AI Transcription Just Got 72% Cheaper. Here Is What That Unlocks.

Estimated read time: 12 minutes

A few years ago, having an hour of audio turned into accurate text cost roughly what a sandwich costs. Human transcription services charged by the minute and were worth it, because the alternative was doing it yourself at four times real time. Then machine transcription got good, and the price started falling. In early September 2026, Microsoft’s MAI-Transcribe-2 was reported at around ten cents per hour of audio, a cut of about seventy-two percent against what was already cheap.

Ten cents an hour is not a discount. It is a different category of thing. At that price, transcription stops being a service you buy for important recordings and becomes a background process you run on everything. That shift is worth thinking about carefully, because most small businesses will respond to it by recording far more and reading far less, which is a way to spend money and gain nothing.

TL;DR

Transcription cost has fallen far enough that the constraint on recording your business conversations is no longer money. It is attention, consent, and storage risk. The businesses that get value from this will not be the ones that record the most. They will be the ones that pick two or three specific questions they want answered from their own conversations, record only what serves those questions, and build a habit of actually reading the output.

Start with sales calls or support calls, not everything. Get consent properly. Set a deletion schedule before you set a recording schedule. And treat the transcript as raw material for a decision, never as the deliverable.

What ten cents an hour actually means

Put a number on your own operation. Say you take eight customer calls a week averaging twenty minutes. That is roughly two and a half hours of audio, about eleven hours a month, call it a hundred and thirty hours a year.

At ten cents an hour, transcribing every customer conversation you have in a year costs about thirteen dollars.

Widen it. Add every internal meeting, every supplier call, every interview you run when hiring. Triple the number. Still under fifty dollars a year. Compare that to what a business typically spends on a single month of any SaaS subscription and the conclusion is obvious: the raw conversion of speech to text has effectively stopped being a budget line.

This is the part worth internalizing. When a cost drops by an order of magnitude, the useful question is never “how much will I save.” Nobody was spending enough on transcription for the savings to matter. The useful question is “what was not worth doing at the old price that is worth doing now.” That is where the value is, and it requires different thinking than cost-cutting does.

Five things that become affordable

1. A searchable archive of what customers actually say. Not your summary of what they said. Their words. When you can search every conversation for a phrase, you find out that eleven people asked about the same thing in different language, and that your website answers a question nobody asks while ignoring the one everybody does. This is the single highest-value use and almost nobody does it, because until recently it required either a transcription budget or a call-recording platform priced for enterprises.

2. Sales call review that is not vibes. Most small business owners do not know why they lose deals. They have an impression. With transcripts you can compare the calls that closed against the calls that did not and find the actual divergence point, which is frequently not where anyone guessed. This is what expensive revenue intelligence platforms have sold to large sales teams for years, and the underlying data is now nearly free to produce.

3. Content that is not invented. If you publish anything, the hardest part is knowing what to say. A year of customer conversations is a content calendar that writes itself, in language your customers already use. A transcript of you explaining something well to one person is a better first draft than a blank page, and pairs naturally with the workflow in our guide to AI writing tools for small business.

4. Onboarding material generated from real work. Instead of writing procedures from memory, record yourself doing the task while narrating, transcribe it, and edit. The result is more accurate than a written-from-scratch document because it captures the exceptions you would have forgotten. This is the fastest route out of the situation where critical knowledge lives in one person’s head.

5. Accessibility and reach on anything you publish. Captions on video, transcripts under podcast episodes, text versions of webinars. This was always the right thing to do and was often skipped on cost. At current prices, skipping it is a choice about effort rather than money. Search engines index the text, which is a side benefit rather than the reason.

The trap: transcripts nobody reads

Here is the failure mode, and it is close to universal.

A business gets excited, turns on recording for everything, and within two months has four hundred transcripts sitting in a folder. Nobody has read any of them. The folder is now simultaneously a source of guilt, a storage cost, and a data protection liability. The project quietly dies and the recording stays on, which is the worst of all possible states.

The reason this happens is that cheap transcription removes the friction that used to force a decision. When each hour cost real money, you asked whether this recording was worth transcribing, and that question did useful work. Removing the cost removes the question, and the discipline goes with it.

So put the question back. Before you record anything on an ongoing basis, write down the specific thing you want to learn from it. “Why do prospects go quiet after the quote.” “Which feature confuses new customers most.” “What our top three objections actually are.” If you cannot name the question, you are collecting, not researching, and collecting has costs even when the processing is free.

A good rule: start by recording one category of conversation for one month, with one question attached, and block an hour on the calendar at the end of that month to read. If you do not use the hour, stop recording. The habit is the whole asset.

Transcription is the cheap part now

Worth being clear about what is and is not at ten cents an hour. That price is for raw speech-to-text: an API call that returns a wall of words.

Almost everything that makes a transcript useful sits above that layer and is priced differently. Speaker identification, summaries, topic extraction, sentiment, searching across a whole archive, and pulling structured data out of unstructured conversation all cost more, and if you are doing that work with a language model, the processing can cost several times the transcription itself. The consumer meeting tools most small businesses actually use bundle all of this and charge per user per month, which is a perfectly rational way to buy it and has not gotten dramatically cheaper just because the underlying transcription did.

The practical implication is that for most small businesses, the headline price drop will not show up as a smaller invoice. It will show up as products including transcription in tiers where it used to be an upsell, and as generous free allowances. That is still real value. Just do not expect your meeting tool’s bill to fall by seventy percent, because transcription was never the expensive part of what it sells you.

If you do want the raw-cost economics, that means handling the pipeline yourself: recording, calling an API, storing output, and building whatever search sits on top. That is a reasonable weekend project for a technical owner and an unreasonable one for everybody else. Most people should buy a tool and use the falling costs as leverage at renewal.

Consent, recording law, and the part people skip

This section is the one that gets skimmed and it is the one that can actually cost you money.

Recording a conversation is regulated, and the rules vary by jurisdiction. In the United States, some states require only one party to consent, while others require every party to consent, and the applicable law when a call crosses state lines is not always obvious. In other countries the framework may be stricter still, and if you handle data belonging to people in regions with comprehensive privacy law, recording them creates obligations around purpose, retention, and access that go well beyond asking permission.

A few things that are true almost everywhere and are worth doing regardless of what the minimum standard is:

Say it out loud at the start. “I record these calls so I can focus on the conversation instead of taking notes. Is that alright with you?” It takes four seconds, it is on the recording, and in practice almost nobody objects. The people who do object tell you something useful.

Have an answer for “what do you do with it.” Who can see it, how long you keep it, and whether it goes to any third party. If your transcription runs through an outside service, that service is a third party, and some vendors use submitted data differently than others.

Do not record what you do not need. Payment details, health information, and anything a customer shares in confidence should not be sitting in a text file on a laptop. If a call turns in that direction, pause the recording.

This is not legal advice and the specifics genuinely depend on where you and your customers are. If you are going to record systematically, a short conversation with a lawyer in your jurisdiction is cheap relative to getting it wrong.

Storage is a liability, not an asset

An archive of customer conversations feels like an asset. In a breach, it is the worst thing you own.

Text is compact and searchable, which is exactly what makes it valuable to you and to anyone who gets unauthorized access. A folder of transcripts is a plain-language record of what your customers told you, often including things they would not want repeated. The reputational damage from that leaking is not proportional to the file size.

Set the retention policy first, before you record anything:

Decide how long. Ninety days covers most analytical uses. If you cannot articulate why you need a two-year-old sales call, delete it at ninety days.

Automate the deletion. A policy that depends on someone remembering is not a policy. Most tools have a retention setting. Use it.

Extract before you delete. The insight is the point, not the transcript. Pull the objection patterns, the recurring questions, and the good phrasing into a document you keep, and let the source material expire.

Know where it lives. If transcripts sit in a vendor’s cloud, in a shared drive, and in someone’s downloads folder, you have three copies and one deletion process. This is the same dependency question we worked through in our vendor continuity guide, with sharper consequences.

Where machine transcription still fails

Accuracy on clear audio with a single speaker is now good enough that most people stop checking. That is fine until it is not, and the failure cases are predictable.

Crosstalk. Three people on a speakerphone talking over each other remains hard. Speaker labels drift and sentences get attributed to the wrong person, which matters enormously if you are using a transcript to settle what somebody agreed to.

Industry vocabulary and proper nouns. Product names, supplier names, and technical terms specific to your trade get mangled with total confidence. Many tools accept a custom vocabulary list. Adding twenty terms takes ten minutes and fixes most of this.

Accents and non-native speech. Performance is not uniform across accents, and the gap is real. If your customer base is linguistically diverse, test before you rely on it, and be aware that a lower-accuracy transcript for some customers is a quiet form of unequal service.

Numbers. Prices, dates, quantities, and order numbers are exactly what you most want to be right and exactly where a small error is least visible. Never use a transcript as the record of a financial term without checking the audio.

The rule of thumb: transcripts are excellent for finding things and unreliable as evidence. Search them freely. Verify anything you would act on.

A ninety-minute starting plan

Minutes 0 to 15. Pick one conversation type and one question. Sales calls and “why do we lose” is the highest-yield combination for most businesses. Write the question somewhere you will see it.

Minutes 15 to 30. Check what you already pay for. Your video conferencing tool, your phone system, and your CRM may already include transcription you have never switched on. Buying a new tool before checking this is how small businesses end up with duplicate subscriptions.

Minutes 30 to 45. Set retention and permissions before the first recording. Ninety days, automated, and a clear answer to who can access it.

Minutes 45 to 60. Write your consent line and practice saying it until it sounds natural rather than legal. Add your twenty custom vocabulary terms while you are in the settings.

Minutes 60 to 75. Record and transcribe one real call. Read the whole thing. You will learn something about how you talk that no amount of theorizing would have surfaced, and you will find out whether the accuracy is good enough for your audio setup.

Minutes 75 to 90. Put a recurring appointment on your calendar for four weeks out, titled with your question. That appointment is the entire project. Everything before it is setup.

The price of turning speech into text has collapsed and will probably keep falling. What has not changed, and will not, is that information only becomes valuable at the point somebody reads it and decides to do something differently. That part still costs an hour of your attention a month, and that is the only part that was ever scarce.

FAQ

Is ten cents per audio hour what I will actually pay?

Only if you call a transcription API directly and handle everything else yourself. Consumer meeting tools bundle summaries, speaker labels, search, and storage, and charge per user per month. Expect the falling raw cost to show up as better free tiers and fewer upsells rather than as a smaller bill.

Do I need consent to record a business call?

The legal minimum varies by jurisdiction, and in some places every party must agree. Announcing it at the start of the call is the practical standard, takes four seconds, and almost never causes a problem. Check the rules that apply where you and your customers are.

How accurate is AI transcription now?

Very good on clear audio with one or two speakers, and meaningfully worse with crosstalk, heavy background noise, strong accents, and specialist vocabulary. Numbers and proper nouns are the most common errors. Treat transcripts as a search index, not as a record of what was agreed.

Should I record internal meetings too?

Only if you will read them. Recording internal meetings also changes how people speak in them, which is sometimes a benefit and often a cost. Start with customer conversations, where the information is genuinely unavailable elsewhere.

How long should I keep transcripts?

Ninety days covers most analytical uses. Extract the patterns you care about into a separate document and let the raw transcripts expire automatically. An archive you are not using is a liability rather than an asset.

What is the single best use of this for a small business?

Searching your own customer conversations for the questions people actually ask, in the words they actually use, and then fixing your website and sales materials to answer those. It is unglamorous, it takes an hour a month, and it outperforms almost everything else you could do with the same time.

Related Coverage

• Best AI Writing Tools for Small Business – turning raw transcripts into content people will actually read.
• AI Agents, Explained – what sits above transcription once the text exists, and what it costs.
• How to Start a Business With AI – where cheap AI tooling genuinely changes the economics, and where it does not.

Faceted Media Magazine covers business, AI, and entrepreneurship for the people building what’s next.