Customer Experience News is a weekly newsletter and video about the most important news and discussions for Customer Experience and Customer Support Leaders.
This is all the weekly news you need in around 5 minutes.
First, an apology. The newsletter went quiet last week. Spring Break turned into a mountain biking injury, which turned into a week of catching up.
Things are moving again, including my fractured elbow.
Our main story today is Zendesk Plans to Acquire Forethought.
Zendesk plans to acquire Forethought:
What Changes for Support Teams
On March 11, Zendesk announced a definitive agreement to acquire Forethought, an agentic customer service startup whose AI agents handled more than a billion monthly customer interactions in 2025.
Zendesk describes it as the biggest deal the company has made in two decades.
Forethought’s core capability is what Zendesk is calling the Resolution Learning Loop: a self-improving system that uses resolved conversations as training inputs. Teams don’t tune prompts manually. Volume and complexity are what improve the system, and every resolved interaction raises the baseline. Zendesk says their AI customers are routinely achieving 80%+ automation rates today, and that they expect AI agents to handle more service interactions than humans by year-end.
Voice is the bigger story
The part of this acquisition that hasn’t gotten enough attention is voice. Zendesk describes voice as “historically one of the hardest channels to scale and learn from” and frames the Forethought acquisition as the fix. Unstructured speech becomes training data. Sentiment markers from tone of voice, previously impossible to capture at scale, become inputs to the learning loop. A customer calling in mid-refund can have that refund processed during the call. When the AI can’t resolve the issue, it hands off to a human agent with full conversation context already in place. For support teams that have treated voice as a channel apart from everything else, that represents a meaningful change in what’s possible.
What changes for Zendesk customers
Forethought’s self-learning mechanism gets embedded into Zendesk’s Resolution Platform as an additive layer rather than a rebuild. The practical effect should be better autonomous resolution rates over time without teams having to continuously update workflows as new scenarios emerge. Whether that plays out as described depends on how cleanly Forethought’s architecture integrates with Zendesk’s existing AI agent product.
What changes for non-Zendesk customers
Zendesk’s stated position is the opposite of what acquisitions in this space typically produce. They have explicitly committed to extending Forethought’s capabilities to teams on other platforms, with the language that:
Progress toward agentic service should not be limited by the systems teams use today.
That is an unusual public commitment for an acquirer to make. The skepticism is not about the intent but about the execution: roadmap decisions after close rarely align with acquisition-day statements, and Forethought customers on Salesforce or Freshdesk should watch product updates closely over the next two quarters.
Watch for Zendesk’s integration timeline announcement and whether Forethought’s multi-platform availability survives past the deal close.
UPDATES AFTER POSTING:
In related news, Decagon launches Duet, your agent building partner that enables teams to build self-improving agents. Read more →
MyAskAI posted on LinkedIn, “Zendesk just admitted (for the third time in two years) that they can’t build AI support in-house.” They shared that they’ve already shipped automated self-learning amongst other features. Read more →
Kustomer AI Now Runs Inside Zendesk
Kustomer, a customer service platform, released Kustomer AI for Zendesk on March 10, 2026. The integration layers Kustomer’s AI automation on top of a live Zendesk instance without requiring a platform migration. It uses a team’s existing Zendesk knowledge base to resolve customer inquiries autonomously, deflects incoming tickets, and hands off to live agents inside Zendesk when escalation is needed.
Operational Impact:
For teams evaluating AI upgrades but unwilling to leave Zendesk, this removes a significant barrier. The upgrade becomes a configuration task rather than a migration project. Teams that have invested in Zendesk Help Center content can put that content to work through Kustomer’s AI without rebuilding it elsewhere. Also released this month: AI for Customers 2.0, which introduces a Procedures framework that gives AI agents guided decision paths rather than relying on generative reasoning alone. Both releases together suggest Kustomer is betting on structured AI behavior over open-ended responses.
Implementation Considerations:
Setup requires connecting your Zendesk instance through Kustomer AI settings, which means administrator access and API credential management. The bigger dependency is knowledge base quality. If your Help Center content is incomplete, outdated, or written for customers rather than AI interpretation, the autonomous resolution rate will reflect that. Procedures in AI for Customers 2.0 can constrain AI behavior, but they require upfront configuration effort to define the decision paths correctly.
Intercom Adds Sensitive Data Redaction at Message Ingestion
Intercom, an AI-powered helpdesk platform, released automated sensitive data redaction. Incoming messages are scanned and redacted before they reach Intercom’s database or appear on any teammate’s screen. Built-in rules cover credit card numbers (last four digits preserved) and Social Security Numbers (fully masked). Teams can add up to 10 custom regex rules to mask business-specific data such as account IDs or policy numbers. The feature works across Web Messenger, mobile SDKs, inbound email, and call transcripts.
Operational Impact:
For teams handling financial services, healthcare, or insurance conversations, this addresses a compliance gap that previously required custom middleware or periodic data purge processes. Customers regularly share payment details or government IDs in support channels not designed to retain that data. Redaction at the point of ingestion is cleaner than retroactive deletion and removes human exposure entirely.
Implementation Considerations:
The built-in rules are toggled on from security settings. Custom regex rules require administrators to build and test their own patterns, which introduces the possibility of misconfigured rules that miss targeted data types or over-mask content agents need to see. Teams should test custom rules against real message samples before treating the feature as a compliance control. The feature also does not account for data shared outside of Intercom’s supported channels.
Plain Overhauls Thread Management with Lock, Merge, and Structured Fields
Plain, a developer-focused support platform, released three thread management updates. Lock threads permanently closes a resolved conversation; if a customer replies to a locked thread, Plain creates a new thread automatically and links the two. Merge threads consolidates conversations about the same issue that arrived across multiple channels into a single parent thread, with replies routed back through each message’s original channel. New thread field types, released March 12, add support for numbers, currencies, and dates directly on threads, with filtering, sorting, and workflow trigger conditions tied to those field values.
Operational Impact:
Lock and merge address a persistent problem in thread-based tools: resolved conversations that reopen when customers reply, and the same issue generating parallel threads across different channels. For teams using Plain as an operational layer, the new field types are the most significant update. Structured data from CRM systems, such as deal value or renewal date, can now drive queue prioritization and automation rules without requiring a separate workflow tool.
Implementation Considerations:
Locking is permanent. The original thread cannot be reopened once locked, which means teams need a clear policy on when locking is appropriate versus leaving a thread in Done status. Merging requires manually selecting parent and child threads. Teams handling high volume across multiple channels should define merge criteria before enabling the feature broadly, otherwise the timeline view on parent threads becomes difficult to follow.
Pluno Launches Custom API Integration for AI Agents
Pluno, a platform that connects customer support systems with issue trackers like Jira and Zendesk, released Custom API Integration. The feature allows teams to connect external REST APIs as tools that Pluno’s Copilot and deflection agents can access mid-conversation. When a customer asks a question that requires live data, the AI retrieves the answer directly from a connected system rather than escalating or redirecting the customer elsewhere.
This differs from MCP (Model Context Protocol), which is an open standard that enables AI agents to discover and connect to tools automatically using a shared protocol. MCP requires the external service to publish a compliant server; Pluno’s Custom API Integration works with any REST API regardless of whether the provider has built MCP support. The tradeoff: Custom API setup is managed and manual where a working MCP server connection can be self-service once the server exists.
Operational Impact:
For support teams that regularly escalate questions that could be answered with a live system lookup, such as order status, account tier, or entitlement checks, this reduces the handoff. The AI handles the retrieval and surfaces the answer without an agent needing to context-switch to another system.
Implementation Considerations:
This is not self-service today. Setup requires sharing API documentation with Pluno and waiting for them to configure the connection on your behalf. Teams should evaluate which lookups are high-frequency enough to justify that effort. The quality of AI responses also depends on what the API returns and how Pluno’s system interprets it. Poorly structured or ambiguous API responses are likely to produce inconsistent answers from the AI agent.
Pylon Tracks Which Customers Were Notified When Features Ship
Pylon, a B2B support platform built for Slack and Microsoft Teams, released Customer Notification Tracking. When a feature request is marked as Done, Pylon scans issue messages, calls, and broadcast records to identify which accounts have already been notified and surfaces that information directly on the issue.
Operational Impact:
For CS and product teams managing feature request follow-through manually, this removes the tracking overhead. The alternative is typically a spreadsheet, a CRM tag, or memory, all of which break at scale. Having notification status visible directly on the issue connects the product feedback loop with follow-through and reduces the risk of customers being missed when a long-requested feature ships.
Implementation Considerations:
The feature works by scanning Pylon’s own records, so its accuracy depends on whether notification activity happened inside Pylon’s channels. Teams that notify customers through external channels, such as direct email or CRM sequences, will not see those notifications reflected in the tracking. The feature is most reliable for teams that have already centralized customer communication inside Pylon.
Gorgias Decouples AI Agent Knowledge from Help Center Visibility
Gorgias, an e-commerce customer support platform, released an update that separates AI Agent article availability from help center customer visibility. Previously, making an article available to the AI Agent required it to also be published for customers, and vice versa. The two settings now operate independently via a toggle in the Knowledge Hub.
Operational Impact:
This opens two workflows that were not previously possible. Teams can now write AI-only articles covering edge-case policies or internal product details without publishing them to the customer-facing help center. Conversely, teams can keep an article published for customers while excluding it from AI Agent context, which is useful when the content is accurate but too nuanced for the AI to apply correctly without producing wrong or misleading answers.
Implementation Considerations:
The toggle lives in the Knowledge Hub, not inside the article’s help center settings. For teams with multiple people managing help center content, this means the AI availability setting exists in a location most content authors will not check by default. Existing articles retain their previous settings automatically, so no content shifts without an explicit change. Teams should audit their current articles to determine which ones benefit from the new split control.
HubSpot Generates Buyer Personas from a Text Description
HubSpot released Make My Persona AI, updating their free buyer persona generator with AI-powered generation. A user describes their ideal customer in plain language and the tool populates all fields automatically: demographics, goals, job responsibilities, and communication preferences. The original step-by-step form remains available as a legacy option. The output is a shareable document in the same format as before.
Operational Impact:
For support operations teams using buyer personas to inform routing logic, tier classification, or self-service content strategy, a faster path to draft personas lowers the barrier to running that analysis without a research project attached to it. The shareable output format makes it easier to circulate personas across teams for alignment.
Implementation Considerations:
AI-generated personas are only as accurate as the description provided. A vague input produces a generic persona. The original Make My Persona tool was already capable of generating plausible-looking but unvalidated profiles; the AI version accelerates the process and the risk simultaneously. A well-formatted persona built on assumptions can be more dangerous than a rough one that reads like a draft. Teams should treat generated output as a starting point and validate against actual customer data, support transcripts, or sales call notes before using it to drive decisions.
HappyRobot Brings Logistics AI to Customer Support
HappyRobot, which raised a $44 million Series B, positions itself as an AI workforce platform for the “real economy,” primarily logistics and supply chain operations. The company’s AI workers handle voice, email, text, and messaging channels across workflows including appointment scheduling, load status checks, payment inquiries, and customer service. Their customer list includes DHL, through a partnership announced in November 2025 targeting hundreds of thousands of emails and millions of voice minutes annually.
What separates HappyRobot from general customer support AI platforms is the industry focus, not the technology. Most AI agent tools in the CX space are built for digital-first businesses with structured help center content and ticket-based workflows. HappyRobot is built for industries where primary customer communication happens by phone, where the counterpart might be a truck driver on a highway rather than someone at a computer, and where the workflows involve live data lookups against systems like transportation management platforms rather than knowledge bases. That is a different build from what Intercom Fin, Zendesk AI, or Kustomer are optimizing for.
Implementation Considerations:
HappyRobot uses a forward-deployed implementation model, embedding their team with customers for weeks-long rollouts. That positions them closer to a managed service than a self-service SaaS platform, which affects cost, ownership, and how long it takes to get to value. The $44M Series B and DHL partnership signal enterprise traction, but teams outside logistics should assess carefully whether the platform’s workflows and integrations are relevant to their operations before running a pilot.
Now this is:
Strategic Support for CX Leaders.
You’ve got ambitious Support targets and new metrics but you’re not sure what to prioritize first.
The list is long, the queue is getting longer, and you don’t have time to step back and think about CX strategically.
What if you could pressure-test your thinking with someone who’s spent 20 years building Customer Support operations?
No pitch, just a conversation.
