A sales team installs an AI notetaker. Every call for the next eight months gets recorded, transcribed, and summarised. The summaries are genuinely good.
Nobody has opened one since the week it was set up.
The transcripts sit in a folder in another application. When a rep prepares for a follow-up call, they open the deal record in the CRM, see the same three-line note someone typed in March, and go into the call with nothing. The transcripts exist and they might as well not.
This is the most common AI implementation failure I see at small companies. The tool works perfectly. The workflow around it was never built.
Why the transcript folder is a dead end
People use the system they already have open. That is the whole principle, and almost every AI tool rollout ignores it.
A rep working a deal opens the CRM. A manager checking pipeline opens the CRM. Someone writing a proposal opens the CRM. If the useful information from the last conversation lives somewhere else, it is not part of anyone's work.
There is a second problem. A transcript is unstructured text. You cannot report on it, filter by it, or count anything in it. Eight months of perfect transcripts and you still cannot answer a simple question like how many deals mentioned a specific competitor last quarter.
So the goal is not better notes. The goal is structured fields on the record, produced automatically, that a person can read in ten seconds and a report can count.
The tools, and what genuinely separates them
Transcription quality is broadly similar across the mainstream tools now. Compare them on where the output goes and how much control you have over it.
HubSpot's built-in call recording and AI summaries. If you already run HubSpot, start here. Availability depends on your subscription tier, so check the current feature list on HubSpot's pricing pages. The advantage is that the output attaches directly to the contact and deal, with no integration to build or maintain when something changes.
Fathom. Records, transcribes, and summarises with very little input from you, and integrates with several CRMs. Suits teams who want to stop taking notes entirely.
Fireflies. Similar recording and summary model, with an API and integrations that make it a reasonable choice when you need to build your own extraction step.
Granola. Built around notes you type yourself. It combines your rough notes with the transcript into a cleaned-up summary, which suits people who think while writing and want the AI to fill the gaps rather than replace the habit.
Otter. The long-standing option, strong on general meeting transcription, less specifically built around a sales pipeline.
Pricing moves, so check the vendor pages directly. And run a real comparison rather than reading reviews: take three actual sales calls, run two tools on them, and look at whether the summary captured the thing that mattered. They differ more on what they consider important than on what words they heard.
Decide what fields the call should fill
This is the step everyone skips, and it is the one that makes the difference.
Before connecting anything, decide what a sales call should produce on the record. Six fields is usually enough:
- Next step and date. What was agreed and when it happens.
- The problem in their words. The exact phrasing the prospect used, not your category for it.
- Budget signal. Whether money was discussed and what was said.
- Decision process. Who else has to agree and what the timeline is.
- Objections raised. Recorded even when they were answered on the call.
- Competitors mentioned. By name.
Create these as properties in your CRM first. In HubSpot that means custom deal properties, mostly text with a couple of picklists. Then the AI has a defined target instead of producing free-form prose that has nowhere to go.
Keep the transcript too, attached to the record. It is useful when someone needs to check exactly what was said. It is just not the deliverable.
Getting the note into the CRM as data
Three routes, depending on what you run.
Native integration. If your notetaker writes to your CRM directly, use it, and spend your effort configuring what it writes rather than building anything. Check what it actually populates. Many integrations attach the summary as a note and update nothing structured, which is halfway there.
Custom extraction. When the native option does not fill your fields, send the transcript from the notetaker's webhook to n8n, Make, or Zapier, run it through a language model with an extraction prompt that returns your six fields as structured output, then write those fields to the CRM. This is a few hours of work and it is the version that actually gives you reporting.
Build the extraction prompt against 20 real transcripts before you connect the write step. Ask for JSON with your exact field names, tell it to return an empty value when something was not discussed, and read the output for all 20. The failure you are hunting for is confident invention: a model that produces a plausible budget figure from a call where budget never came up.
Manual paste. Nobody does this consistently. Skip it.
One thing to get right before any of it goes live: announce the recording at the start of every call and get a verbal yes. Consent law varies by province, state, and country, and some places require all parties to agree. Most tools can post a notice in the meeting automatically. If you sell across borders from BC, treat all-party consent as your standing practice and have a lawyer confirm your specific situation once.
The human step you cannot automate away
Have the AI draft the update. Have a person confirm it before it writes.
AI extraction misreads context in specific, predictable ways. A prospect saying "we spent about 40 thousand on this last year" becomes a budget figure for this deal. A polite "we will definitely take a look" becomes a commitment. A competitor mentioned as an example of what they do not want becomes a competitive threat on the record.
A wrong field in a CRM is worse than an empty one, because it flows into forecasts and reports where nobody questions it again. This is the same trap as any AI output that ships without review: the confident wrong answer costs more than the missing answer.
The review is 30 seconds. The summary appears, the rep reads it, corrects anything wrong, and approves. Do it right after the call while the conversation is fresh, and it never becomes a backlog.
Reps who resist this are usually reacting to the wrong thing. They are being asked to check that a machine understood their conversation, which is a much smaller job than writing notes.
What this gives you once it runs
Three things, none of which are available from a folder of transcripts.
Reporting on the conversation. Which objections appear most often and at which stage. How frequently a specific competitor comes up. Whether deals with a documented next step close at a different rate than deals without one. These are answerable in a CRM report once the data is structured, and unanswerable before that.
Coaching that uses evidence. A manager can see that a rep's calls consistently end without an agreed next step, which is a specific and fixable behaviour. Far more useful than a general observation that follow-up needs work.
Marketing material built from real language. The problem field, filled in with the prospect's own words across a hundred calls, is the best source of website copy you will ever have. It also tells you which objections to answer directly on the site, which is where copy that converts actually comes from. Most B2B websites describe the product in internal vocabulary. Your customers do not use that vocabulary, and now you have a record of what they do use.
That last one is the reason marketing should care about a sales tool at all.
The takeaway
Installing an AI notetaker takes ten minutes. It produces nothing on its own.
Define the fields the call should fill. Get the extraction writing to those fields. Keep a person in the loop before the record updates. Do that and eight months of calls becomes a database of what your market actually says, instead of a folder nobody opens.