This review is part of a larger series of LinkedIn newsletters titled AI in Market Research: Reviews of AI tools, platforms, and solutions that market researchers should use today.
When people hear about Claude, they still ask whether it beats ChatGPT. The honest answer hasn’t changed: it depends on the job in front of you. What has changed is how much of that job Claude can now take on.
What’s new since the last review: The original review ran in July 2024. Since then Claude added Projects, Artifacts, web search, connectors, and an agentic desktop app called Cowork, and the model line moved through several generations to Opus 4.8, Sonnet, and Haiku. The one change researchers need to act on is data privacy: the old “Claude stores nothing” claim no longer holds on consumer plans. Details below, along with a new section on Cowork.
Two years ago the knock on Claude for researchers was the missing GPT builder. If you wanted a synthetic buyer persona you could reuse across a study, ChatGPT’s custom GPTs were the cleaner path. That gap is mostly closed. Claude Projects give you a persistent workspace with its own instructions and reference files, so you can load a study’s transcripts, discussion guide, and stakeholder context once and keep working against it without re-explaining the setup each session. Skills let you save a repeatable procedure, a coding scheme, or a report template and reuse it. Between the two, building a reusable research assistant no longer requires wiring the Anthropic API into a tool like Bubble the way the original review described.
The old file limits are gone in practice too. The 2024 review walked through the five-attachment cap, the 10MB ceiling, and the workaround of pasting interviews into one document. With Projects and context windows now reaching hundreds of thousands of tokens, you can hold a full qualitative study in a single working space. Web search, a toggle in the interface, means secondary research is less constrained by a training cutoff than it was in 2024.
The original review also flagged that Anthropic had promised CRM integrations and collaboration “but the specifics are yet to be seen.” They arrived. Connectors now link Claude to Google Drive, Gmail, Slack, Notion, Microsoft 365, and many other services through an open standard called MCP. Team and Enterprise plans add admin controls, shared projects, and SSO.
Now the correction that matters most. The original review praised Claude for not storing user data on any tier, free included. That is no longer accurate. Since Anthropic’s late-2025 consumer privacy changes, Free, Pro, and Max users can allow their chats to be used to improve Claude, and if that setting is on, Anthropic may retain de-identified data for up to five years for model training. If that setting is off, consumer chats return to the shorter retention path and are not used for training in the same way. The business tiers are the exception: Claude for Work, along with Education, Government, and API access, are not used to train Claude’s models in the same way as consumer chats.
For a research firm handling client material under NDA, the practical rule is simple: keep client data on a Team or Enterprise account, and never drop it into a personal Free or Pro chat. This is a genuine reversal from where things stood in 2024, so check your team’s account settings before the next study.
Everything the original review liked about Claude for day-to-day research work still holds: drafting and improving survey and interview questions, categorizing and summarizing open-ended responses, and outlining and editing reports.
Ratings: Core Claude Product
| Dimension | Rating | Rationale |
|---|---|---|
| Usability | 5 / 5 | Still the most pleasant of the major chat tools to write and think in. |
| Power | 5 / 5 | The GPT-builder gap that cost a star in 2024 is closed by Projects and Skills, and connectors plus web search extend Claude’s reach well past the old limits. |
| Flexibility | 5 / 5 | Handles qual and quant support tasks and adapts to a team’s own templates and instructions. |
| Cost | 5 / 5 | Pro is around $20 a month, in line with ChatGPT Plus; Team is around $25 a seat with a five-seat minimum. |
| AI Washing | None | It’s pure AI. |
Claude Cowork
Most tools in this series live inside a chat window. You prompt, it answers, you prompt again. Claude Cowork works differently. Anthropic built it for the work that piles up around a research project: a folder of forty transcripts nobody has sorted, or a report that needs assembling from six source documents before anyone can read it end to end. You point Cowork at a folder, describe the outcome you want, and it reads the files and hands back a draft for you to review.
Agentic tools like this one follow a pattern we see across B2B tech adoption: the capability shows up ahead of the governance layer that makes it safe to lean on. Here is how that plays out for a research team.
Folder-level permissions
You authorize Cowork one folder at a time. Point it at a study’s working directory and it can read, edit, and create files inside that folder, and it’s designed to leave anything outside that scope alone. Any agent that can execute code carries some risk that a misconfiguration broadens its reach, so treat the boundary as a safeguard rather than a guarantee.
A plan before it acts
Before making changes, Cowork lays out its plan and waits for approval. You can watch each step or walk away and come back to a finished result. For reorganizing a shared drive of interview recordings and discussion guides, that checkpoint matters. You’re not handing over research data and hoping for the best.
The assembly work
Give it a folder of source documents, interview notes, stakeholder decks, and prior reports, and Cowork produces a structured first draft that synthesizes across sources instead of summarizing one file at a time. If you’re building a message-testing report from a stack of verbatims and stakeholder input, that’s the work it was built for. It also handles the adjacent grunt work of renaming, deduplicating, and sorting a messy project folder before analysis starts.
Connectors
Connectors turn Cowork into part of the research stack instead of a local file tool. Point it at a Google Drive folder of transcripts and it can also read Slack threads or pull client emails from Gmail. Add OneDrive, SharePoint, DocuSign, or calendar data and it covers much of what a researcher touches between kickoff and final readout, though research-specific systems remain less represented.
Scheduled tasks
Cowork runs on a schedule as well as on demand. A recurring Friday digest that sweeps a folder of new interviews, pulls out themes, and drops a draft into your team’s working drive is the right mental model, as is a weekly win-loss summary or a rolling review of NPS verbatims. Where it’s still early is operational control: there is a Scheduled tasks page and per-task pause or delete controls, but not the kind of tenant-level calendar or one-click “pause everything” control a research director may want before a major study locks.
Browser automation
Cowork can also work beyond static files. Used well, it acts like a junior desk researcher with more patience than any human: checking a shortlist of competitor sites for copy changes, or pulling public product documentation into a local corpus. Point it at a market landscape spreadsheet and it can help keep that current too. The tradeoff today is that you’re trusting its navigation choices without the click-by-click trace you’d get from a human researcher’s notebook, and fine-grained controls over which domains it may visit or how long it retains what it captures are still emerging.
Where Cowork Could Improve
Cowork’s enterprise controls are still catching up to its capability. Anthropic now offers audit logs, a Compliance API, and enterprise data export and integration pathways for Claude, but granular Cowork actions and content still appear unevenly represented in those controls. For a firm operating under client data-handling requirements, that is a real gap; confirm what your environment surfaces before trusting Cowork with anything under NDA.
It’s also a different surface from Claude’s chat interface, which brings an onboarding curve. Learning the folder-trust model, and how to write a good outcome-level instruction rather than a step-by-step prompt, takes some adjustment. Teams will want to build their own pattern library for good Cowork tasks the way they did for good prompts.
Finally, the connector catalog is strong for general productivity tools but still shallow for the research-specific platforms many teams live in. Survey platforms, CX systems, and bespoke insight repositories still sit mostly outside the first-class connector list. Scheduled tasks and browser automation push Cowork toward repeatable research operations, but without central visibility into everything running across client folders or consistent policy controls at the connector and domain level, larger firms will still treat these features as experimental rather than core.
Ratings: Claude Cowork
| Dimension | Rating | Rationale |
|---|---|---|
| Usability | 4 / 5 | The folder-permission model and plan-before-execute step are well designed, but the shift from prompting to outcome-framing takes practice. |
| Power | 4.5 / 5 | Multi-step execution across files, folders, connectors, scheduled tasks, and browser automation is a different capability tier than the chat-based tools in this series. |
| Flexibility | 4 / 5 | Strong for document assembly, file organization, and cross-source synthesis; still maturing on the compliance, audit, and connector coverage larger research organizations will want before full rollout. |
| Cost | 4 / 5 | Bundled with paid Claude subscriptions starting around $20 a month, with deeper admin controls at the Enterprise tier, rather than sold as a separate line item. |
Conclusion
The judgment calls in a research project were never the bottleneck. The filing, sorting, recurring summarization, and drafting around them were. Claude has spent the last two years getting better at exactly that layer, and Cowork pushes it furthest. The division of labor is the point: a researcher still decides what the findings mean, and the agent clears the runway to get there. Just mind the privacy setting, because the account you use plays a major role in how Anthropic may handle that data.
At Cascade Insights®, we put these tools through real B2B studies as they mature; if you’re weighing where Claude and Cowork fit in your own research stack, that’s a conversation we’re glad to have.
Last updated: 7/14/2026