Claude’s new trick: finished files, not just advice

Anthropic moves from chat to deliverables—while warning users to watch their data.

Anthropic is giving Claude a job beyond talking: creating real documents, spreadsheets, slide decks, and PDFs directly from a chat. The company says the feature is live in preview for Max, Team, and Enterprise customers, with Pro access “in the coming weeks,” and it can be switched on in settings. Anthropic details the rollout in Anthropic’s file creation announcement.

Claude no longer stops at suggesting a pivot table or a slide outline. It generates a multi-sheet Excel model with formulas and charts, turns meeting notes into a formatted report, or converts a PDF into a deck. You describe the goal, upload data if needed, iterate in chat, and download or save to Google Drive. The core claim: fewer hops between idea and artifact.

What’s actually new

Under the hood, Anthropic gave Claude access to a private compute environment where it can write code and run programs during a session. That lets the model perform data cleaning, statistical analysis, chart generation, and file assembly rather than merely drafting text for you to paste elsewhere. It’s agentic behavior, but scoped to file creation.

Enable “Upgraded file creation and analysis” in Settings > Features > Experimental. Then ask for an output—say, a budget tracker with variance analysis or a research memo with embedded visuals. Claude orchestrates the steps, surfaces the draft, and refines it based on your prompts.

Evidence and scope

Anthropic’s own examples span three buckets: turn raw data into analysis with charts and written insights; build spreadsheets with working formulas and multiple sheets; and cross-format work like producing slides from a report or a spreadsheet from invoices. Early coverage from consumer tech outlets aligns with that framing, emphasizing finished, downloadable files rather than in-chat artifacts. The promised capability is broad but pragmatic: structured business documents first, creative layouts later.

You can hand Claude CSVs from a sales system and request cohort analysis with a one-page summary. Or feed it a dense PDF and ask for a 10-slide executive brief. It’s still you setting the objective and giving feedback; Claude handles the glue work.

Competitive context

This pushes Claude into the same arena as ChatGPT’s agent modes and Google’s deep integrations with Workspace, but with a different emphasis. OpenAI has showcased web navigation and tool use; Google leans on collaborative editing inside Docs, Sheets, and Slides. Anthropic is staking out “deliverables from chat” as a first-class outcome, with export rather than co-editing as the default.

Owning the moment when draft becomes file is a subtle land grab. If teams get comfortable generating a “good enough” model or deck inside Claude, traditional app time shrinks and switching costs rise. The question is where the quality and control thresholds sit for different functions.

The enterprise calculus

For buyers, the math is straightforward: a single conversation could compress analyst, designer, and project-manager handoffs into one flow. Marketing and consulting teams may adopt quickly because their documents favor narrative and visuals. Finance and legal will likely test longer, given precision and audit needs.

Where the output is a starting point—first-cut forecasts, take-away decks, monthly trackers—Claude’s speed will tempt. Where the file is the record—board packs, contracts, audited models—teams will keep human-in-the-loop standards and stricter review gates.

Limitations and risk

Anthropic itself flags the big caveat: enabling file creation gives Claude internet access during the process, which may put data at risk. The company advises monitoring chats and avoiding sensitive inputs. That warning should trigger policy updates—classification of permissible data, redaction guidelines, and logging for prompt and output handling.

There are also practical constraints. Conversational assembly can obscure the steps taken to produce a chart or a metric, so reproducibility matters. Teams should pair Claude’s outputs with a light “workpaper” discipline: capture inputs, assumptions, and prompts next to the deliverable. And remember that “private computer environment” is still a vendor-managed runtime; treat it as such in risk models.

The near-term work shift

If the feature holds up in practice, the biggest change is cognitive. Much of office work is overhead: formatting, chart choices, pagination, and consistency. Claude abstracts that layer so people can focus on content and decisions. The trade-off is control versus throughput. Specialists will still prefer native tools for complex, high-stakes work; generalists will default to chat for speed.

Today’s sweet spot is multi-step, medium-fidelity outputs that benefit from quick iteration: financial models with scenarios, KPI dashboards with light analysis, condensed research briefs, and client-ready decks. Tomorrow’s question is governance—how to scale this without sacrificing accuracy, security, and accountability.

Why this matters

• Execution, not advice, is the new moat. AI vendors are competing on who can collapse the most steps between prompt and deliverable inside trusted workflows.
• Policy will dictate ROI. Productivity gains hinge on clear rules for what data can be used, who reviews outputs, and how results are traced and audited.