for everyone
Meta puts an AI agent in the Muse app and WhatsApp that books travel, fills out forms, sends email, negotiates bills, and checks out on the customer's behalf. It keeps working after the app is closed. The next step on the curve: no setup, no self-build. Customer AI agents go from built by early adopters to done for everyone.
730K → 2.5Mdownloads in 5 days, then 13 days. Ahead of Claude (400K) and Grok (200K) over the same windowSensor Tower via CNBC - Today
Google Universal Cart goes live in Search and Gemini
The commerce infrastructure is operational. Open to anyone. Scaling now. Your customers are already using it.
- What's next
Personal AI assistants going default
Apple Intelligence, Alexa+, Gemini Assistant, agentic capabilities rolling out as the default experience on devices your customers already own. From “available” to “everywhere”.
Why this is speeding up.
Each protocol removes a barrier that used to stop AI from acting on your customer's behalf. Read information. Talk to another AI. Pay for things. Complete a purchase. Act when nobody's watching.
Two years ago, none of the infrastructure existed. Most of it shipped in the last twelve months. Adoption is compounding: the protocols count downloads in the tens of millions, and Meta's Muse passed 2.5 million downloads in under two weeks. It's open source, supported by Google, Anthropic, OpenAI, Mastercard, Visa, PayPal, American Express , and built into devices your customers already own. Claude. ChatGPT. Gemini. Alexa+. Muse.
The infrastructure is built. The adoption curve is steepening.
The framework
Four modes. One that changes everything.
Every service interaction now sits in one of four interaction modes. Most CX investment is in Mode 2. Most CX risk is in Modes 3 and 4, the ones where the customer brings AI to the conversation.
Mode 1 · Human ↔ Human
The model we know.
Declining as a share of all interactions.
Mode 2 · Human ↔ Brand AI
The model most companies are building.
Chatbots, IVR, self-service. Customer still human.
Mode 3 · Customer AI ↔ Human
The model almost nobody is designing for.
Customer's AI calls your contact centre. Already happening.
Mode 4 · Customer AI ↔ Brand AI
The model that changes everything.
Two AI agents negotiate outcomes on behalf of their principals. No human in the loop. The fastest, most efficient, and most dangerous mode, because without shared structure, every interaction is a negotiation from scratch. Policy gaps exposed instantly.
- Status
- Already live in early deployments
- Cost
- High tokens, slow resolution without structure
- Solution
- The Service Handshake (open standard)
Learn more about the 4 modes of service.
Get in touchThe Service Handshake
An open standard for service conversations when AI joins in.
Mode 4, AI talking to AI, only works at scale with shared structure. The Service Handshake is an open standard that defines what each party must declare before any AI-involved service interaction: their goals, what they can do, what data they can share, what happens when things go wrong. Published under CC BY 4.0. Free to use. Free to build on.
SH-1.2 · DOI: 10.5281/zenodo.23181969 · Licensed CC BY 4.0
Goals
What this party is trying to achieve, in priority order. Including explicit non-goals.
Options & Constraints
What this party can and cannot do. Hard limits, soft limits, jurisdictional boundaries.
Data Permissions
What data can be shared, for what purposes, with what retention rules.
Fallback Rules
What happens when confidence is low, goals conflict, or harm risk is detected.
Cost Parameters
What the interaction is allowed to cost in time, tokens, or money.
Authority Level
Who the agent acts for, who benefits, what it can commit to.
Workshops & consulting
Get the conversation started in your business.
Small groups. Senior rooms. Working sessions on the four modes, the standard, and the operating decisions your organisation is steadily deferring. Register your interest and one of the team will be in touch.
- Practical working sessions with your leadership team
- Audit your Dual CX readiness in depth
- See how a customer AI interacts with your service, live
- Map your service against the Dual CX framework
Field notes
What we've been finding out there.
Our own field work, real calls, real orders, real brands, alongside the latest writing from the wider network. Practitioners who see Dual CX from the floor, not the deck.
Happy for you to use our work — just credit “Dual CX. CC BY 4.0” and link back. How to credit us.
From the field
Field test · ID 003
The AI called the contact centre. Nobody knew.
We built a customer-side AI voice agent and pointed it at a real UK contact centre. Three calls, three human agents, none detected it. What surprised us wasn't verification, it was that the AI set the temperature of the interaction.
Read the field test
Mode 4 · Customer AI ↔ Brand AI
One message. No phone calls. 4.2 seconds to resolution.
What happens when a customer's AI contacts a brand's AI directly, and both parties speak the Service Handshake. Aligned handshakes resolve in seconds. Misaligned ones escalate cleanly, with full context inherited.
See both demos
The State of Dual CX · Q1 2026
We audited 100 UK brands. None were ready.
A field audit across 17 sectors. Zero brands had a published AI-interaction policy. 73% had no fallback path when a customer's AI reached them. The gap isn't a technology gap, it's a design and governance gap.
Read the audit
Latest from the network
Browse the library →The transparency deadline has passed. What Article 50 actually changes for a contact centre.
Article 50 of the EU AI Act came into force on 2 August 2026, with the Commission's final guidelines adopted on 20 July. A working note for contact centre leaders: what changes on the queue, what a leader needs on the record now, and how the transparency logic reads across to Modes 3 and 4.
Maria McCann
Neos Wave · August 2026
The CX blind spot that's going to cost you: introducing Dual CX
The framing piece. Why the customer is now the second designer of your service, and what CX, service, and product leaders need to see before they finish their next AI roadmap.
Maria McCann
Neos Wave · 2026
The moment trust breaks: inside Jess's escalation
The customer-side story of an agentic commerce failure. Operating models have to account for what customers actually experience when AI transactions go wrong.
Martin Hill-Wilson
Brainfood Consulting · May 2026
The Dual CX library.
A growing collection on agentic commerce, contact centre readiness, the business case, and Dual CX framing. Filter by topic.
Add to the conversation.
Practitioner, researcher, or operator with something to add? We'd like to hear from you.
Before you go further
Not sure where you are on the curve?
Take the 2-minute Dual CX readiness diagnostic. Eight questions, no email required to see your score. Get a scored view of where your service breaks when AI enters the conversation.
Take the diagnosticNo sign-up. Answers stay in your browser until you choose to share.
Get in touch
Talk to Dex, or drop us a line.
Dex can answer most questions about Dual CX, the four modes, the audit, and the Service Handshake, by voice or chat. For anything else, the form is the fastest way to reach the team.
Option 1 · Talk to Dex
Ask anything, by voice or chat.
Dex is a voice agent built on the Neos Core. Get the headline figures, field stories, and routes, fast. See what good Mode 2 looks like in practice: declared, honest, useful.
- Dual CX, the four modes, the audit, the Service Handshake
- Pointed to the next session or workshop, if relevant
- Voice or text, your choice
Dex is an AI assistant. Conversations may be reviewed to improve the service.
Option 2 · Get in touch
Drop us a line. One of the team will be in touch.
The series
Join us for live sessions on LinkedIn.
Maria McCann and Martin Hill-Wilson are restarting the LinkedIn Live series on Dual CX — the shift, the modes, and what it means for how service is designed and run. Watch the short precursor, and leave your details if you’d like to hear when the next session is scheduled.
Episode 2 is now live. See all episodes →
Precursor · 7 minutes
All episodes →Follow the series
Get a heads-up before the next live session.
Name and email, that’s it. We’ll write when a session is scheduled — not before, not otherwise.