Straight answer: AI notetakers transcribe words accurately but they invent meaning, and that gap between what was said and what the summary claims was agreed is where small businesses get into real trouble. Use them for notes, never for decisions, and always tell the person on the other end of the call that a bot is listening.
What these tools do now
Three years ago a "notetaker" was just a transcript. Now Otter, Fireflies, Fathom, Read AI and Zoom's own AI Companion do a lot more than that. They join the call as a visible bot (usually with a name like "Fireflies Notetaker" sitting in the participant list), record audio, transcribe it, then generate a summary, a list of action items, sometimes sentiment scoring, and increasingly they push all of that straight into your CRM without you touching anything.
That last bit is why I started using one in the first place. I run consulting calls back to back some weeks, four or five a day, and I was losing details. Not big things, small ones: what someone's budget was, which competitor they mentioned, whether they said "next quarter" or "next year." So in early 2026 I picked three tools and ran them for three months.
The three months, in actual numbers
I tracked this because I wanted to know if it was worth the money, not just the novelty.
- 47 client and prospect calls recorded across the period, average length 34 minutes
- Three tools tested: Fireflies (paid tier, roughly £8/month at the time), Fathom (free tier), and Otter Business (around £16 per user/month)
- Transcript word accuracy across all three: consistently above 90 percent for clear audio, dropping to around 70 to 75 percent on calls with strong accents or bad connections (and I say this as someone with a Hebrew-accented English who gets mangled by voice tech constantly)
- Summary accuracy, meaning "did the three-bullet AI summary correctly represent what was decided": this is the number that matters, and across my own spot-checking it was wrong or misleading on roughly 1 in every 8 calls
That last figure is the one worth sitting with. Not the transcript. The transcript is good these days, it is close to a solved problem. The summary is not, and most people never check the summary against the actual transcript because the whole point of the tool was to save them the time of doing that.
The call that nearly went wrong
Here's the specific one. A prospective client, a director at a small manufacturing firm, was talking through a project on a discovery call in March. She said something like: "we could probably stretch to fifteen hundred a month if this goes well, but to start with we'd want to keep it lower while we test it." Fireflies' summary read: "Client agreed to a monthly retainer of £1,500."
That summary went straight into my CRM. If I hadn't listened back to the actual recording before sending my proposal, I would have quoted based on a figure she never committed to, and either undercut myself on the low starting number she'd asked for, or looked greedy quoting the ceiling as if it were the agreed price. Neither is a great first impression. The AI hadn't lied exactly, it had compressed a conditional, hedged sentence into a clean decision because clean decisions are what summaries are built to produce. It got every word in the transcript right and still got the meaning wrong.
This is the uncomfortable part that most write-ups about these tools skip: the risk isn't the AI mishearing you, it's the AI understanding you perfectly well and then flattening nuance into false certainty because a tidy bullet point reads better than "maybe, under conditions." Businesses that run their pricing, their scope, or their promises off these summaries without checking the source audio are one bad quarter away from a client dispute they can't fully explain.
The consent problem nobody mentions
The other issue is quieter and it's about trust, not accuracy. Most of these bots announce themselves by name in the call, which technically satisfies transparency requirements under GDPR in the UK and EU. But "a bot named Fathom Notetaker joined the call" is not the same as a client understanding that their voice is being recorded, transcribed, stored on a US server, and potentially fed into a summary that gets forwarded to your whole team, your VA, and sometimes synced into a CRM that a third party can access.
I had two clients in those three months ask, mid-call, "is that thing recording?" after they'd already been on the call for ten minutes. Neither had clocked the participant list. That's not their fault, it's a easy thing to miss on a laptop screen, and it told me the tools' idea of "consent" is doing the legal minimum, not the decent minimum.
What I do now, and what I'd recommend to any small business owner using these tools, is say it out loud at the start of every call: "just so you know, I've got an AI notetaker running so I can focus on the conversation rather than typing, happy to turn it off if you'd rather." It takes four seconds and it has never once made a client uncomfortable. Not saying it, and having them notice halfway through, has.
Where they earn their keep
None of this means don't use them. Here's where they've saved me time,:
- Action items after sales calls. Even with the occasional wrong summary, having a rough list within two minutes of hanging up beats me trying to remember five things while jumping into the next call.
- Client onboarding calls where there's a lot of detail (logins, brand guidelines, contact names) that used to get lost in my own scribbled notes.
- Handover to a VA or team member who wasn't on the call. A transcript they can search is far better than a two-line email from me trying to remember the gist three hours later.
- Spotting patterns across calls. Fireflies in particular will tell you how much you talked versus the client, which is a useful gut check if you suspect you're doing too much pitching and not enough listening.
A simple rule set if you're going to use one
After three months, here's what I do, not what the tool's marketing tells you to do:
- Say it out loud at the start of the call, every time, no exceptions.
- Never send a proposal, quote, or contract detail based on the AI summary alone. Read the transcript for anything involving money, dates, or scope.
- Turn it off for calls that involve anything sensitive: staff issues, a client venting about a personal problem, or anything you wouldn't want sitting on a US server indefinitely.
- Set a deletion policy. Most tools default to keeping recordings forever. I set mine to auto-delete after 90 days because I don't need a permanent audio archive of every sales call I've ever had.
- Check the summary against the transcript at least once a month, start to finish, so you know your actual error rate rather than assuming the tool is more reliable than it is.
If you're not technical and the idea of picking apart AI tools like this feels like a lot, this is exactly the kind of thing I cover with clients in plain terms, no jargon, over on how I use AI day to day as someone who isn't a developer. You don't need to understand the model to know when to trust it and when not to.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Who shouldn't bother with one at all
If you're in a regulated field where a misquoted number or misattributed decision has real legal weight (financial advice, legal services, anything involving health information), I would not run an AI notetaker on client calls at all right now, not until the summary accuracy problem improves. The transcript is fine for your own records if you're clear it's just raw text. The auto-generated summary is the part I'd switch off entirely in those industries, because a wrongly summarised "agreement" is exactly the kind of thing that ends up in a complaint file.
For everyone else, sole traders, consultants, small agencies, the tools are worth the £10 or £15 a month, as long as you treat the summary as a rough first draft written by an enthusiastic junior who occasionally makes things up with total confidence, rather than as a source of truth.
The bit that changed how I run calls
The most useful side effect of testing this for three months wasn't the notes at all. It was that knowing every word was being transcribed made me tighten up how I speak on sales calls. Less hedging, less "maybe we could look at," more direct statements of price and scope, because I knew a machine was going to try to summarise whatever I said and I'd rather it had something clean to work with. Oddly, that's the same discipline that made the month my website started paying me again work as well as it did: being specific instead of vague, on the page and on the phone, is what converts people. The AI notetaker didn't teach me anything new about marketing. It just made the cost of being vague immediately visible, in black and white, five minutes after every call.
It's a similar lesson to the one I learned watching 569 AI-written blog posts get indexed at a 5 percent rate: volume and speed from AI tools feel productive, but the thing that determines whether any of it holds up is whether a human checked the output against reality before trusting it. A transcript nobody reads is the audio version of a blog post nobody proofread. It looks like work. It isn't finished work.
If you're weighing up whether tools like this belong on your website too, the same principle applies as it does on a call: anything that captures information from a real person, whether that's a chatbot, a form, or an interactive calculator qualifying leads before they ever book a call, needs a human checking what it's telling you before you act on it, not just trusting the neat summary at the end.
Frequently asked questions
Do AI notetakers breach GDPR if I use them on client calls?
Not automatically, but you need a lawful basis to record and process someone's voice data, and simply having the bot announce its name in the participant list is a weak version of consent. Saying it out loud at the start of the call, telling clients where the data is stored, and having a deletion policy puts you on much firmer ground.
Which AI notetaker is most accurate for UK accents?
In my own testing across 47 calls, Fireflies and Otter performed similarly on clear audio, both above 90 percent word accuracy, but all three tools I tried dropped noticeably on strong regional accents or poor connections, sometimes below 75 percent. None of them is meaningfully better than the others on this specific problem yet.
Should I trust the AI-generated summary of a client call over my own notes?
No. Treat the summary as a rough first draft, not a record of fact. In my three months of testing, roughly 1 in 8 summaries misrepresented what was agreed, particularly around price and scope, because the AI tends to turn hedged or conditional statements into clean, confident decisions.
Is it rude to tell a client an AI is recording the call?
No, the opposite. In my experience, saying it plainly at the start of the call takes a few seconds and clients almost always appreciate the transparency. What damages trust is a client noticing a recording bot ten minutes into a call that nobody mentioned.
Related: looking for an AI consultant for small business? Here is how I can help.
Related reading: AI Notetakers Are Quietly Costing You Client Trust. Here's What I Do Instead and Should You Let an AI Notetaker Into Your Client Meetings?.
