Agentic CX
We’ve all had the moment where you’ve typed out the perfect, specific question into a support chat, and what comes back is next to useless or sends you in a circle of FAQs.
For years a support bot existed to answer. It pointed you at what is meant to be the right article, or the right form, and hoped that was enough. The new generation does something different. It acts. It changes the address, processes the refund, moves the subscription, verifies the account, and reaches for a human only when it should. There is a name for this now, agentic CX. Software agents that hold a real conversation and complete the task behind it, across chat, voice, email and real humans.
A wave of well-funded vendors is promising high resolution rates, brand-perfect tone and easy deployment. This article takes a closer look at where agentic CX has landed, what is still just a promise, and where it's hopefully going.
Why it’s different this time (and who’s behind it)
We’ve been promised the helpful robot before, so it’s natural to be sceptical. The constantly evolving models mean we can now reason through messy, half-formed requests, where older bots could only pattern-match keywords (that’s the simple version). They can be plugged into the systems behind the brand to do the ‘thing’, whether that’s processing a refund, moving a subscription or verifying an account.
Most of these agents run on the same widely available AI from OpenAI, Anthropic and Google, so the brain is not what sets them apart. The value sits in everything wrapped around it. e.g. the connection into your systems so a refund just happens, the checks that catch a wrong answer before a customer sees it, or the unglamorous job of plugging into old contact-centre kit.
A fresh crop of companies, built from the ground up for this kind of work, are now setting the pace. Sierra, for one, has gone after big consumer brands with a managed, omnichannel agent, and prices by the resolution, charging only when an issue is genuinely sorted. PolyAI has put its energy into voice, and making a phone call feel natural at scale. Germany’s Parloa leans into control and governance, letting an enterprise simulate and supervise an agent before it goes anywhere near a customer. Decagon pitches itself as a platform any team can wield, letting non-technical staff shape agent logic in plain language while engineers keep control of the guardrails and integrations, all unified across chat, voice and email.
The incumbents, the Salesforces and Intercoms, are bolting agents onto their existing suites, strong on data and install base. It's easy to assume they are a step behind, but Forrester's most recent ranking of these platforms puts them in the Leader tier, with the buzziest challengers a rung below as Strong Performers. The incumbents own the system of record, and they are closing the gap faster than the hype suggests.
4 key areas for CX leaders to consider
The death of deflection and paying for outcomes. Deflection only ever made sense because the old bots couldn’t really resolve customer queries. A decision tree cannot process your refund or change your plan, so the best it could manage was to nudge you towards an answer and hope you gave up before reaching a person. That stops being the goal the moment an agent can genuinely complete the task. Once it can reach into the billing system, verify who you are and make the change, there is nothing left to deflect.
The alternative to deflection has opened the door to a different commercial model, and it is spreading fast. Several of the new players charge by the resolution, so the vendor only earns when the ask is genuinely sorted. Zendesk moved its AI agents to per-resolution pricing in 2026, Intercom's Fin charges around a dollar a resolution, Decagon prices on resolved interactions, and at least one analyst now calls outcome-based pricing a market standard.
In principle that is a healthier bargain than paying per seat or per conversation. The devil, as ever, is in the detail. Who decides what counts as “resolved”, and what about the agent that half-helps and then passes you on? When the same vendor both delivers the outcome and scores it, that definition is doing a lot of quiet work, and it is worth pinning down before you sign.
The rise of the brand agent. If an agent handles 70% of your interactions, that agent is embodying and becoming your brand for most people, most of the time. Its tone, its manners and its judgement now live in a configuration screen. One benefit of these new ways of working is that increasingly they’re able to remember everything. With SiriusXM’s “Harmony”, the relationship starts to live inside the platform itself. She has been handling SiriusXM listeners since early 2024, which makes her one of the longest-running agents of her kind. Alongside the everyday jobs of subscriptions, payments and billing, she can reset the satellite signal to your car radio from space in the middle of a chat, and she knows that a listener asking for “Howard” means Howard Stern. Late in 2025 SiriusXM became the first business to give its agent a lasting memory, so Harmony can remember a listener from one conversation to the next. SiriusXM calls it their highest-rated and lowest-effort service channel, though the hard numbers behind that are not public.
If the agent is your brand, though, a configuration error becomes a brand crisis, live on your own storefront. Gap found this out in December 2025. Days after launching a Sierra-powered chatbot on Gap.com for product questions, it was steered off-scope into inappropriate topics. It turned out to be a coordinated effort to jailbreak the agent, combined with a guardrail that had been misconfigured on the vendor's side. The controls were fixed and the agent stayed live.
Voice. Everything so far has been roughly chat-shaped, yet the shift to spoken agents changes the design problem entirely, from tone and timing to how a brand should actually sound out loud. The harder and more interesting part is listening. A voice carries more than its words, and companies like Hume have built their approach around reading the feeling in how something is said, measuring the spectrum of emotions. For a brand that means an agent that can hear confusion or frustration building and choose to soften, slow down or pass the call to a human before a customer boils over. This feels particularly exciting and an area rich for brands to come to the fore and really support customers. It is also the least mature of the channels at the moment, though changing rapidly. Hopefully gone are the days of yelling in frustration as the calm voice asks, “did you mean, book tickets for the cinema in Aberdeen?”
Rewriting the approach to journeys. For years we’ve drawn the journey by hand. We’ve mapped the funnels, designed the screens and built the menus and forms to shepherd people down the path we expected them to take. The new agents learn the journey from what customers say and do. Sierra Ghostwriter is a good example of this, reading back over thousands of real conversations to surface the recurring snags and the questions nobody ever designed a screen for, then feeding all of that straight back into how the agent behaves, “It allows companies to design, test, and optimize production-ready customer experience AI agents purely through conversational natural language”.
We’ve only scratched the surface
A few other threads that already deserve some attention…
Designing for the worst moment. A support experience is only really judged by how it treats someone in distress, or with an accessibility need, and how cleanly it gets them to a human who can help. This is where ‘AI empathy’ and understanding will be tested - imagine trying to manage a loved ones pension or banking details after they die and ‘computer says no’ - Companies like Hume are developing ‘articicial intelligence that understands and response to human emtions’, one to look at. What’s left for people is the hard, knotty, often emotional work, helping humans when and where they’re needed most. Perhaps us humans become the premium tier, or luxury option? Refined roles, higher-skilled and more valued.
What sits underneath the understanding. An agent’s read of a customer is only ever as good as the semantic layer it draws on, which is where ontologies earn their keep. An ontology is a map of how a business actually fits together (simply put), its products, customers, accounts and the relationships between them, written down in a way a machine can reason over. It’s what lets an agent understand that a missing payment connects to a specific account and policy. The richer and more accurate that map, the more genuinely an agent can reason about a customer. We dug into this in our piece on ontologies and personalisation.
Build or integrate. Behind every smooth agent there sits a pile of decisions about whether you buy a platform or build your own. Monzo has written openly about engineering its own customer-operations agent, and it is a useful read for anyone weighing this up. Most of the effort went into everything around the model, grounding answers in knowledge their experts had vetted, building guardrails to catch a hallucination before it reached a customer, and standing up a heavy evaluation setup to prove the thing was safe in a regulated business. They had a human check every message before the agent spoke to customers directly, then sampled less as confidence grew. They began by answering questions, then moved to doing the work, with flows that find a missing refund or order a replacement card after a loss or fraud. The agent itself is the small part, with the work sitting in the knowledge, the tools, the evaluation and the human oversight.
Reactive turns proactive. Give an agent memory and a live view of your systems and it stops waiting to be asked. Imagine if your flight slips, and it has already rebooked you and messaged you to confirm the new booking details. SiriusXM giving Harmony a lasting memory is the early signal. The stuff of sci-fi and creepily helpful.
Where next and the risks?
The risk of change has always been there, whether it’s cost of implementation (not simply plug-and-play), disruption, re-platforming, and the chance of getting the ‘tool’ selection wrong. The other consideration, as always with a shift like this, is people. The same agents that resolve the easy queries also remove the work a lot of customer-service roles are built on. Forrester expects something like half of customer-service jobs to go to AI by 2030. Enabling and redeploying people onto the knotty, emotional work the agents cannot touch, is the way forward.
As has always been the case, the baseline for "a good customer interaction" is still set by the last best experience they've had anywhere, your competitors, other industries or companies. The risk here is standing still while your competitors reset what customers expect.
I’m optimistic about what the future holds and the continuing evolution of the helpful customer agent bots. Long live the bot.
