Acedly AI on Zoom: Real-Time Interview Copilot for Zoom Calls (2026)
How Acedly AI works on Zoom — hidden from screen sharing, sub-200 ms end-to-end latency, grounded in your résumé and the job description. Verified stealth across Zoom share modes, what to test before your next Zoom interview.
Devon Park
Head of Research, Acedly
What a Zoom interview assistant actually is
A Zoom interview assistant is a desktop application — almost never a browser tab — that sits beside your Zoom client during a live interview. It does three things in order: it captures the interviewer's audio off the system loopback, it transcribes and reasons about the question, and it renders an answer on a surface that is excluded from Zoom's screen-sharing pipeline. The whole loop has to fit inside the silence between the end of a question and the moment a candidate would normally start talking — about 200 milliseconds.
The reason this category is shaped the way it is comes back to Zoom itself. Zoom is the dominant Western interview platform in 2026, and the way Zoom interviews are run — share-screen-as-default for technical rounds, gallery view for panel rounds, breakout rooms for system design — sets the constraints every assistant has to satisfy. A tool that works on a generic web call but breaks the moment a Zoom interviewer asks "can you share your screen?" is not a Zoom interview assistant. It is a demo.
The defining number for this category is end-to-end latency, and the right target is sub-200 ms — measured from the last syllable of the interviewer's question to the first character of the answer appearing on your screen. Acedly's median on consumer hardware is roughly 98 ms. Anything past 250 ms and you visibly fall behind the rhythm of the call.
Why Zoom is still the dominant Western interview platform in 2026
Zoom's share of professional interview minutes hasn't moved much since the 2022 peak. Recruiters' calendar invites still default to Zoom for the first round, the technical round, and frequently the loop debrief. Microsoft Teams owns the enterprise loop in finance and consulting; Google Meet owns startups and product orgs; but for the long tail of mid-market hiring — the SaaS company, the Series B, the cross-border recruiter pulling candidates from three continents — Zoom remains the path of least resistance.
A few patterns are specific enough to Zoom that they shape what an assistant has to handle:
- Share screen is the default for technical rounds. Recruiters expect candidates to share either a coding sandbox tab or, increasingly, the entire desktop. The latter is the riskier case for any interview assistant — a tool that hides one window but reveals when "Entire Screen" is selected is half-built.
- Gallery view vs. speaker view changes how often the candidate's webcam is centred. In a panel round on Gallery view, the recruiter often glances at the candidate's eye gaze. An assistant that lives on the same screen as the webcam pushes your gaze off-camera; an assistant on a second monitor keeps you looking forward.
- Breakout rooms get used for system-design rounds. A senior engineer hops into a breakout room with the candidate to whiteboard. Audio routing through breakout rooms is identical to the main room from the OS's perspective, but some browser-tab tools fail because they treat breakout rooms as new sessions.
- Zoom for Government and Zoom for Education builds exist. They have slightly different update cadences and occasionally different screen-share UIs. An assistant that only verifies on the consumer Zoom build will eventually surprise you on a federal contractor interview.
The takeaway: Zoom-specific does not mean "works on Zoom." It means "verified against the actual Zoom surfaces a candidate runs into."
How an AI interview assistant works on Zoom specifically
The end-to-end pipeline of a real-time copilot is the same on every meeting platform — capture audio, transcribe, ground, infer, render — but each link bends to the platform underneath it. On Zoom, three pieces are platform-specific: the audio path, the rendering exclusion, and the integration with Zoom's own AI features.
Audio capture: macOS vs. Windows
On macOS, the assistant uses Core Audio's loopback interface or, on macOS 14+, the ScreenCaptureKit audio APIs to subscribe to the system audio output. This means it hears whatever Zoom is playing into your speakers — the interviewer's voice — without needing kernel extensions or virtual audio devices. The Zoom desktop app routes interviewer audio through the standard Apple audio path, so a properly built assistant captures it the same way Apple's own screen-recording tools do.
On Windows, the equivalent is WASAPI's loopback mode. The assistant opens an IMMDevice in render-loopback mode and reads the same buffer that Zoom is writing to your speakers. The Zoom Windows client cooperates with this; it does not insulate its audio from system loopback the way some DRM-protected applications do.
The practical implication: a native desktop assistant captures the Zoom audio cleanly. A browser-tab tool cannot, because browsers don't have access to system loopback without explicit user permission to "share a tab with audio" — and you cannot ask the recruiter to enable that. This is the single biggest reason browser-tab interview tools are a dead end on Zoom.
Why a browser-tab tool fails on Zoom
Browser tabs see the audio of their own page, not the system. A browser-tab interview tool either has to ask you to run Zoom inside the browser (which loses features like gallery view and reactions), or it has to rely on your microphone picking up the interviewer through the speakers — which is noisy, slow, and breaks if you use headphones. Native desktop tools sidestep both problems. The whole reason an assistant should be a native app is that this audio path is the floor, not the ceiling.
Rendering: how it stays out of Zoom's screen-share pipeline
Zoom's screen-sharing system, on both desktop platforms, is built on the OS's window-capture APIs. On macOS that means CGWindowList and ScreenCaptureKit; on Windows it means the Desktop Duplication API and the Graphics Capture API. The assistant opts out of these by setting NSWindowSharingNone (macOS) and SetWindowDisplayAffinity(WDA_EXCLUDEFROMCAPTURE) (Windows). When Zoom asks the OS for a window list, the assistant doesn't appear in it. When Zoom captures the entire desktop, the OS skips the excluded window in the capture buffer.
This is the same mechanism that hides DRM video from screenshots. It is an OS-level guarantee, not an application-level trick. A correctly built assistant cannot accidentally show up in a Zoom share — but a tool that doesn't use these flags can never be hidden, no matter what the marketing claims.
Integration with Zoom AI Companion
Zoom AI Companion (formerly Zoom IQ) is Zoom's own auto-summary feature. It transcribes the meeting audio and produces a summary the host can review afterwards. The thing to understand is what AI Companion can and can't see. It transcribes the meeting audio — the same audio Zoom is already routing — which means it sees what the interviewer said and what you said into your microphone. It does not see your local UI. It does not see your second monitor. It does not see Acedly. AI Companion is a transcript service, not a desktop monitor.
The Zoom-specific stealth checklist
Stealth on Zoom is a binary, not a spectrum. The interviewer either sees the assistant or they don't, and there are six specific Zoom surfaces where the test happens. A serious Zoom interview assistant passes all six on the build of Zoom your interviewer is actually using.
- Excluded from Zoom's "Share Screen → Window" picker. When the candidate clicks "Share Screen," Zoom shows a grid of available application windows. The assistant must not appear in that grid. This is the most common failure mode for half-built tools — they hide during the share itself but appear in the picker, which gives the candidate two seconds to see the wrong row before they realise the mistake.
- Hidden when sharing the entire desktop. "Entire Screen" share on Zoom captures everything the OS shows on that monitor. The assistant must be excluded at the OS-capture level so that even an Entire-Screen share doesn't reveal it. This is where
NSWindowSharingNoneandWDA_EXCLUDEFROMCAPTUREactually pay off. - Doesn't appear in Zoom's "App Window" share list. Distinct from the picker — this is the in-share toolbar that lets the candidate switch which window is being shared mid-interview. An assistant that hides on the initial share but reappears in the switcher is one click away from being seen.
- Doesn't appear in Zoom recordings. Local Zoom recordings and cloud recordings both pull from the same capture buffer the host's view sees. Capture-excluded windows do not appear in either. The implication for the candidate is that an interview replay — including one that AI Companion auto-generates — never shows the assistant.
- Hidden in Zoom's in-meeting reactions and cursor highlights. Some Zoom features (the spotlight cursor, the "draw on shared screen" annotation overlay) render on top of the shared region. A correctly hidden assistant is below all of that — it is excluded at the source, not just visually layered.
- Compatible with Zoom for Government / Zoom for Education builds. These builds have separate FedRAMP-compliant or FERPA-compliant codepaths. The window-capture APIs they use are the same, so a properly built assistant works there too — but verification has to happen on the actual builds, not just on consumer Zoom.
The right way to verify any of this is not to read the marketing page. It is to start a Zoom call with a friend, share your screen each of the three ways (one app window, the desktop, the secondary monitor), and have them tell you what they can see. Anything visible to the friend is visible to the recruiter.
Comparison: how Zoom interview tools actually differ
Most products that show up under "Zoom AI" search results are one of four things, and only one of them is a real Zoom interview assistant. Here's the comparison we use internally.
| Feature | Acedly | Browser-tab AI | Screen-recorder copilots | Generic AI chat |
|---|---|---|---|---|
| Median end-to-end latency on Zoom | ~98 ms | ~600–900 ms | Post-call only | ~2–4 seconds |
| Stealth on Zoom screen share | Excluded at OS capture level | Browser tab only (fails on Entire Screen) | Visible in recording playback | No (just another window) |
| Reads coding sandbox on screen | Coderpad, HackerRank, LeetCode, etc. | Limited to same browser | Yes, but post-call | Manual paste only |
| Grounded in your résumé and JD | Yes, by default | Sometimes | N/A (post-call) | Only if you paste |
| Works in Zoom breakout rooms | Yes (audio path is identical) | Often resets session | Yes (records the room) | N/A |
| Visible to Zoom AI Companion | No (local UI only) | No | No | No |
The honest read of this table is that browser-tab tools and generic AI chat are not in the same category as a real Zoom interview assistant. They share keywords but not capabilities. Screen-recorder copilots are useful for self-review after the fact, not in real time. Acedly's category is the narrow one where all four columns — latency, stealth, grounding, and screen reading — have to be solved simultaneously.
What to do in the 10 minutes before your Zoom interview
Five minutes of setup beats a great copilot during an awkward call. The candidates who get the most out of any Zoom interview assistant are the ones who treat the launch as a checklist, not as something to figure out under pressure.
- Test screen share with a friend on the actual Zoom build you'll use. Twenty minutes before the call, open a Zoom test meeting, share your screen three ways (One Window, Entire Screen, Second Monitor), and confirm the friend sees nothing they shouldn't. This is the single highest-value thing you can do.
- Practice your hotkey twice in silence. The most common in-call failure is fumbling the hotkey because you've never used it under stress. Two clean repetitions burn the muscle memory.
- Verify your microphone input level. Zoom's auto gain control will sometimes amplify a desk-tap to "speech" volume. Open Zoom audio settings, speak a sentence, and check the input meter is in the green band.
- Choose Gallery view if it's a panel round, Speaker view if it's a 1:1. Gallery view is the default for panels because the recruiter expects to see all interviewers. Speaker view keeps the question-asker centred — better for technical rounds where you watch the interviewer's face.
- Move Acedly to your second monitor. This is the biggest tell-control move. If Acedly is on the same monitor as the Zoom window, your eye line drifts. On a second monitor, your gaze stays on the camera and the assistant stays peripheral.
- Close anything you don't want shared by accident. Even though Acedly is excluded from screen share, your Slack window and your draft email are not. Standard hygiene.
- Have your résumé open in plain text. Not for the assistant — Acedly already has it. For you, in case the recruiter asks something résumé-specific and you want to glance.
Privacy: what Zoom's recording sees vs. what your interviewer sees
The two layers of "what gets seen" on Zoom are easy to confuse and the difference matters for any honest evaluation of an interview assistant.
The live screen-share layer is what the interviewer sees in real time. Everything in this article — the six stealth surfaces, the share picker, the entire-desktop share — is about this layer. A properly built assistant is invisible here at the OS-capture level. The interviewer cannot see Acedly during the call.
The recording layer is what gets saved if the host has cloud or local recording enabled. Zoom recordings — including the auto-generated AI Companion summary — are produced from the same capture buffer the host sees. Capture-excluded windows do not appear in the recording either, because the OS never put them in the buffer to begin with.
Zoom AI Companion is worth being explicit about, because candidates often worry. AI Companion transcribes the meeting audio (interviewer + candidate microphone) and produces a summary. It does not have access to the candidate's local screen, the candidate's second monitor, or any process running on the candidate's machine. It is a transcript service. The relevant property of Acedly is that everything it generates lives in the candidate's local UI and never enters the meeting audio — Acedly does not speak into your microphone.
The honest framing: a recruiter who reviews a Zoom recording afterwards sees exactly what they saw live. They do not get a privileged second pass that reveals the assistant. The capture exclusion is the same mechanism in both cases.