Resemble AI launches native real-time deepfake detection in Microsoft Teams

Why Resemble AI is targeting video calls

Resemble AI argues that the common assumption that people would notice a fake on a video call is getting harder to rely on. The company points to several cases. In 2024, an Arup employee authorized 15 transfers worth $25.6 million after joining a call in which his CFO and colleagues were all deepfakes. In 2025, a Singapore finance director transferred more than $499,000 after a deepfake executive call. In February 2026, a cloned voice helped move roughly €95 million out of Fideuram, part of Intesa Sanpaolo.

The post also cites Gartner's September 2026 survey of 297 CISOs. According to the figures Resemble AI quotes, 41% had dealt with a deepfake used in a social-engineering attack on an employee phone call in the previous year, and 36% had seen one on a video call.

Resemble AI says organizations need to be ready to detect AI and fraud across four use cases: CEO fraud, candidate fraud, payment fraud, and IT Help Desk impersonation.

How Resemble Detect works in Teams

Resemble Detect for Microsoft Teams combines identity matching and enrollment, audio and video deepfake detection, and fraud alerts inside Teams. It includes four of the company's models: Detect for spotting AI in video or audio, Identity for identity matching, Signal for fraud detection, and Intelligence for plain-language explanations of the other results.

Once an administrator enables it, Teams meetings are covered automatically. Before a call, users can optionally enroll a face and voice sample tied to their Microsoft 365 account. The company says enrollment is a one-time process that takes a few minutes. During the call, the system analyzes audio and video streams continuously and checks whether the person on camera and microphone matches the enrolled identity.

• If something is flagged: the meeting host receives an alert naming the participant and the stream that triggered it. The host can ask the person to verify, remove them, or involve the security team.
• After the call: flagged events are summarized in the Teams chat. A report with timestamps, frame-level visual analysis and audio-segment scoring can be generated for closer review.

There is no separate bot in the meeting, although organizations can add a visible one if they prefer. Participants do not need to download anything. Deployment goes through the Teams admin center.

The identity check also compares each participant with the Microsoft 365 account they signed in with. If someone joins on a colleague's account, it flags an account mismatch. Unenrolled people appear as "identity not found", and the organizer can add them once they are confirmed as authentic.

Accuracy claims, add-ons and data handling

Resemble AI says it publishes its detection benchmarks openly. On the Podonos leaderboard in August 2026, its audio model ranked #1 with 99.47% accuracy across 4,524 test files, with a 0.4% miss rate and a 0.7% false-alarm rate, according to the company. It reports 98.2% accuracy for video deepfake detection, based on an internal benchmark of 1,000 files. The company also lists a 0.12s real-time factor, support for 54 languages, and testing against more than 250 generative models.

Signal and Intelligence are premium add-ons. Signal listens for suspicious intent, such as an urgent request to pay a new bank account, and labels what it hears as clean, suspicious or fraud as the call goes on. Intelligence explains in plain language why a part of the call was flagged. The company says Signal works without transcription or recording of conversation content.

On data, Resemble AI says media is kept only for participants flagged by the system, so security teams can review the event. It states that customer data is not used to train its models, that the platform is SOC 2 Type II certified, and that it is designed to support GDPR and HIPAA requirements.

What it means for organizations

Resemble AI positions the product for organizations where video calls drive decisions about money, access, sensitive information or executive approval. It names financial services, legal, healthcare, critical infrastructure and government, while saying the underlying risk is broader. For CISOs, the company describes it as an extra layer on top of existing identity controls, aimed at what happens after someone passes the initial checks to join a meeting.

Organizations that run sensitive meetings on Teams can request a walkthrough from the company. Pricing and availability details were not given in the source.