Customer support has quietly become one of the most important proving grounds for real world artificial intelligence. When support stops being a cost center and starts driving product growth, you get a preview of what an always on, automated market research engine can look like in practice. Loom, now part of Atlassian, is a concrete example of that shift. Its embedded AI support now resolves 80 percent of inquiries without human escalation, demonstrating how scaled automation can reshape both customer experience and support operations.
From ticket queues to an AI driven growth engine
For most of the past decade, digital support followed a familiar pattern. Customers opened tickets, human agents responded, and any insight into user needs was extracted later, if at all, through manual tagging and quarterly reviews. The rise of chatbots promised efficiency, but early systems were brittle, scripted, and rarely trusted with anything beyond basic FAQs.
Loom took a different path once it integrated Atlassian Customer Service Management and embedded an AI support experience directly into its product. The AI agent now resolves about 80 percent of help seeking inquiries without human escalation, drawing on Loom documentation and troubleshooting content to answer questions and ask clarifying follow ups when needed.
This is not just a convenience feature. The shift is associated with an 11 percent reduction in churn across the customer base and a 10.7 percent increase in video creation after the launch of the in product AI support widget. In the first year alone, more than fifty thousand support conversations ran through this embedded AI agent.
Those are important deltas for any software as a service business. Churn reduction compounds over time, and added video creation maps directly to Loom’s core usage and revenue metrics. The key point is that AI support is no longer measured only on deflection and cost. It is evaluated on retention and engagement, which aligns support operations with product and growth strategy rather than leaving them as a reactive service function.
Support as a structured insight stream
The deeper innovation is not just resolution rates. It is how Loom treats every AI mediated interaction as structured research data. Integrated with Atlassian Customer Service Management, Loom’s AI support engine unifies ticket based conversations and in product help sessions into a single analytics stream that can be segmented by cohort, feature, and outcome.
Every AI support interaction becomes structured research, unified into a single, segmentable analytics stream.
Routine questions are absorbed by automation, which lowers escalation volume and frees human specialists to focus on complex qualitative issues that require judgment and nuance. From an insight perspective, that unified stream behaves like a high throughput sensor network. Each interaction is time stamped, labeled, and tied to metadata such as plan type, usage patterns, and eventual outcomes like expansion or churn.
The same system that answers a question about how to share a video is also quietly logging that a specific feature confused a new user segment in a particular week. Over thousands of conversations, patterns emerge in friction points, missing documentation, and emerging feature demand.
This is where generative models make a difference. Instead of relying only on manual tagging, AI can cluster conversations by theme, sentiment, and user goal, then surface anomalies such as a sudden spike in questions about a recently shipped feature. That lets product and research teams treat support as a living funnel analysis tool and as an early warning system for churn, not just a log of solved tickets.
How async research became Loom’s default mode
Loom extended this automation mindset into customer research itself. Rather than centering the process on live thirty minute calls and long surveys, the company has built an async first workflow that mirrors how its customers already use the product.
Loom was acquired by Atlassian in 2023 for about nine hundred seventy five million dollars, which reinforced its role as the async video layer inside a broader collaboration suite. Public commentary and product blog posts outline a consistent research pattern. A small research team with a single digit headcount manages input from millions of users by leaning on three main async modes.
The first is recorded video reactions, where users respond to feature concepts or interface changes by recording a Loom that captures both their screen and narration. The second is in product voice notes, short audio clips that allow users to describe what they were trying to do at the moment they hit friction. The third is AI moderated text follow ups. When a user clicks something like I had trouble with this, a contextual AI agent asks two or three targeted questions to capture the why while the experience is still fresh.
This approach replaces many traditional research rituals. An async customer interview becomes a structured conversation between a participant and an AI moderator, completed in two to ten minutes rather than a scheduled half hour call. Completion rates are often higher than live sessions, frequently above eighty five percent, because participants can respond on their own schedule without installing new tools or joining a meeting.
Transcripts flow directly into AI systems that summarize, tag, and link qualitative feedback to behavioral telemetry. In practice, Loom can circulate feature mock videos to beta cohorts, collect reaction Looms that show exactly how people interact with new flows, and synthesize usability issues in hours rather than weeks.
The output is richer than text survey responses because it combines screen capture, tone of voice, and spoken context, while still scaling far beyond what a small human research team could handle manually.
Generative AI across the research life cycle
Although the most visible parts are support chat and video responses, the same generative capabilities extend across the entire research life cycle. Before a study even starts, AI assistants can summarize prior literature, internal memos, and previous interview findings related to a problem space.
During data collection, they moderate text based interviews, ensure coverage of key topics, and adapt questions based on earlier answers. Afterward, they process long transcripts, code open ended responses, and cluster themes without weeks of manual effort. The connection back to product telemetry is particularly important.
When a user records a Loom describing why they abandoned a workflow, the AI can link that qualitative explanation to click level data from the same session. That closes the loop between what people say and what they do. The same underlying models also power Loom specific features that turn raw video into structured artifacts.
Loom AI now provides automatic titles, summaries, chapters, and AI workflows that convert a recording into a text document, bug report, or message that can be sent directly into tools like Jira and Linear. For research, that means every participant recording can also become a concise brief for designers, engineers, and stakeholders who do not have time to watch the full video.
Looking forward, teams are starting to experiment with synthetic personas and simulated datasets that extend empirical findings. A model trained on real support and research data can generate realistic but privacy safe variants of user behavior for scenario planning and stress testing designs. Used carefully, this can accelerate early stage exploration, as long as decisions remain grounded in actual observed data.
Why this matters for businesses and the research industry
For technology companies, Loom’s approach highlights a shift from episodic research to continuous sensing. In the older model, teams ran quarterly surveys, occasional usability tests, and ad hoc customer interviews. Insights arrived in bursts and often went stale before they could influence roadmaps.
In the emerging model, every support conversation, every I had trouble click, and every reaction video contributes to an always current map of user needs. This has several implications.
- Product teams can iterate faster because they see the impact of changes in near real time, not after the next scheduled round of research.
- Support operations become a strategic function, feeding prioritized backlogs with evidence rather than anecdote.
- Smaller research teams can cover much larger user bases, since AI absorbs the repetitive work of scheduling, transcription, and first pass analysis.
- Investors and leaders can tie research activities directly to hard metrics like churn, activation, and feature adoption, as Loom has demonstrated with its churn and engagement improvements.
The model also challenges traditional market research vendors. Panels, long surveys, and extended interviews will not disappear, especially for exploratory work or regulated industries. But as more product companies embed AI into support and feedback channels, a growing share of insight will come from passively collected, in context data rather than standalone studies.
Risks, limitations, and what to watch
None of this is automatic progress. There are real risks and open questions. Data quality is an obvious concern. If the majority of conversations are handled by AI, any systematic misunderstanding or hallucination can mislead both users and internal analytics.
Even with an 80 percent AI resolution rate, the remaining 20 percent may hide edge cases and nuanced needs that are easy to overlook. Maintaining a healthy feedback loop between human agents and AI systems is critical. Bias is another issue. Async research flows favor people who are comfortable recording themselves or typing detailed responses.
Loom mitigates some of this by offering multiple modes, including text first interviews for users who do not want to turn on a camera or microphone, but some voices will still be under represented. Organizations need explicit strategies to reach less vocal or less engaged segments.
There are also governance and privacy questions. Turning every interaction into structured data is powerful, but it raises expectations around consent, data retention, and acceptable use. Companies must be clear about what is logged, how it is anonymized or aggregated, and who can query it. Regulatory scrutiny of automated decision systems is increasing, and market research is unlikely to remain exempt.
Finally, there is the risk of over automation. If teams rely too heavily on AI summarized insights, they may miss the subtle context that emerges only in occasional deep, human led conversations. The most effective setups pair continuous AI mediated sensing with targeted, high depth studies led by experienced researchers who can challenge assumptions and interpret signals in context.
Key takeaways and what comes next
Loom’s work with Atlassian Customer Service Management shows that AI embedded support can do much more than cut ticket volume. It can improve retention and product usage while turning support channels into a rich stream of structured market insight.
Combined with async first research workflows built around video, voice, and AI moderated text, a small team can keep up with millions of users in close to real time. For businesses, the lesson is straightforward. Any interface where customers ask for help, express confusion, or react to changes is also a potential research instrument.
The question is whether you treat those moments as one off transactions or as part of an integrated insight engine that informs product, marketing, and strategy. Over the next few years, expect to see more companies follow this pattern. Support, research, and product analytics will converge into shared platforms where generative AI handles the heavy lifting of summarization, coding, and pattern detection, while humans focus on framing the right questions and making the consequential calls.
Those who get the balance right will ship better products faster and understand their customers more deeply, not because they run more surveys, but because every interaction counts as research. For now, Loom offers a clear preview of that future and a reminder that the most transformative AI systems often emerge not from flashy separate tools but from careful integration into everyday workflows.
Conclusion
The way companies learn from customers is changing fast. For years, market research meant surveys, interviews, and occasional deep dive studies that took weeks or months to complete. Now product and growth teams expect fresh insight every day, not once a quarter. Systems like Evidence Loom, built on modern AI research tools such as Perplexity Sonar, point to a future where conversation data from online communities becomes a continuous source of structured intelligence rather than a noisy stream of anecdotes.
From classic market research to continuous insight streams
Traditional market research was designed for stability. Teams defined a study, recruited participants, asked a set of questions, and produced a report that guided decisions for months or years. That approach works well when markets move slowly and communication channels are limited. It struggles when customer expectations shift weekly and new competitors appear overnight.
The first big change came with digital analytics and social listening. Web tracking, search trend analysis, and social media monitoring gave companies more frequent signals, but the data was often shallow. A spike in traffic or a trending keyword hinted at interest, yet rarely explained underlying problems or motivations.
Community platforms changed the game by capturing long, unfiltered conversations about products, tools, and everyday frustrations. Over the past few years, market researchers and growth teams have increasingly treated these conversations as an informal focus group at scale. Tutorials now show how to scrape hundreds or thousands of discussion posts with automation tools, then feed them into large language models for synthesis. This shift laid the groundwork for dedicated AI agents that turn raw talk into structured insight.
What systems like Evidence Loom actually do
Evidence Loom sits in a fast growing category of AI systems that automate qualitative market research by continuously monitoring and interpreting large volumes of discussion data. The idea is straightforward but powerful. Instead of copying posts into spreadsheets and reading them manually, an AI agent collects them, filters out noise, and organizes useful content into themes, sentiment patterns, and decision ready summaries.
Open projects and commercial platforms already illustrate how this works in practice. One project focuses on trending topics and sentiment, fetching popular articles and discussions, extracting key phrases, and checking interest against search trends to validate demand. Commercial tools promise to scan thousands of posts and comments, then deliver structured reports with buying signals, competitor mentions, and feature comparisons in under sixty seconds. Some platforms even track dozens of communities in real time and feed insights into a unified database of pain points and opportunities, with users reporting weekly time savings in the range of ten to fifteen hours compared with manual research.
Evidence Loom follows the same general pattern, but with a stronger focus on turning this automated analysis into ongoing market research for product and growth teams. Where a traditional study gives a snapshot, a system like this builds a live feed of customer language, emerging needs, and reactions to new features or campaigns. That feed is organized and labeled so teams can query it just as they would query a data warehouse, but with qualitative evidence instead of pure numbers.
How the underlying AI pipeline works
Although implementations differ, the core pipeline behind Evidence Loom and similar agents tends to have four stages.
1. Collection
Conversation data is gathered from targeted communities using application programming interfaces, automation workflows, or dedicated connectors. Tutorials show automations that fetch hundreds of posts for specific topics, filter out promotional content, and focus on genuine problem descriptions. Other platforms continuously monitor more than fifty communities and related sources such as review sites and work marketplaces, storing everything in a structured repository.
2. Cleaning and filtering
Not all posts are useful. Effective systems remove spam, short or empty messages, and content that does not match defined themes. Some workflows apply basic rules such as minimum engagement thresholds or recency windows to keep signals current. Others rely on language models to distinguish real pain points from chatter.
3. Interpretation and synthesis
Once the data is clean, language models step in. Open source projects already demonstrate pipelines where models detect sentiment, classify posts into thematic clusters, extract recurring issues, and suggest potential solutions or content ideas. Commercial tools extend this by highlighting buying signals, tracking competitor mentions, and generating strategic recommendations for marketers and product leaders.
4. Delivery and integration
The final step is to make insight usable. Some workflows push synthesized findings into documents, spreadsheets, or dashboards that teams can browse or query. More advanced platforms turn conversation data into living knowledge bases that can be explored through natural language search across topics, segments, or time periods.
Evidence Loom fits into this ecosystem by focusing on structured, decision ready summaries that map directly to product and growth questions. Instead of handing over raw transcripts or vague sentiment charts, it organizes findings around issues, opportunities, segments, and potential feature responses.
The role of Perplexity Sonar in this shift
Perplexity Sonar is part of a broader move to combine retrieval, large language models, and agent like behavior for research tasks. Built on modern open models, it can ingest and connect information from many sources at once, then surface patterns through conversational queries. For systems like Evidence Loom, that means the underlying research engine is not just summarizing text, but actively looking for corroborating signals, conflicting evidence, and emerging themes across a wide set of documents.
This matters for trust. One of the central challenges in AI powered research is hallucination and shallow synthesis. Tools that only summarize a single source can miss important context or invent details. By contrast, a retrieval oriented engine can ground its answers in multiple documents, highlight specific passages, and reveal the reasoning behind a pattern. When that engine is wrapped in a workflow that continuously monitors conversation data, the result is closer to a real research partner than a static dashboard.
Why this matters for product and growth teams
For product managers and growth leaders, the promise of tools like Evidence Loom is less about novelty and more about changing how decisions are made day to day. Several platforms already report substantial time savings when teams switch from manual reading and spreadsheet tagging to automated pipelines. Makers of one such tool describe marketers and founders saving more than ten hours per week by automating competitive analysis, sentiment tracking, and opportunity discovery across community discussions and review platforms.
Beyond saving time, continuous AI research can help teams stay closer to real customer language. When conversation feeds are organized by problem themes, work flows, and outcomes, teams can hear how people actually describe their struggles rather than relying solely on internal terminology. Tutorials show workflows that categorize posts along dimensions such as experience, tooling, work flow, and cost, then turn the results into visual maps of opportunities and pain points. When this kind of analysis becomes a routine input into planning, feature design and messaging tend to align better with what customers actually care about.
There is also a structural benefit. When market research shifts from episodic projects to an ongoing feed, insight becomes part of everyday work instead of a rare event. Growth teams can track sentiment shifts after a campaign, product teams can monitor responses to new releases, and leadership can see emerging issues before they become crises. Over time, this can create a culture where decisions are continuously informed by real stories and patterns rather than occasional high level reports.
Risks, biases, and limitations
Despite the promise, automated AI market research is not a magic mirror of reality. It carries several risks that experienced teams need to manage.
First, communities are not representative samples. Conversation platforms over index certain demographics, interests, and regions. If an agent is trained on a narrow set of communities, it may generalize from a specific cluster of power users or hobbyists and miss silent segments with very different needs.
Second, language models are still fallible. Projects that rely on models for classification and summarization show that they can misinterpret sarcasm, over simplify nuanced complaints, or collapse distinct themes into a single category. Automated sentiment analysis can exaggerate negative or positive feelings if context is thin. Without human review, this can lead to overconfident conclusions.
Third, ethical and privacy concerns matter. Even when data is public, teams need clear standards for what is appropriate to collect, store, and use. Blending conversation data with internal records can raise questions about consent and transparency. Regulatory expectations around data use continue to evolve, and mature organizations will want governance frameworks that match those expectations.
Finally, there is a risk of over automation. When research becomes a dashboard, teams may stop talking directly with customers. In practice, the best outcomes appear when automated insight complements interviews, surveys, and direct conversations, not replaces them. Human researchers are still needed to design good questions, interpret subtle signals, and challenge assumptions that agents cannot see.
How to adopt these systems responsibly
For teams evaluating Evidence Loom or similar tools, a few practical principles can keep adoption grounded.
1. Start with clear questions
Define the decisions you want to inform. For example, early stage product validation, feature prioritization, or messaging refinement. Then configure the system to focus on communities and themes that match those questions.
2. Design transparent pipelines
Document how data is collected, filtered, and analyzed. Make sure stakeholders can see what sources are included, what rules are applied, and where models are used for classification or synthesis.
3. Combine automated insight with human review
Treat AI generated findings as a starting point. Build routines where researchers or product owners review synthesized themes, check examples, and adjust labels before major decisions are made.
4. Monitor bias and coverage
Regularly inspect which communities and topics your system is over sampling or ignoring. Adjust sources and weights to avoid tunnel vision on a single segment. Consider periodic complementary studies, such as targeted interviews, to test whether automated signals match reality.
5. Align with data ethics and governance
Work with legal and compliance teams to define acceptable use of public conversation data. Ensure storage, access, and sharing are controlled, and communicate clearly with teams about how and why these systems operate.
The road ahead
Evidence Loom and similar agents are part of a wider trend that is likely to reshape market research over the next few years. Instead of treating insight as something produced on demand by dedicated projects, more organizations will build ongoing intelligence layers that sit beneath day to day work. In that world, the most valuable research tools will be those that combine strong retrieval engines, capable language models, and honest, transparent workflows that show their reasoning rather than hiding it.
If teams approach these systems with curiosity and rigor, they can move beyond one off reports and build a living picture of their market that evolves alongside their product. The smartest organizations will use automated conversation analysis not to replace researchers, but to free them to ask better questions, design sharper studies, and focus on the decisions that matter most, while staying grounded in the voices of real people who gather in places like reddit
