Atlas Now Speaks Files, Images, and Voice—Without Losing the Plot
Attachments, multimodal context, and voice land in Atlas conversations—governed, scoped, and ready for real work.

Key takeaways
- Atlas conversations now support file attachments (images, PDFs, spreadsheets, audio) with full context retention across turns—no more copy-pasting or losing the thread.
- Voice integration (speech-to-text and text-to-speech) is live, with Deepgram as the default and Gemini fallback, all behind a single provider abstraction for future flexibility.
- Attachments are conversation-scoped, governed by permissions, and never auto-promoted to org-wide Knowledge—because context should be controlled, not viral.
- Rhonda (the anonymous surface) gets limited access: images and small docs only, with hard caps and short retention. No voice. No surprises.
You can now attach files to Atlas conversations—without breaking governance
Here’s the reality: most AI chat tools treat attachments like an afterthought. You upload a PDF, the model reads it, and then—poof—it’s gone. Or worse, it gets dumped into some nebulous "knowledge base" where it lives forever, ungoverned, unscoped, and ready to haunt your next unrelated prompt. That’s not how work happens.
Atlas now lets you attach files directly to conversations. Images, PDFs, DOCX, spreadsheets, CSVs, even audio. The model gets the full context, and it sticks around for later turns. No more copying tables into prompts. No more losing the thread because the attachment vanished after the first reply. Just work.
✦ How Atlas handles attachments
Small files: inlined in full. Large files: chunked and embedded, with a summary always inlined. Hybrid context means the model never loses the plot, even if the file is 50MB. Magic-byte sniffing on every surface, with real antivirus scanning only on the anonymous Rhonda surface (because no one should trust random uploads). Attachments are conversation-scoped—no auto-promotion to org-wide Knowledge. Governance isn’t an afterthought; it’s the default.
Voice lands—but not the way you’ve seen it before
Voice in AI tools usually feels like a gimmick. You talk, it transcribes, and then you’re back to typing. Or worse, it’s a one-way street: you can speak, but the AI can’t talk back without sounding like a robot reading a script. Atlas does it differently.
Speech-to-text and text-to-speech are now live, with Deepgram as the default and Gemini fallback. All behind one provider abstraction, so we can swap engines later without breaking your workflow. The goal isn’t just to add voice—it’s to make it useful. That means real-time transcription, natural-sounding responses, and a design that doesn’t foreclose bidirectional voice in the future.
Oh, and Rhonda? No voice. Hard stop. The anonymous surface gets images and small docs, with Turnstile-gated access and short retention. Because if you’re not willing to attach your name to it, you don’t get to talk to it.
The system around the model is what makes this work
A model can read a file. A model can transcribe voice. But without a harness, that’s just capability without coordination. You end up with twelve tabs open, a meeting to align on what the AI actually said, and a spreadsheet to track whether the work got done. That’s not progress—that’s a tax.
Atlas treats attachments and voice as first-class citizens in the conversation. The model gets the context. The harness retains it. Permissions and approvals apply. Audit logs record what happened. And when the work is done, you can measure whether it moved the needle—not just how many tokens it burned.
✦ Why this isn’t just another AI chat feature
Most tools add attachments and voice as checkboxes. Atlas adds them as governed, scoped, and measurable parts of the workflow. That’s the difference between intelligence and operational AI. One is a model. The other is a harness.
What you can do now—and what’s still coming
Today, you can upload files to NyLi, Rhonda, and Atlas Owner conversations. The model will see the contents, and the context will persist across turns. You can also use voice for transcription and text-to-speech, with the flexibility to switch providers later. That’s the shipped reality.
What’s not here yet? Realtime bidirectional voice. Org-wide Knowledge promotion for attachments. And—let’s be honest—plenty of edge cases we’ll find once this hits real workflows. But the design is built to evolve. The harness doesn’t assume the current stack lasts forever. It assumes work will keep changing, and the system needs to adapt without breaking.
The point isn’t to add features. It’s to make the intelligence operational. That means governed access, measurable outcomes, and a system that learns from what actually works. Not just another chat window with a file upload button.
See it on your own data.
Connect your tools and Atlas shows you what matters.
Frequently asked questions
Can I use attachments in org-wide Knowledge, or is it conversation-only?
Attachments are conversation-scoped only. They don’t auto-promote to org-wide Knowledge, because context should be controlled, not viral. If you want to make a file available to the whole org, you’ll need to upload it through the proper Knowledge channels—with permissions, approvals, and audit logs intact.
What happens if I upload a malicious file?
Magic-byte sniffing runs on every upload, and real antivirus scanning happens on the anonymous Rhonda surface. For NyLi and Atlas Owner, the assumption is that you’re uploading trusted files—but the system still checks. Governance isn’t about trusting users; it’s about building a system that doesn’t break when someone makes a mistake.
Why no voice on Rhonda?
Rhonda is the anonymous surface. If you’re not willing to attach your name to a conversation, you don’t get to use voice. It’s that simple. The goal isn’t to maximize usage—it’s to build a system where capability is governed, scoped, and accountable.
Keep reading
Related resources
Why Your AI Agents Keep Hitting the Same Wall—And How to Fix It
BlogWhen Your AI Strategy Is Just a Model with Nowhere to Go
BlogPast the Spreadsheet: Making AI Accountable, Not Just Another File
BlogMeasuring Work Differently in the Age of AI
BlogQuestioning the Work Itself: What Should Even Exist to Automate
BlogNot Everything Should Be Automated — Deciding What Deserves It
Newsletter
The consolidation memo.
Practical insights on AI, operations, and the future of business software. No fluff.