For as long as small businesses have owned a recorder, the constraint was never the recording. It was that somebody had to sit down afterwards and listen to it.
Microsoft AI put a new price on that job this week. On September 3 it released MAI-Transcribe-2, which turns an hour of recorded audio into text for ten cents.
The short version: AI transcription for small business is now cheap enough that cost is not the thing stopping you, and it has not been for some time. Microsoft’s ten cent hour undercuts the cheapest managed competitor by five cents, not by an order of magnitude. The useful consequence is not a saving on your bill, because you almost certainly do not buy transcription by the hour. It is that capturing what gets said in your business now costs less than the meeting you would hold to decide whether to capture it. That moves the real question somewhere most owners have not looked.
What Microsoft actually shipped
The claims are specific, which makes them checkable. Microsoft says MAI-Transcribe-2 ranks first on FLEURS, a multilingual speech benchmark, across 60 languages with an average word error rate of 5.2 percent, and second on Artificial Analysis’s independent word error rate leaderboard. On speed it claims 10 times faster processing than OpenAI’s GPT-Transcribe, 7 times faster than ElevenLabs Scribe v2, and 5 times faster than Gemini 3.5 Transcribe. The model is available through Microsoft Foundry, the MAI Playground, and OpenRouter, and it handles speaker separation, word-level timestamps, keyword biasing, and automatic language identification.
One asterisk belongs on the accuracy headline. That 5.2 percent error rate is higher than the 3.7 percent reported for the previous version, because the new figure averages across 60 languages rather than a narrower set. Broader coverage, slightly higher average. That is a sensible trade rather than a regression, and Microsoft is careful not to claim it beats every rival on raw accuracy. But “most accurate in the world” and “a worse error number than last time” are both true sentences about the same model, and it is worth knowing which one you are being sold.
How much does AI transcription for small business actually cost?
Against Microsoft’s own previous model, the cut is dramatic. VentureBeat notes the ten cent price is 72 percent below the $0.36 an hour Microsoft charged when it shipped the first model in this line five months ago. Against the wider market, the cut is a nudge.
Here are the list prices per hour of audio as they stand: AssemblyAI’s value tier at $0.15, OpenAI’s mini transcription model at $0.18, AssemblyAI’s flagship async tier at $0.21, ElevenLabs Scribe v2 at $0.22, OpenAI’s larger model and Whisper at $0.36, and Deepgram’s Nova-3 at $0.46. The floor moved from fifteen cents to ten. That matters enormously to a company transcribing a hundred thousand hours a month. It is invisible to you.
Run the arithmetic on a real business. A service company taking forty customer calls a week, averaging six minutes each, produces roughly seventeen hours of audio a month. At ten cents an hour that is $1.73. At the most expensive price on the list above, it is about $8. The spread between the cheapest and the dearest option in the entire market, applied to a real month of a real small business, is six dollars.
So the interesting number in this story is not the price. It is the ratio between that price and what you are charged.
Why the tool still costs more than the transcription
If the raw input costs a couple of dollars a month and your meeting notes subscription costs considerably more than that, the difference is not transcription. You are paying for the recording app, the storage, the search, the summaries, the sharing permissions, the integrations with your calendar and your CRM, and somebody to answer the phone when it breaks. Those things have real value. They are simply not the thing named on the tin.
This is worth internalising because it should change how you compare tools. Vendors in this category compete on accuracy percentages, and accuracy is now the least differentiated part of the product. Every serious option is in the same narrow band, and the last two percentage points will not decide whether the tool earns its keep. What decides that is what happens after the transcript exists: whether it is searchable six months later, whether the summary is good enough that somebody reads it, whether it lands in the system where the work actually happens. We have made a version of this argument before about the collapsing price of reading a document, and the shape is identical. When an input becomes nearly free, the value moves to what surrounds it.
The same logic applies to the ten cent price itself. Microsoft describes it as a limited-time offer running to the end of the year, and has not published what the standard rate will be afterwards. That is a familiar pattern in AI pricing this year, and we have argued that the date beside the price now matters more than the price. If a vendor tells you they are switching to this model to hold your subscription cost down, the fair question is what happens in January.
What changes when listening costs nothing
The bottleneck does not disappear. It moves, and it is worth knowing where it lands.
It lands first on attention. A transcript nobody opens is a more expensive form of forgetting, because it carries the illusion of a record. The businesses that get value here are the ones that decided in advance what they were looking for: the three recurring objections in sales calls, the promise somebody made about a delivery date, the detail from the site visit that never made it onto the quote.
It lands second on where the words live. An archive of every customer conversation is a genuine asset and a genuine liability in the same folder. Knowing who can search it, how long it is kept, and what happens to it if you change tools is now a real decision rather than a theoretical one.
And it lands, usefully, on your people. The honest case for transcription has never been that it replaces the person taking notes. It is that nobody can listen properly to a customer while writing down what the customer is saying. Handing the recording job to software gives that attention back to the person in the room. That is a small change that customers can feel, and it is the reason this particular capability tends to earn its place faster than flashier ones. If you are weighing where to start, we sorted the options by the job you actually need done.
The thing to check before you record everything
Cheap capture makes blanket recording tempting, and this is the one place where moving fast can cost you properly. Consent law is not uniform. Most states allow recording when one party to the conversation consents, which includes you. Roughly a dozen require every party to agree, and the exact list is genuinely contested at the margins because several states treat phone calls and in-person conversations differently. California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington appear on every version of that list.
None of this is an argument against recording. The remedy is small and well established: say at the top of the call that it is being recorded, and keep the practice consistent rather than selective. The mistake worth avoiding is switching on capture everywhere because it now costs $1.73, and discovering the rule afterwards. Check your own state, and check the state your customer is sitting in.
What should a small business do about this?
Four things, none of them urgent. Start by finding out what you already pay for, because transcription is bundled into most phone systems, video call platforms, and note-taking apps, and a surprising number of owners buy a second tool for something they own already. Second, pick one conversation type worth capturing rather than capturing everything; sales calls and site visits are usually where the missing information lives. Third, decide the question before you decide the tool, because “what do I want to know that I currently forget” is a better selection criterion than any benchmark. Fourth, if a vendor is pitching a price built on this week’s promotional rate, get the post-January number in writing.
The pattern underneath all of this is one we keep running into. A capability gets cheap, the headline treats the price as the news, and the actual change is that a constraint you had stopped noticing quietly lifted. We wrote a fuller guide to reading those pricing pages and to what is genuinely worth paying for, and the median small business spending on AI is a great deal lower than most people assume.
Frequently Asked Questions
How much does AI transcription cost a small business?
Far less than most owners expect. Raw API rates now run from about $0.10 to $0.46 per hour of audio depending on the provider, which means a service business generating seventeen hours of recorded calls a month is looking at somewhere between two and eight dollars of actual transcription cost. In practice you will rarely pay that directly, because transcription arrives bundled inside a phone system, a video platform, or a note-taking subscription, and those are priced per user per month. The gap between the two numbers is what you are really buying, and it is mostly storage, search, summaries, and integrations rather than the transcription itself.
Is MAI-Transcribe-2 worth switching to?
Probably not as a decision you make directly, unless you or a contractor build on transcription APIs yourself. For everyone else the model reaches you only if the vendor behind your existing tool chooses to adopt it, and the saving lands on their cost line rather than your invoice. If you do buy at the API level, the speed advantage is the more interesting claim of the two, because faster processing is what makes transcription feel immediate rather than something you collect later. Treat the ten cent rate as promotional and confirm the standard price before building a budget on it.
Do I need permission to record and transcribe customer calls?
It depends on where you and the customer are. Most US states operate on one-party consent, meaning your participation in the conversation is sufficient, but roughly a dozen require all parties to agree, and a few treat phone calls differently from in-person conversations. California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington appear on every version of the all-party list. The practical answer is to announce recording at the start of every call and apply it consistently, which satisfies the stricter rule everywhere and costs you one sentence. Confirm the specifics for your own state before you turn on blanket recording.
Will AI transcription replace my admin staff?
That is not what this technology does, and the framing gets the value backwards. Transcription removes the mechanical act of writing down what was said, which was never the part of an administrator’s job that required judgment. What it hands back is attention: nobody listens well to a customer while simultaneously taking notes, and the person freed from the keyboard is the one who catches the detail that turns into repeat business. The businesses seeing real returns here use the recovered time on follow-up and on the work that was quietly falling through the cracks, not on reducing the number of people.
The question we keep circling back to is this one. If every conversation in your business had been transcribed and searchable for the past year, what would you go looking for first? We would genuinely like to know, because the answer is usually the tell for where the real gap is.
