Customer experience
How agentic AI will change customer experience design forever
TL;DR: Customers are increasingly taking their questions about your products to third-party AI assistants instead of your website or app. Building your own conversational experience is essential to remaining part of that decision-making process. The 3/3/3 methodology helps you validate, prototype and launch one in just over three months.
Your customers are still visiting your site, opening your app, calling your service line. The traffic numbers look fine. What’s changed is what happens before they arrive.
“What’s the best account for my situation?” “Which of these three products fits what I need?” “How do I resolve this?” Questions like these used to come to your website, your call center, your branch. Increasingly, they go to ChatGPT, Claude and Gemini instead, where customers get a complete, personalized answer in seconds in a conversation you're not part of, with an outcome you'll still be held to.
With ChatGPT alone handling over 2.5 billion messages a day from 900 million weekly active users, customers have already imported the expectation of conversational immediacy into every interaction they have.
The instinct, reasonably, is to close the gap: offer that same conversational experience inside your own channel, powered by your own data and governed by your own controls, and reclaim the customer interaction. Every vendor pitch is built on some version of that promise. Almost none of the resulting deployments survive contact with real customers.
What works is proving the concept before committing to it. The 3/3/3 methodology is our approach to taking a product from concept to a production-ready minimum lovable product (MLP) in just over three months, progressively validating technical feasibility, business value and customer fit.
To understand why this approach works, it helps to start with why most conversational pilots fail.
Why most conversational pilots don’t survive contact with customers
Three hard problems sit between a conversational demo and a production-grade customer channel. These aren’t interface problems. They determine whether conversation can carry the weight of a real business interaction.
01. Speed that keeps customers engaged
Human conversation has a pace. Cross a threshold of delay and the other party assumes something has gone wrong. That threshold is well under a second. The benchmark in your customer’s head is now ChatGPT, Claude and Gemini, all of which already respond in under a second.
What it costs you. If your channel can’t hold that pace consistently, customers don’t complain. They simply stop using it and go back to the channel where they got the answer faster, which is no longer yours.
Why it’s genuinely hard. The obvious approach, routing every customer message through a large AI model, is both too slow and too expensive to operate at scale. Run every reply through the heaviest AI in the stack and your unit economics collapse before you’ve even launched. This isn’t a fringe observation. In June 2025, Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, and was direct about the underlying reason:
Many use cases positioned as agentic today don’t require agentic implementations.
The cost and complexity of getting speed right at scale, without burning the budget, is one of the reasons most pilots never become production channels.
02. The trust boundary around a channel that can actually act
A conversational channel that can only talk is a novelty. A conversational channel that can act on your customer’s behalf, moving money, changing an address, raising a claim, booking a service or committing a transaction, is a real channel.
What it costs you. Most organizations today respond to that risk by saying no. Pilots stay in read-only sandboxes, unable to actually do anything useful for the customer. But a read-only conversational channel isn’t a customer channel. It’s a brochure with a microphone. You get all the cost of building it, none of the business value.
Why it’s genuinely hard. The moment a conversational AI is allowed to take action, it enters the same risk envelope as every other serious channel in your business, with none of the security, audit and compliance controls that took decades to build around call centers, apps and branches. Giving a generative AI the keys to your systems is not something your CISO, your regulator or your general counsel will sign off on by default, and they’re right not to. It’s an engineering, risk and governance problem that has to be solved together, which few vendors know how to do.
03. Answers your business can stand behind
This is the one that keeps CFOs and general counsels awake. Generative AI, left to generate, will tell your customers things that are wrong with the same confidence it tells them things that are right. For a regulated institution answering a customer’s question about their money, their coverage, their eligibility or their rights, it’s a career-ending risk for whoever signed it off.
What it costs you. One wrong answer, spoken with confidence, is all it takes. A misquoted premium, an incorrect balance, a hallucinated policy or a wrong eligibility decision. The customer acts on what they were told. The liability lands on you, not on the vendor whose AI invented the number.
Why it’s genuinely hard. You can’t build a customer channel on a system that might hallucinate. It has to be built so that hallucination is structurally impossible for the answers that matter: the numbers, the policies, the entitlements and the decisions. That’s a design problem the agent-first vendors on the market today don’t have a clean answer to. Solving it requires rethinking the relationship between the AI and your systems of record from the ground up, which is slow, expensive and unforgiving of mistakes. It’s also what separates a conversational pilot from a channel your CFO can put her name to.
Speed, trust and certifiability. Each one is hard on its own. Solving all three together, under real customer load, is what moves a conversational channel out of the demo phase and into production.
How to build a conversational channel that works
The typical response to a shift this size is to commission a strategy, build a roadmap and plan a multiyear program. It is the wrong response here. The shift is moving too fast, and the only learning that matters comes from putting something real in front of customers. By the time a traditional program lands, its assumptions will be out of date and competitors who moved faster will have taken position.
At Thoughtworks, we advocate for an agile, iterative approach instead. Rather than making large upfront commitments, organizations should move in fast, evidence-driven increments, continuously validating that the technical approach works, that the business case holds and that customers actually want the experience before investing further.
We structure that approach through our 3/3/3 delivery motion:
Three days | Concept: Is this real for us? Align executive sponsors on where your customers are moving, which conversations matter most to your business and what the shape of a first move looks like.
Three weeks | Prototype: Can we actually do this? Build a working prototype of a real customer conversation. Prove the speed. Prove safe integration. Prove the answers stand up. Test with users. Emerge with a business case grounded in evidence.
Three months | Build and launch: What happens in the market? Put an MLP into the hands of customers. Measure actual impact on the relationship, the economics and the competitive position. Iterate from evidence.
It progressively de-risks the investment across the three dimensions that decide whether a product actually succeeds — desirability, feasibility and business viability — in a cadence every executive can commit to.
This is possible due to AI/works™, Thoughtworks’ agentic development platform that sits underneath this cadence. The platform turns business needs into dynamic specifications and working code through coordinated AI agents. Because it relies on established components and built-in guardrails, teams can accelerate development while maintaining security and compliance.
It's about not moving fast for its own sake, but finding out early which ideas were never going to work, instead of spending years finding out the expensive way.
The question for your board
The question at the top of your organization shouldn't be which AI tools to procure or which models to standardize on. Those are execution questions, and the answers will change every six months for years. The question that matters now is a strategy question, and it has a short shelf life:
Are you present, as a business, in the way your customers now want to be served? Or are you about to watch your relationship with them thin out in favor of whoever is?
For most organizations, the honest answer today is no, and the plan to change it doesn’t exist yet.
That’s a competitive position, and it’s a deteriorating one. Pick the single customer conversation that matters most to your business. Put something real in front of customers before the end of the next quarter. Then scale from a position of evidence.
Your customers have already moved. The only question left is whether you move with them.
Sources and further reading
