How do I get written records of calls with customers?

This page may contain affiliate links from which Blur Studio earns commissions.

When a customer disputes what was agreed, "I said" versus "they said" is a bad place to be. A written, searchable record of the call settles it in seconds instead of becoming a he-said-she-said argument, and it also catches the small details nobody writes down in a CRM note: the exact price quoted, the date promised, the "yes, go ahead" that authorized the work.

Check consent rules before you record anything

This is the step people skip, and it's the one that can turn a helpful record into a legal problem. Recording laws vary by country and, in the US, by state. Some require only one party (you) to consent; others require every party on the call to know and agree. If you operate across multiple states or countries, the safest default is a short, consistent disclosure on every call ("this call may be recorded for quality and training purposes") rather than assuming your home state's rule applies everywhere your customers are. Check your own jurisdiction's current rule before you record a single call, not after. For a US-based service business, this usually means checking whether your state is "one-party consent" (your own knowledge of the recording is enough) or "all-party consent" (everyone on the call needs to know), and if you take calls from customers in a different state or country than the one you operate in, the stricter of the two rules is the safer one to follow by default.

How much does it actually cost to transcribe calls?

Three real options, priced as of this check:

• Otter: free for up to 300 minutes a month with a 30-minute cap per conversation. Pro is $8.33 a month on annual billing (or $16.99 monthly) for 1,200 minutes. It's built as a meeting-and-call app with its own bot that can join calls directly.
• Rev: AI transcription runs $0.25 a minute, so $15 for a one-hour call. Human-verified transcription, if you need higher accuracy for a dispute, costs $1.99 a minute, or about $119 an hour.
• ElevenLabs Speech to Text (Scribe v2): $0.22 an hour through the API, roughly $0.0037 a minute, a fraction of either of the above. It handles files up to 3GB or 10 hours long, labels up to 32 separate speakers, and adding your own vocabulary (product names, part numbers) costs an extra $0.05 an hour.

The catch with Scribe v2 is that it's a developer tool, not a finished app: you (or someone you hire) need to get the recorded audio file to it and the resulting text back out, usually through a small script or an automation tool. ElevenLabs' plans also include some Speech to Text access on the Free tier, though it's limited: Speech to Text draws from the same shared credit pool as every other ElevenLabs feature, at roughly 330 credits a minute, so the Free plan's 10,000 monthly credits cover about 30 minutes of transcription a month before you'd need a paid plan or the pay-as-you-go API rate above.

A worked example: what this costs at real volume

Take a five-person plumbing crew answering roughly 40 customer calls a week, averaging 6 minutes each: about 4 hours of audio a week, 16 to 17 hours a month. On Rev's AI rate that's $240 to $255 a month. On Otter's Pro tier, 1,200 minutes a month covers that volume comfortably for a flat $8.33 on annual billing or $16.99 monthly. On Scribe v2's API rate, the same 17 hours costs roughly $3.74 a month, the cheapest by a wide margin, but only once someone has built or bought the small piece of automation that gets the recorded file to the API and the transcript back out. For that crew, the honest math is that Otter is close to free relative to the time it saves, and genuinely less total effort than wiring up an API to save a few dollars a month. The API only wins once call volume is high enough, or frequent enough, that the time spent building the connection pays for itself, and for most small crews that point is well past where they are today.

What you actually get back, and how to make it useful

Scribe v2's output isn't a block of plain text; it's structured data. Each segment of speech comes with a speaker label (so you can tell the customer's words from your rep's), a timestamp, and a confidence score per word. That structure is what makes a transcript archive actually searchable later, rather than a folder of text files you'd have to read one by one to find anything. If you're building this yourself, keep that structure in whatever you store it in, a database field rather than a flattened text file, since flattening it at the first step throws away the exact detail that makes a dispute easy to settle: who said what, and when.

One genuinely useful recent change, per ElevenLabs' own changelog (28 September 2026): transcript editing is now supported for both batch and real-time transcription, so a misheard word or name doesn't have to stay wrong in your record forever. If a transcript gets a part number or a quoted price wrong, which is exactly the kind of error that matters most in a dispute, you can now correct it in place rather than treating the first AI output as final.

What I'd actually tell a small business owner here

If you're doing this because one dispute went badly and you want it to never happen again, don't overthink the tool. Otter is the fastest way to get a working system today: it joins calls, transcribes automatically, and needs no building. The $0.22-an-hour API rate on Scribe v2 only matters if you're transcribing enough volume that the per-minute saving adds up to real money, or if you already want the transcript to flow into something else (a CRM note, a searchable archive, an alert when a customer says a specific word). For a five-person crew doing a few dozen calls a week, that volume usually isn't there yet, and the time you'd spend wiring up an API to save a few dollars a month isn't worth it. The API becomes the right call once you're doing hundreds of calls a month or want the transcript to trigger something automatically.

How to set this up, step by step

1. Confirm your recording-consent rule and add a short disclosure line to your call flow or on-hold message.
2. Decide whether you want a ready-made app (Otter, Rev) or a building block (Scribe v2 via API). If you're not sure, start with the ready-made option; you can always move the volume to an API later.
3. If you're recording through your phone system, confirm it can export audio files in a common format (WAV or MP3). Most VoIP and call-tracking systems can. Poor audio can be improved first: seecleaning up audio before transcribing it.
4. Run a test call through whichever tool you pick, and check the transcript against what was actually said, especially names, numbers and prices.
5. If accuracy on specific words is off (part numbers, product names), add them as custom vocabulary if your tool supports it. On Scribe v2 that's the keyterm prompting feature, at an extra $0.05 an hour. More ideas are instopping transcripts getting technical terms wrong.
6. Store the transcripts somewhere searchable and set a retention policy; a pile of text files nobody can search is barely better than the raw audio.

One practical detail that trips people up: Scribe v2 returns a transcript, not a summary. If what you actually want is "what did we agree to" in two sentences, you need a second step, either a summarization tool or a person, reading the transcript. Don't expect the transcription step alone to answer "what happened on this call." A separate guide coversgetting a call summary without relistening.

When to skip all of this

If you're only trying to settle one dispute and don't expect it to recur often, a one-off human transcription from Rev at $1.99 a minute is simpler than setting up any ongoing system, and you get a transcript a reasonable person would trust in a dispute without having to build or manage anything. It's also worth skipping any of this if you already outsource your call answering to a service: most answering services and virtual receptionists log call notes or provide recordings as part of what you're already paying for, so check what you have before building a parallel transcription system for calls someone else is already handling.

How I'd approach this

Where this usually gets interesting for me is past the point of "just transcribe the call." A business with real call volume often wants the transcript to do something: flag calls where a price was quoted, pull every call mentioning a specific product issue, or feed a weekly summary to an owner who doesn't have time to listen to calls. That's a genuinely useful custom build, reading from whatever's cheapest to transcribe with and surfacing the two or three things that actually matter, rather than a pile of searchable text nobody reads. If you're already past the "which app do I pick" stage and into "I want this to tell me something," that's the conversation worth having with me directly. To skip the manual step, see how totranscribe every call automatically.