Ten releases from the first half of 2026, all pointed at the same thing: what a customer-facing agent does once it is live in production.
If you run a customer-facing AI agent, the hard part starts at go-live. Getting an agent to answer a common question is close to solved. Getting it to run the refund, pull the customer's order history, hand off to a person at the right moment, and leave a record you can defend to an auditor is a different problem. That is the distance between a demo and a production operation, and it is the distance Zowie is built to close.
Over the past year the platform ran more than 100 million conversations for enterprises including Allianz, KRUK, and Decathlon, much of it regulated work. Zowie is seven years into running customer-facing agents in production. The first half of 2026 went into ten releases, and none of them lengthened a feature list for its own sake. Each one makes the agent do more once it is live: more context before a conversation, more action inside it, and more control after it goes into production.
One decision holds all of it together. The model handles the conversation; the Decision Engine handles the business logic. Refunds, claims, eligibility, and routing run as rules against real data, and the model never decides them. "Mostly right" can pass for a casual reply. It cannot pass for a refund.
Before the conversation: give the agent context to work from
1. Moments: shape the interactions most likely to break
January 28, 2026
Every agent hits the same handful of moments over and over: the opening hello, the question it cannot answer, the customer who asks for a human, the end of a process. They look trivial in a spec, and in production they are exactly where the experience breaks. A weak fallback puts the customer in a loop; a clumsy handoff burns goodwill; a finished process leaves the customer with nowhere to go. Moments turn each of those points into a predefined, configurable scenario, so you decide how the agent behaves in the situations most likely to go wrong before they reach a transcript or an escalation queue. Less guesswork at runtime, more consistency across millions of interactions.
2. Interaction Setup: the agent knows who it is talking to before the first message
March 27, 2026
A good agent should not start cold. Before the first message it can already know the customer's market, their plan, whether they have an open order, and which escalation path they need. Interaction Setup pulls that data from your systems into the conversation before it begins and segments the customer by plan, region, value, status, or language. What matters is where the logic lives: segmentation runs through the Decision Engine, so the agent applies a rule against real data instead of inferring "this looks like a VIP" from the customer's wording. In practice that is a retailer treating a first-time buyer differently from someone with three open orders, and a bank applying the same routing across regulated workflows every time. It is the line between an agent that reacts to text and one that orchestrates around context.
3. Native Shopify, Gladly, and Dixa integrations: less plumbing between pilot and production
March 30, 2026
An agent in isolation is worth little. It needs the order system, the tickets, the knowledge base, the customer record, and a path to escalate, and every missing integration is another week between pilot and production. In March we deepened the Shopify, Dixa, and Gladly integrations, so agents read existing knowledge, work with commerce and ticketing data, and open or escalate tickets without a custom build in the way. Enterprise AI projects rarely die because the model was too weak. They die because the plumbing never reached production quality. This shortens that path.
During the conversation: let the agent do more than answer
4. Custom Visual Aids: show the answer instead of narrating it
March 10, 2026
Some information does not belong in a paragraph. An order summary, a claim status, a payment receipt, a set of product options: all of it lands better on screen than read aloud. Custom Visual Aids let the agent render your own UI components inside the conversation, on live data, in your design system. That changes what the conversation is for. The agent stops being only an answer machine and becomes a place to act. For commerce the effect is measurable: Total Wine used Zowie for product discovery and reported 4x higher conversion and 20% higher average order value. In regulated work it is about trust: show the receipt or the policy status instead of asking the customer to take the agent's word for it. The platform places these components into the conversation on your live data; the model does not invent them. And they matter most on voice, where an agent cannot read out an itemized receipt or ten product options without losing the customer. The visual carries that detail to the screen while the call stays spoken and simple.
5. Voice configuration in the UI: tune how the agent sounds without an engineering ticket
May 20, 2026
Voice breaks differently than chat. A poor chat reply can be reread or quietly escalated; in voice, the timing, the pauses, and the way the agent handles a noisy line all reach the customer in real time, and the same agent can sound sharp or broken depending on those settings. Tuning them used to mean an engineering ticket. Since May, voice configuration lives in the Zowie UI, so the CX and ops people who understand the customer can adjust speed, vocabulary, and noise handling without waiting in a queue. Engineering still governs the platform; the day-to-day adjustments move to the people who live in the experience. The larger win is that voice does not splinter into its own silo, with its own controls and its own operational debt. One platform, every channel.
After it goes live: see what happened and control what changes
6. Supervisor: one view of how the agent is actually performing
February 3, 2026
Automation you cannot see into is a liability. Once an agent handles thousands or millions of conversations, "we'll read some transcripts" and a spreadsheet stop being a plan. Supervisor went generally available in February as one view of interactions, success rate, transfer rate, and breakdowns by intent, knowledge base, process, and contact reason. From any number on the screen you click straight into the conversations behind it. That is the part that counts: aggregate metrics tell you something is off; they do not tell you what. Supervisor gives you the path from signal to evidence, and it lets leadership tell whether the agent is improving for real reasons or just containing more tickets. That matters most in regulated work, where you have to prove how each interaction was handled rather than report an average.
7. CSAT in Supervisor: resolution and satisfaction, side by side
March 9, 2026
Every automation program can walk into the same trap. Resolution climbs, costs drop, transfers fall, every number on the dashboard turns green, and customers are quietly getting more annoyed, because resolved and happy were never the same thing. CSAT in Supervisor puts the two next to each other. You watch satisfaction day over day, compare it across periods, and break it down by intent, knowledge base, process, and contact reason. When CSAT dips you are not staring at one number guessing what happened; you can see that the billing-dispute flow or one stale help article is dragging it down, and fix that one thing. "Is the agent any good?" is an argument. "The refund intent dropped four points this week" is a task.
8. Channel-level performance: find the resolutions a blended average hides
March 31, 2026
Averages are where operational problems hide. An agent can be strong in chat and weak in email, and a blended number makes everything look stable while something is plainly broken. Channel-level breakdowns pull that apart: how the agent does on each channel, and where it is leaking resolutions. For a multi-channel operation this is not optional, because voice, chat, email, and social carry different expectations, latencies, and escalation patterns. What matters is where it is failing, and what to fix first.
9. Audit Log: who changed the system, when, and what changed
May 12, 2026
Governance is not only about how the model behaves. It is also about who can change the system, when they did it, and what changed. The moment an agent touches customer-facing processes, that history stops being a nice-to-have. Audit Log records changes to roles, permissions, and teams, each entry carrying a timestamp, an actor, and the before-and-after state, filterable across both your users and Zowie staff. For regulated industries this is a requirement, and SOC 2, GDPR, DORA, and the EU AI Act all point the same way. It is worth separating two things people conflate. Audit Log covers the administrative layer: who changed the system. Traces covers the execution layer: what the agent did when it ran. Production AI needs both, and they answer different questions.
10. Cowork: change a live agent without breaking it
April to June 2026
The more an agent matters, the more people work on it: CX adjusting flows, ops documenting edge cases, product tuning the experience, engineering reviewing routing. Without guardrails that activity gets risky, from two people editing the same flow to a staging push that quietly ships a persona change no one meant to make. From April to June we shipped a run of updates for exactly this. Flow notes let teams document decisions on the canvas, sticky notes in April, then color-coded and resizable ones inside the Decision Engine in May. Live collaboration shows who else has a flow open, warns you when it has changed under you, and puts a three-way diff in front of you before you save. Staging diffs lay out every changed flow, routing rule, and persona setting before anything publishes. None of this is cosmetic. It is change management for AI agents. For KRUK, running under regulatory constraints across its markets, that control is part of the product. Getting the first agent live is the easy milestone. Changing a live system safely, every week, is the real one.
What ten releases add up to
Read together, the ten releases are one move: Zowie becoming a stronger control layer for the AI agents enterprises run in production. The year started from a single premise, and it still holds. Answering a common question from a knowledge base is close to solved; the ceiling is in the process work, the refunds, claims, account changes, and billing disputes that depend on real business logic. Zowie's answer there is architectural. The model handles language, the Decision Engine executes the logic, Traces records the path, Supervisor surfaces quality, Audit Log tracks every change, and Agent Connect keeps the stack open enough to fit what you already run. Six months made each of those deeper: agents that do more inside the conversation, show more of what they did after it, and can be changed safely while they run regulated work at scale.
Book a technical walkthrough, or read the full product updates.
