AI recording pendants prove people will pay to stop losing conversations. Most businesses lose the same data on every sales call and never fix it.
Quick answer: AI recording pendants such as Bee, Plaud and Omi cost about $50 to $180 and turn conversations into transcripts, summaries and tasks. Their real lesson for businesses is that structuring conversation data, not recording it, creates the value. Sales calls, support chats and meetings can be captured, structured and routed into the CRM automatically, with consent and human review.
A small, odd product category is proving something important about business: AI recording pendants. Clip one to your shirt and it listens through the day, transcribes what was said, works out who said it and pulls out the action items. The cheapest, Bee, launched at $49.99 plus a $19 monthly subscription, and Amazon bought the company. People pay for these devices out of their own pockets to fix a low-stakes problem: forgetting what was said. Most businesses have the same problem in their sales calls, support chats and meetings, at far higher stakes, and still have not fixed it.
AI pendants are small wearable microphones that send audio to an app, where speech-to-text and a language model turn it into transcripts, summaries and to-do lists. Hardware runs from about $50 to $180, usually with a subscription for heavy use, and the category has already seen two big-tech acquisitions and one high-profile shutdown.
| Device | Price and plan | Status (October 2026) |
|---|---|---|
| Bee (wristband or clip) | $49.99 plus $19 a month | Amazon agreed to buy Bee in July 2025; shown by Amazon at CES in January 2026, run separately from Alexa for now |
| Limitless Pendant | $99 | Acquired by Meta, announced 5 December 2025; no longer sold, existing users supported through 2026 |
| Plaud NotePin S | $179, with 300 transcription minutes a month included | Independent; says it has more than 2 million users and $100 million in annual recurring revenue (June 2026) |
| Omi | $129 offer price (regular $179), no subscription required | Independent; open source, claims 300,000+ users |
| Humane AI Pin | $699 at launch | Shut down on 28 February 2025; assets sold to HP for $116 million |
Two details matter for business readers. First, Plaud's growth came from a professional audience: the company describes its users as professionals in more than 170 countries and has launched a team product. Second, Plaud's figures are company-reported, and the same article notes it did not disclose profitability or split hardware from subscription revenue, so treat the numbers as a direction of travel rather than audited fact.
Because the hardware is a delivery route for something more valuable: a continuous, structured record of what a person hears and says. Meta's statement on Limitless said the team would help it "accelerate our work to build AI-enabled wearables", and Amazon's Alexa leadership told TechCrunch that Bee and Alexa would eventually come together.
The microphone is also moving into products people already wear. EssilorLuxottica reported selling 7 million Ray-Ban and Oakley Meta smart glasses in 2025 alone, more than triple the previous year. The amount of conversation that can be captured is about to grow, for individuals and for companies.
The failures are as instructive as the deals. Humane tried to replace the phone and was switched off about ten months after launch. The pendants that survived do one narrow job well: they remember conversations for you.
It is in what happens after. Recording has been cheap for decades; what is new is the step that turns a messy hour of talk into a summary, a task list and a searchable record sorted by speaker.
Bee makes the point by design: after transcribing a conversation, it discards the audio. Nobody buys it for the audio file. They buy it for the structure: who promised what, by when. TechCrunch notes the trade-off, that without the recording you cannot play a conversation back to check accuracy, which matters a great deal for work use.
That distinction, capture versus structure, is exactly the gap most businesses have. A recorded sales call sitting in a folder is no more useful than a call that was never recorded. A call whose objections, budget, timeline and next step land in the CRM automatically is a different asset.
More than most leaders assume, because a growing share of business conversation happens in unplanned calls that nobody minutes. Microsoft's 2025 analysis of Microsoft 365 signals found that 57% of meetings are ad hoc calls without a calendar invite, and that employees are interrupted by a meeting, email or notification every two minutes during core hours.
| Finding | Figure | Source |
|---|---|---|
| Meetings held as ad hoc calls with no invite | 57% | Microsoft Work Trend Index, June 2025 |
| Interruptions in core hours | One every 2 minutes, about 275 a day | Microsoft Work Trend Index, June 2025 |
| Change in meetings after 8pm | Up 16% year on year | Microsoft Work Trend Index, June 2025 |
| Sales reps' time spent on non-selling tasks | 70% | Salesforce State of Sales, 5,500 respondents, 2024 |
| Organisations where less than half of CRM data is accurate and complete | 76% | Validity, State of CRM Data Management 2025 |
| CRM users who saw revenue loss from poor data quality | 37% | Validity, 2025 |
Put those together and the picture is clear. Conversations are frequent, fragmented and unscheduled. Salespeople already spend most of their week on admin, which per Salesforce includes meeting preparation. And the system meant to hold the record, the CRM, is incomplete at three in four organisations in Validity's survey of 602 CRM users. The objection a prospect raised, the exact phrase a churning customer used, the detail that should have shaped a proposal: almost all of it evaporates when the call ends.
Asking people to take better notes does not fix this. That is precisely the habit the pendant replaces for individuals. The fix is to make capture and structuring automatic.
None of this needs a wearable. It needs the pendant's principle applied to the channels you already run: phone, video calls, WhatsApp, email and support chat. A workable pipeline has six steps:
This is ordinary workflow automation, doing the one job manual process has always done worst. The pendant makers worked out that capture is cheap and structure is where the value sits. The businesses that automate the second part will simply know more about their customers than competitors relying on memory.
Usually yes, if you tell people and have a lawful basis, but the rules differ by country and, in the US, by state. Treat consent as part of the design, not an afterthought.
Federal law requires the consent of at least one party to a conversation. About 11 states, including California, Florida, Illinois, Massachusetts, Pennsylvania and Washington, generally require all parties to consent, according to the Reporters Committee for Freedom of the Press. A business calling customers across states should default to announcing the recording to everyone.
Singapore's PDPC says an audio recording may comprise personal data if an individual can be identified from it, including from their voice, and the PDPA's consent obligation applies to collecting, using and disclosing personal data. The same guidance notes that an employee acting in the course of employment is treated as acting for the employer, so the company carries the obligation. Penalties for breaches of the data protection provisions can reach S$1 million or 10% of annual turnover in Singapore, whichever is higher. For a fuller primer, see our guide to data privacy in the AI era.
Consent gets harder when software joins the meeting. In August 2025 a proposed class action accused Otter.ai of recording conversations without permission; NPR reported that by default the tool asks the meeting host for permission to record but not the other participants. Whatever the outcome, the lesson for buyers is practical: check who your tools ask, what they tell participants and whether recordings are used to train models.
Regulation and risk also shape availability. When Limitless joined Meta, it stopped service in the EU, the UK, Brazil, China, Israel, South Korea and Turkey on the same day. It gave no reason on that page, but businesses with customers in those markets should check a vendor's regional support before building on it.
Good enough to be useful, not good enough to trust blindly. A 2024 study of OpenAI's Whisper speech-to-text model found that roughly 1% of transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio, and 38% of those hallucinations included explicit harms such as invented associations or false authority.
One per cent sounds small until you multiply it by every call in a sales team's month. That is why the review step in the pipeline above is not optional for anything that drives a decision, and why keeping audio for a short, defined period is worth the storage cost.
Machine-reported numbers need reconciling against reality in every channel, not just transcripts. On an Indian D2C food and handicrafts store we manage, Meta's reported purchases ran roughly 30% above the store's actual orders. We now check platform figures against the system of record before acting on them.
The closest thing to a pendant we have built for a business is a CRM that captures conversations and outcomes without relying on anyone's memory. For a Kolkata real-estate developer, we built a CRM where leads from Meta ads, WhatsApp and walk-ins land in one pipeline. WhatsApp messages fire automatically on site-visit confirmations and sale closures, booking forms and sale agreements generate as PDFs, and real sale outcomes are sent back to Meta's conversion API. That last step matters most: ad optimisation learns from actual sales, not just form fills. It is the business version of extraction. A conversation that ends in a booking becomes structured data that improves the next campaign.
The second lesson is less glamorous. One hosted chatbot we run went dark, and the cause was not the model. It was a CDN security rule and an expired credential. Conversation-capture systems fail the same way: a changed phone-system setting, an expired API token or a full storage quota silently stops the pipeline, and nobody notices until a manager asks why the CRM has no call notes for a week. We now monitor the plumbing (DNS and CDN rules, tokens, database quotas) as closely as the AI output.
Start with the conversations that already carry revenue: discovery calls and support contacts. Record them with clear notice, extract a fixed set of fields into your CRM, and have a person review anything that changes a price or a promise. Do not buy wearables for staff; the phone system, meeting platform and WhatsApp already capture most business conversation, with clearer consent. Judge any tool on three things: how it handles consent for every participant, whether it writes structured fields to your system of record, and whether you can check its output against the audio. If you want this built around your own CRM, our AI CRM team can scope it.
Mostly no, by design: pendants capture in-person audio around the wearer. For businesses it is usually simpler to capture calls and video meetings through the phone system or meeting platform, which already know who is on the call and can play a consent message to every participant.
Treat it as a policy decision, not a personal choice. Under Singapore's PDPA, an employee acting in the course of employment is treated as acting for the employer, so the company carries the consent obligation. Set a written policy covering which devices are allowed, how participants are told, where recordings are stored and when they are deleted, and prefer company-managed tools.
Start with the fields that change what someone does next: next step and owner, budget, timeline, decision-maker, objections raised, competitors mentioned and, for support, the stated reason for contact or cancellation. Keep the list short and fixed so the data is comparable across calls, then add fields once the first set is reliably accurate and used in reporting.
Keep audio for a short, defined period. Transcription models still hallucinate occasionally, and without the recording nobody can check a disputed summary. Bee discards audio after transcription, which TechCrunch notes makes it impractical for many work uses. A retention period of a few weeks, stated in your privacy notice, balances verification against data-protection obligations.
They overlap. Note-takers summarise individual conversations for the person who was there. Conversation intelligence tools aggregate many calls to show patterns, such as which objections lose deals or which phrases precede churn, and write structured data into the CRM. Most businesses get more value from the second, because it turns individual memory into shared, measurable knowledge.