MCP-enabled assistants can now build and manage NEXT workflows. Describe the outcome you want, let the assistant discover available step types, validate the proposed configuration, and create the workflow without manually wiring every step on the canvas.

Catch invalid workflow variables before execution​

The workflow editor now flags references to variables that do not exist. Problems appear both on the canvas and in the step configuration, helping you repair renamed, removed, or mistyped references before they cause a failed run.

Include verbatim quotes in cluster results​

Ask for quotes in Cluster mode and NEXT can include a sample of verbatim feedback with each cluster summary and mention count. Quotes remain in their original language and include source annotations where available, adding direct customer evidence to the quantified themes.

See every data source shaping a chat​

Chats now show all configured data sources in one place. Alongside the focus query, you can see selected memories and connected sources such as Google Drive or Confluence, making it clearer what context NEXT is using to produce an answer.

Ask NEXT to create a PowerPoint or export a result to PPT from a regular chat—there is no separate mode to find first. NEXT shows progress while it builds the presentation and returns a real .pptx file ready to download and edit.

Search directly inside transcripts​

Transcript search is back with controls designed specifically for reviewing records. Search stays scoped to the transcript, highlights every match, and lets you move between results with next and previous controls while keeping count of where you are.

Understand where every memory came from​

Memory details now distinguish the insight itself from its supporting evidence and show how the memory was generated. Follow links back to the analyzer, workflow, exact workflow run, record passage, or other memories that produced it.

AI queries can now combine memories with uploaded files, surveys, web search, Google Drive, and Confluence. NEXT decides which sources are relevant, applies the appropriate focus, and brings the evidence together in one answer. You can also attach a file and compare it directly with what customers have already shared.

Reuse previous workflow runs while testing​

Load an earlier workflow execution back into the editor with Copy to editor. Its outputs become available for testing, so you can iterate on a later step without manually reconstructing the state that led to it.

Highlights become memories​

The NEXT interface now uses memories for the insights previously called highlights. The new name better reflects that these items can contain synthesized understanding as well as direct excerpts, while keeping them available throughout search, chat, and workflows.

Duplicate workflows from the overview​

Every workflow now has a Duplicate action on the overview page. Create a complete, independently editable copy—including steps, connections, and settings—without opening the original workflow first.

You can now run a single step when its input is pinned, or execute a workflow only up to the step you are working on. Validate changes without waiting for the entire workflow, making complex processes much faster to build and debug.

See every execution exactly as it ran​

Workflow executions now preserve the version of the workflow that produced them. Even after steps are added, removed, or reordered, opening an earlier run shows its original graph and results, giving you a reliable historical view when troubleshooting.

Count matching memories inside a workflow​

The Read Memories step now exposes the total number of memories in focus as an output variable. Use the count in conditions, reports, or later calculations without adding a separate query.

Workflows move into the main navigation​

Workflows now appear directly in the left-hand menu instead of being hidden in Settings. Open a workflow from the menu, provide any required inputs, and run it from its own page. The item is shown only to people with the appropriate permissions.

Customer Twins can now be generated from the accounts and tags already organized in NEXT. Choose a source Twin, select the accounts you want to represent, and create multiple grounded Twins without configuring each one from scratch.

Let Customer Twins present workflow results​

Workflows can send generated assets, such as HTML reports, to a Customer Twin. During a conversation, the Twin can open the available content and talk the audience through it, combining a live persona conversation with the latest structured analysis from NEXT.

Predictable tag suggestions for keyword jobs​

Keyword tagging jobs now suggest unused tags in a consistent alphabetical order. When several tags in a group are available, the same one is proposed each time, making job configuration more predictable.

Find newly created API tokens instantly​

API tokens are now ordered by creation date with the newest token first. The creation action has also moved to the top of the page, so a token appears exactly where you expect immediately after creating it.

Pin the output of an earlier workflow step and reuse it while testing downstream changes. Generate an expensive report once, adjust the remaining logic, and rerun from the pinned result instead of repeating the entire workflow every time.

Remove workflow connections in one click​

Workflow connections now show a delete control when you hover over them. You can remove the connection directly instead of selecting a line and pressing Delete, reducing the risk of accidentally removing an entire step.

Run sanitization immediately​

Selecting Sanitize now on a record now starts sanitization immediately. Manual requests no longer wait for a scheduled background run, so the action behaves as expected when you need sensitive information removed right away.

Use newly created tags sooner​

New tags now appear in search suggestions much faster, including tags that do not have memories yet. This removes the delay between creating a tag and using it in searches, workflow output, or a newly configured teamspace.

Workflows can now follow multiple execution paths. Use True and False outputs from an IF step for conditional logic, or connect one step to several downstream steps that can run in parallel. Complex processes no longer need to be forced into a single linear sequence.

Inspect executions inside the workflow editor​

Open a past execution directly on the workflow canvas to see its sequence, the path it followed, and the result of each step. Keeping the run and its workflow graph together makes it faster to verify behavior and investigate unexpected results.

Focused chats honor your selection​

When you start a chat from a selected memory or report, NEXT now reliably keeps that selection as the conversation’s focus. Requests such as “summarize this” use the context you chose instead of falling back to the broader dataset.

Preserve document formatting in records​

Imported documents such as PDFs can now retain headings, lists, and section breaks instead of being flattened into transcript-style blocks. The result is easier for people to read and gives AI Chat clearer structure to work with.

Add sticky notes directly to the workflow canvas to explain logic, document decisions, or leave guidance for teammates. Notes can be moved, resized, and edited inline, and they never affect workflow execution.

Better insights from PDFs and presentations​

NEXT now uses a specialized analyzer for PDFs and PowerPoint presentations. Instead of looking for conversational quotes that may not exist, it extracts meaningful statements from headings, lists, and slide content, producing more useful memories from document-based research.

The new Publish Web Page step turns workflow output into a hosted page with a shareable link. Publish rich reports and dashboards without routing HTML through webhooks or maintaining a separate hosting tool. A workflow can update the page again on later runs, making recurring reports straightforward to maintain.

Cleaner workflow updates in Slack​

Slack digests from NEXT are now shorter and easier to scan. Story links no longer create unnecessary previews, annotations remain available on the Story page, and updates with highlights show a single player instead of filling the channel with multiple video embeds.

Tidy a workflow canvas in one click​

Use Tidy up workflow from the canvas menu to automatically arrange a workflow from left to right. Steps and branches align into a readable layout while sticky elements stay out of the way, helping large workflows remain understandable as they grow.

Start a chat from a shared highlight​

Shared highlight pages now let you open a new chat grounded in that highlight. Recipients can move directly from reading an insight to asking follow-up questions, without hunting for the original conversation or rebuilding its context.

The new Compose Text workflow step combines fixed text with values from earlier steps. Build predictable reports, emails, and summaries with placeholders while avoiding the cost, latency, and formatting variation of asking an AI model to assemble text whose structure you already know.

Use fast, focused AI calls in workflows​

The new Ask AI Model step sends a prompt directly to a model without starting an agentic chat. Choose a text response or a simple Yes/No result that can drive workflow decisions. It is a faster and more predictable option for focused generation and classification tasks.

Copy formatted chat answers into your documents​

Copying from chat now includes rich-text formatting alongside plain text. Paste an answer into Microsoft Word, Outlook, or Google Docs and tables remain real tables, while headings, lists, bold text, and links retain their formatting.

Pick Google Drive and Microsoft folders visually​

File-based subscriptions now use native Google Drive and Microsoft folder pickers. Browse and search Google Drive, OneDrive, and SharePoint locations during setup instead of finding and pasting paths or folder IDs manually.

The new Breakdown Table step creates summary tables directly inside a workflow. Group data by a tag group, add time-based columns such as This Quarter or Last Year, and choose how values should be calculated. The result is a Markdown table ready for reports, emails, documents, or further AI processing.

See how data moves through every workflow step​

Workflow executions now expose the context, output, and variables for each step. You can inspect what a step received, what it produced, and what it passed downstream, making complex workflows considerably easier to understand and troubleshoot.

Reference step outputs with stable names​

Workflow outputs now have readable references such as chat1.completion and googleFile1.markdown. These references remain stable when you reorder or remove other steps, so workflows are easier to maintain and less likely to break as they evolve.

Shared highlights provide more context​

Shared highlight pages now show the associated account beside the date and display descriptions for both text and media highlights. Recipients can understand who a highlight relates to and why it matters without switching back to the source item or asking for additional context.

Tasks can now recognize account-related criteria in your request and focus the analysis on matching accounts. Ask about a particular customer, region, or industry and NEXT uses the relevant account fields to narrow the underlying data automatically, giving you more relevant answers with less manual filtering.

Workflow executions update in real time​

The Workflow Executions page now refreshes automatically while you have it open. New runs and status changes appear in the background about every ten seconds, so you can follow progress and spot failures without repeatedly refreshing the page. Executions also have a dedicated URL that is easier to bookmark and share.

The HubSpot-native agent now imports emails alongside notes. Select Emails in the Type dropdown during setup and emails start syncing into your customer memory — subject line, message body, and original timestamps intact. You can filter emails with the same rules you use for notes, including direction (inbound or outbound) and deal stage. If your teams are having meaningful conversations over email that never made it into the intelligence layer, that blind spot is now closed.

Workflows can pause and resume​

You can now pause any workflow without losing your configuration. Hit pause from the workflow list or the detail page header, and execution stops cleanly. Resume when you're ready — same setup, same state. Previously, testing or troubleshooting a workflow meant deleting and rebuilding it. Now you can stop, investigate, adjust, and restart without any rework.

Resilient long-lived workflows​

Workflows can now run reliably across hours or days. When a workflow hits an execution limit — an LLM rate limit, a timeout, or any transient failure — it pauses at the last completed step and picks up from there once the issue clears. No full restart, no lost progress. This makes large-scale processes practical: backlog enrichment across thousands of tickets, multi-step approval chains, or any workflow where the work simply takes time.

Workflows can pull Google Docs as context​

Workflows can now fetch content from Google Docs on demand and pass it as clean Markdown into downstream steps. You can pull an entire doc or target a specific section by its heading — structure like headings, lists, links, and tables comes through intact. Organizational context that lives in Google Docs can now feed directly into workflow logic without manual copy-paste.

Real-time task progress​

Long-running tasks now show what's happening as they work. Instead of a static spinner, you see each step as it completes — retrieval, generation, enrichment, storage — with a percentage when available. If something fails, the progress indicator switches to a clear error state showing the last step and what went wrong. You know where things stand at every moment.

Smarter natural-language date ranges​

Date filters now handle natural language the way you'd expect. Typing "last 500 days" returns highlights from 500 days ago through today — not a single day 500 days in the past, which is what used to happen. Specific references like "yesterday" or "last Monday" still match that single day. Plain English, correct results, no special syntax needed.

Prefilled tasks via URL​

You can now deep-link into NEXT with a pre-loaded task prompt. Add ?prompt=Your text here to any NEXT URL and the task input opens with that text ready to go — the user can review and edit before running. This is especially useful for orchestrators building workflows.

We partnered with Llamaindex for PDF import and transcription — a major upgrade in processing quality. The capability also offers an EU deployment, so if you need EU data residency, PDF processing now stays within the region. Llamaindex has been added as an official sub-processor.

Clusters now quantify in one shot​

Running a task in cluster mode used to be a two-step process — you'd get clusters with a rough signal-strength estimate, then click Compute to get exact counts with linked details. That second step is gone. Clusters now return fully quantified in a single run, with final numbers and linked details from the start. We removed the Compute button and the estimate badges entirely.

Tasks show the exact reporting period​

Tasks now display the precise calendar date range of the data behind the analysis. A weekly report, for example, correctly reflects the intended Monday–Sunday window from your applied data focus — instead of drifting by a day or resolving relative ranges inconsistently. Small addition, but it makes summaries easier to verify and share with confidence.

We added a G2-native agent that brings G2 reviews into your customer memory. Review data now feeds directly into the system's reasoning and enterprise workflows — closing a blind spot if you've been tracking G2 separately.

Enhanced HubSpot-native agent: pipeline filtering​

The HubSpot-native agent can now filter note imports by deal pipeline. Add a filter during setup and only notes from that pipeline get imported, keeping your memory clean.

Focus stays sticky across follow-ups​

When you set focus to a specific dataset — a cluster, a recording, or manually selected highlights — that focus now stays sticky across follow-up questions. No automatic broadening, no mid-task prompts to re-confirm. The experience stays grounded in the data you chose, and you can narrow further without losing your starting point.

Workflow monitoring​

As you run more and more workflows, you need visibility into what's happening. Every workflow run now shows live status — running, succeeded, or failed — and you can open any run to inspect progress step by step.

Reactions in workflows 🙌​

Workflows can now react inside the tools your teams use. A new Reaction step adds emojis to GitHub and Linear tickets during a workflow run — eyes when processing starts, rocket when it's done. A backlog enrichment workflow can signal to engineering that a ticket has been analyzed; a triage workflow can mark that classification is complete. It's a lightweight, creative way to make workflow progress visible without anyone checking a separate system.

More stable Claude Desktop connection​

The MCP connection between Claude Desktop and NEXT is more reliable — faster cleanup of idle connections, fewer timeout errors, and smoother reconnections.

Multi-step workflows can now use the output of one task as the input for another task — for example, referencing {{task_1:message}} in a downstream step pulls in the exact result from the earlier one. This makes it straightforward to build workflows where each step builds on the last, instead of duplicating instructions or working around unreliable handoffs between steps.

Improved reliability of in-line source annotations for tasks​

Task responses now always include highlight annotations , even on longer answers — so every claim stays traceable back to the original source. Previously, some responses would drop their source tags entirely, making it harder to verify at a glance. Annotations now appear in-line as expected, and clicking one still opens the original highlight for full context.

Focus dialog always reflects the current selection​

Opening the Focus dialog now shows the actual applied focus — not leftover edits from a previous session. If you reopen the dialog after making changes or switching context, it's prefilled with what's currently active, so adjustments start from the right baseline instead of requiring you to mentally reconstruct what's set.

The Planner now runs each step in the right order — and only asks for your input once.

• Narrowing always happens first. The Planner used to occasionally skip the narrowing step and jump straight into building clusters, only to realize it didn't have the data it needed. That loop is gone. The Planner now narrows down the relevant highlights before doing anything else, so plans land on the first attempt instead of cycling through retries.
• One confirmation, not two. Confirming your focus used to be unreliable — in some cases you'd confirm once, only to be prompted again after clusters were already built. Switching between tasks or reopening one could also wipe out the response that was already on screen. Focus confirmation now sticks: once you confirm, your answer and clusters persist, and moving between tasks won't reset anything or trigger a second approval.

Focus shows exactly what data the agent is working with​

When a task runs, the focus indicator now reflects the actual scope of the data — not an assumption or a premature label.

• Scope appears only after narrowing completes. You used to see labels like "Entire teamspace" before the agent had finished determining scope, which made it unclear whether the results were truly teamspace-wide or still loading. The focus label now waits until narrowing is done — or until you set it yourself — so what you see is what the agent actually used.
• Teamspace-wide searches read clearly. When a task covers the full teamspace, the thinking panel now says so directly instead of showing "0" — which previously made it look like nothing was found. Narrower scopes still show exact highlight counts where available.

Smart suggestions when a focus returns no data​

When a focus combination returns zero data, the task now suggests ways to adjust — broadening the timeframe or relaxing less critical tags — so you can find relevant data without manually experimenting with different filter combinations. Instead of a dead end, you get a starting point for your next try.

Step-specific focus narrowing​

We released automatic focus narrowing for automation steps. A new checkbox in the step setup tells the agent to select the data it works with based on the current input rather than a fixed time window — so a ticket about onboarding pulls in onboarding-related evidence, not a random sample of your data. Off by default; turn it on when the step benefits from precision over breadth.

Smarter vector search defaults​

Vector search now only kicks in when it measurably improves on your existing filters. When tags and dates already capture the intent — like country tags on a "top pain points in France and the Netherlands" request — the extra semantic search is skipped. Generic terms like "customer" are ignored, and sentiment phrases like "wow moments of delight" map to the right feedback category instead of being treated as a literal search.

Follow-ups that preserve your context​

Follow-up tasks now inherit the active time frame, filters, and dataset by default instead of recalculating scope from scratch. A follow-up like "Which collections were mentioned?" stays within the original window and dataset until you explicitly change it ("switch to April," "include all projects").

Agents know when to act — and when not to​

AI decision steps can skip actions when there's nothing meaningful to add​

We added an AI decision step you can place between any two automation steps to evaluate whether the output is worth acting on. In backlog enrichment, for example, this means a comment only posts to your GitHub discussions, Jira tickets, or Linear issues when the agent actually found relevant customer evidence — no more empty updates cluttering threads.

The Task Planner only promises what it will actually do​

We fixed a mismatch where the task planner described steps — like narrowing the data scope or running a context search — that were turned off and never executed. The plan now reflects your configuration: what you read is what the agent will do.

No more duplicate or looping plan steps​

The task planner now detects and collapses duplicate steps. Rewrite-style requests produce a single step instead of stacking the same instruction multiple times, and re-running the planner won't re-add work that's already there.

Medallia integration​

Medallia survey responses now import directly into your workspace​

We added a Medallia integration so you can bring in NPS scores, post-interaction surveys, and journey-stage feedback alongside your existing support tickets, reviews, and call transcripts. Agents can draw on survey sentiment and support conversations in the same pass. For teams already using Medallia, experience data becomes an integral part of how you drive CX, marketing, and product performance — not a separate reporting layer.

Platform & experience​

Descriptions and prompt previews on automation step cards​

You can now add a short description to each automation step, displayed under its name. "Create task" steps show a prompt preview (up to ~200 characters) and the focus target — an account name or segment — so the intent is visible at a glance without opening the step.

Running and paused states now visible in the task list​

We added real-time status indicators to the task list. A spinner appears while an agent is running; an "Awaiting approval" pill shows when a step needs your confirmation. Both update live — no refresh needed.

Library pages load without resetting the sidebar​

The library navigation frame now stays in place while only the page content reloads. Switching between Home, Highlights, Data and Tasks is noticeably faster with no more sidebar resets between pages.

Following last week's Jira release, the same autonomous enrichment is now available for Linear. When a ticket is created or updated, the NEXT AI agent retrieves relevant customer evidence and comment directly in Linear with quotes, quantified data, and links back to the source. Instructions to connect Linear.

HubSpot deal notes in your intelligence layer​

HubSpot deal notes now sync into NEXT AI daily as highlights — notes from deal calls, objections, feature requests that were previously locked inside your CRM. Each note is linked to its deal and automatically tagged with owner, deal phase, pipeline, deal size, and more, so you can filter and cross-reference sales insights alongside surveys, support tickets, and transcripts. If you already have HubSpot connected, you may be asked to reauthorize once to grant an additional permission.

Survicate surveys now supported​

Connect Survicate and your survey responses flow into NEXT AI as highlights — free-text and multi-choice answers combined into a single, readable entry per response. Platform tags (iOS, Android) and language are applied automatically, and rating answers are normalized so filtering and scoring work consistently across all your survey sources.

Modes built into prompt templates​

You can now set the analysis mode (Clusters, Hypothesis, Podcast, and others) directly in a prompt template. When someone starts a new analysis from that template, the right mode is already applied — no manual selection, no risk of defaulting to the standard mode and getting the wrong workflow.

Filter-aware data retrieval​

When you set tags and date ranges in your Focus, the agent's search query used to restate those same constraints. The agent now recognizes the manually set filters and searches only for the context that's left.

Regression-ready auto-narrowing playground​

The data auto-narrowing playground is now a complete simulation environment. Define scenarios with sample prompts and the tags you expect, run them all at once, and get a clear pass or fail for each case. You can see exactly which tags were applied or missed, reuse the same scenarios whenever the narrowing logic changes, and trust the results — the playground runs the same data pipeline as production.

On-demand automation runs and bulk subscription cleanup​

Two operations improvements:

• scheduled automations now have a "Run" button that executes immediately with the current configuration — no need to wait for the next scheduled run or edit triggers to test.
• Cleaning up subscriptions no longer means deleting one at a time — select multiple, confirm, and remove them in a single step.

Jira automation is here. Whenever an issue is created or updated, the NEXT AI agent automatically retrieves, compiles and computes relevant customer evidence, then posts it as a comment on the ticket — customer reels, real quotes, quantified data, and links back to the source. Every ticket arrives pre-loaded with the context your team needs to make confident decisions, without anyone having to search for it.

Connect your Jira instance via OAuth, select the projects you want to monitor, and the agent handles the rest.

Focus you can trust​

When you set a focus, you're telling the agent exactly what data matters. This week, four changes make sure it listens — and that you can always see what's in scope.

• Highlight count stays visible. Previously, confirming your focus hid the highlight count and left the area feeling cramped. The focus area now keeps the count on screen in a cleaner two-line layout with better spacing between tags. At a glance you know how much data the agent is working with and where that data is coming from.
• Your selections stick. Setting a focus manually used to be undermined by auto-suggestions that widened the scope and replaced your choices. Suggestions now layer on top of what you already selected — your tags and filters remain the foundation, and the system adds from there.
• Smarter defaults from tag groups. When setting the focus, the system now recognizes which tags belong together. Instead of requiring every tag to match (which over-narrows results), related tags are offered as alternatives — for example, "Product Issue or Pain Point." You get broader, more useful defaults without having to rearrange chips by hand.
• Consistent refinement across every flow. Behind the scenes, the way the agent narrows focus has been rebuilt into a single, dedicated step. Your existing scope always serves as the starting point, and any refinement — whether it's a new tag, a date range, or an updated description — is added on top rather than rewriting what you set. This applies uniformly, so focus behaves the same way no matter where you're working.

New mode: Evidence​

Until now, every analysis in NEXT AI assumed you already had a hypothesis to test or a structure to cluster around. But many questions start broader: "What are customers saying about checkout?" or "Show me everything related to pricing." Forcing those into modes designed for data crunching added friction and muddied the results.

A new Evidence Mode gives you a dedicated space for open-ended exploration. Point the agent at a topic and it gathers all relevant evidence into a single, organized view. No hypothesis required, no clustering to compute — just everything that matches, in one place. It's available alongside Hypothesis and Cluster modes, so you choose the approach that fits the question.

Excel exports for clusters​

We've upgraded highlight downloads in clusters from a CSV—workable, but clunky to open and format—to native Excel (.xlsx) files. Same fields, same column order, but the file opens cleanly in Microsoft Excel or any spreadsheet app with no import dialogs or encoding issues.

Improvements to agent reasoning​

In Evidence and Hypothesis Mode, the agent now draws a hard line between structure and substance: only steps backed by actual evidence produce meaningful output. When there's nothing to cite, the response says so plainly instead of generating plausible-sounding detail (hallucinating). You can trust that every claim in the output traces back to a real source — and when the data isn't there, you'll know.