AI meeting assistants for accountants and bookkeepers: what works

Every client meeting creates a second, smaller job. The file note. The follow-up email. The two or three things somebody promised to do. Nobody bills it, and it gets done at 6pm, on a Sunday, or not at all.

So “what is the best AI note taker” is the wrong opening question. The useful one is how your firm runs the process: what ends up in the file, who checks it, and where it lands so nobody re-types it. Those conversations are already the most under-used data in a firm.

This is the spec every tool is measured against, and almost none meet it. A file note a reviewer accepts answers five things: what the client asked, what you told them, the facts and assumptions the advice rested on, what changed about the scope, and who owns the next step by when.

Generic recaps give you topics and action items. “Review VAT issue” is a topic: no question, no owner, no date. The scope conversation is worse. A client mentions a new entity, somebody says “we can look at that”, and the engagement widens with no written trail and no fee attached. A summary drops it, because nothing in it knows that phrase is commercial. So judge every tool on whether its output arrives shaped like the record you are obliged to keep. If not, our prompt patterns for accountants are the cheaper fix.

Prep gets skipped, so the first ten minutes go on reconstruction while the client waits. Tools holding your meeting history can write that in advance: a pre-meeting brief built from prior conversations with that client plus the calendar invite, which is why recurring clients get the best ones. Keep brief and agenda apart. A brief is internal: what to raise, what is outstanding, what worried the client last time. An agenda is client-facing, sent ahead.

The most useful move here needs no vendor. Write a real agenda into the calendar invite instead of “Catch up”. Every tool that reads the invite produces better output, and every human who reads it arrives prepared.

Suite-native

Microsoft Teams Premium and Zoom Workplace, where Zoom’s meeting AI now lives after it retired the AI Companion brand in June 2026. As at September 2026, Microsoft’s intelligent recap needs a Teams Premium or Microsoft 365 Copilot licence on top of a Microsoft 365 subscription and a Teams licence; Zoom includes meeting summaries with eligible paid Zoom Workplace plans and sells its agentic tier separately as ZoomMate.

Teams carries the strongest governance story here, and a structural one: the recording never leaves the Microsoft 365 tenancy your IT already administers. Zoom’s is contractual, not architectural. Both fail on reach, being one platform only. Bundled transcription is where practitioners complain loudest: accents, cross-talk and firm vocabulary come back mangled, and a lodgement name transcribed wrong is a file note nobody trusts.

Generalist

Otter.ai, Fireflies and Fathom (fathom.ai, unrelated to the reporting tool of the same name). Mature, cheap, cross-platform, hard to beat for a searchable record. Sales-intelligence tools get suggested in the same breath and should not be: Gong and Chorus are built for deal review, not compliance, and Motion is a scheduler, not a notes tool.

They fail on output shape: a meeting artefact, not a firm artefact, and nothing reaches practice management without middleware somebody owns.

Accounting-native

Tools built for accounting and bookkeeping firms rather than for meetings in general. The difference is what comes out: a file note shaped the way a reviewer expects, a drafted follow-up email, engagement letters and minutes, and a flag on the scope signal a client drops in passing, written into practice management, not an inbox.

They fail on reach too, and on youth. A specialist covers only the meetings its integrations reach, so you still need a process for the client on a platform it does not join. And a young vendor’s file-note template bets on their idea of a file note, not yours.

Before anyone books a demo, delete the shapes your constraints rule out. IT policy forbids third-party bots: suite-native, argument over. Clients on three platforms: suite-native is gone. Outcomes must reach practice management untouched: generalists are gone, unless you will own middleware.

Three routes get an outcome to where work happens. Copy and paste is the suite-native default, and the first thing skipped on a busy Thursday. Middleware wires a generalist into your task tool with Zapier, an API or an MCP connection, then becomes something one person owns forever, as in proposal creation with Ignition. Native integration writes into practice management itself.

Before you shortlist, write down where a meeting outcome has to end up: a job in Xero Practice Manager (XPM), a document in FYI, a task in Kloud Connect, a checklist in AccountKit, a work item in Karbon. In a US firm: Financial Cents, Canopy, TaxDome, Jetpack Workflow, Firm360 or Mango Practice Management. If a tool cannot reach it, someone in your firm is the integration, and their time is a cost, not a saving. That destination matters more than transcript quality: a transcript nobody reads is worse than none, a discoverable record nobody has checked.

Per seat breaks where everyone from graduate to partner sits in front of clients, so firms ration licences to a few partners, which is how a tool quietly dies. Pooled hours with unlimited users inverts that, until an onboarding wave or a filing deadline burns the pool. Free and bundled is a real starting point: Fathom’s free tier records and transcribes without limit and rations AI summaries each month, and an eligible paid Zoom Workplace plan already includes them. Prove the habit there first: the cheapest line item is often the most expensive, because the cost arrives as re-typing.

Vendors genuinely diverge here. Each entry was read from the vendor’s own policy or trust documentation in September 2026, and those change.

• Otter.ai says it trains its proprietary AI on de-identified audio recordings and transcriptions, which may contain personal information. A documented opt-out exists on Enterprise plans, arranged through an account manager rather than a self-serve toggle, and Otter states Enterprise workspaces are excluded from model training by default. It is also defending consolidated US privacy litigation over recording participants without consent, where an August 2026 order let core claims proceed: plausible enough to be heard, not proven.
• Fathom says it does not authorise third-party AI providers to train on your meeting content, and that it may use de-identified data from meeting content to train its own in-house models, with an opt-out in your account settings.
• Fireflies states that it does not use personal information for AI model training and contractually prohibits its vendors from doing so.
• Zoom states it does not use customer audio, video, chat, screen sharing or other customer content to train its own or third-party AI models. The content still sits in Zoom’s cloud, not yours, and Zoom names outside providers including Anthropic, OpenAI, Google and Perplexity among the subprocessors behind its AI features. That assurance is contractual, not a tenancy boundary.
• Teams Premium inherits the Microsoft 365 tenancy and compliance boundary your IT already governs.

Notice how much work “de-identified” does there. It is not the claim “we do not train on your content”, and when the audio is a client’s tax position, that gap belongs with your risk partner.

Assume it works. Within a fortnight it surfaces the exchanges your file notes used to lose, the scope conversation above among them. A firm that flags those and does nothing has bought a better record of work it still gives away. Decide three things in week one.

Who reads the flags. Not the person who took the meeting, who said the words and will rationalise them. Put them in front of the engagement owner, or whoever runs job budgets, weekly and in a batch. Minutes as a list; never, as a notification.

What the re-scope conversation sounds like. Short, factual, early. “You mentioned the new entity on Tuesday. That sits outside what we agreed in March. Do you want us to pick it up?” Within days that reads as attention; at the end of the job, as a surprise.

How you decide variation or write-off. Did the client ask, or did someone here volunteer it? Volunteered work is yours to carry. Once, or a pattern? Ten minutes once is goodwill; ten minutes monthly is a service you have started delivering without saying so. Can you name it as a distinct piece of work in one sentence? If you cannot name it you cannot vary it, and the honest call is a write-off and a tighter scope next time.

A bookkeeping practice runs more meetings than an accounting firm of the same size, and shorter ones. Monthly check-ins, a BAS or VAT query, a payroll change, a client walking through a pile of receipts. Many are fifteen minutes on the phone, not an hour on Teams. That changes what the tool has to do well.

The record is an approval, not advice. A bookkeeper’s file note rarely captures recommendations. It captures what the client approved or told you: reclassify those transactions, add the new employee from the first of the month, the supplier invoice is personal and stays out of the business. When a client later says “I never agreed to that”, the note is the answer. Judge a tool on whether it pulls out the approvals and the facts the client gave you, with a date, and sends them to the client in a short follow-up they can correct.

The phone matters more than the meeting link. If half your client conversations happen on a mobile, a tool that only joins scheduled video calls covers half your work. Check whether it records a phone call or an in-person chat with consent, not just a Zoom or Teams meeting.

Where it lands is a task, not a document. Most of what a bookkeeping call produces is work: chase the missing statement, set up the new payroll item, lodge after the client confirms. If the output cannot become a task in the system the team actually works from, the bookkeeper will re-type it, and the saving disappears.

Price by meetings, not by people. Small bookkeeping teams take a lot of short calls each. Per-seat pricing is usually fine for a team of three; it is the pooled-hours plans that a month of BAS or VAT deadline calls can burn through. Count your meetings for a typical month and a deadline month before you choose.

The five vendor questions below apply unchanged. So does the judgement about which conversations to capture at all: our list of client meetings worth recording works for a bookkeeping practice too, and if you are setting one up, start with how to start a bookkeeping business.

Put these to the vendor in writing, and the answers into your firm’s one-page AI policy.

1. Consent. How does the tool announce itself and capture consent from every participant, including the client, who is not your user? US states, the UK and Australia set different floors.
2. Training. Is our audio or transcript used to train your models or anyone else’s, identified or de-identified? Get a yes or no in writing.
3. Retention and deletion. How long is anything kept, can we set our own period, and what happens when we ask for deletion or leave?
4. Residency and subprocessors. Which country holds the data, and which third-party AI providers see the transcript on the way to a summary?
5. Access. Who inside our firm sees whose meetings? Client conversations should not be firm-wide reading by default.

A good answer arrives as documentation: a current SOC 2 Type II report readable under NDA, a GDPR-ready data processing agreement naming subprocessors, and a HIPAA business associate agreement where offered. A vendor that replies in adjectives has not answered. Our guide to AI for accounting firms sets meeting notes beside the other candidates.

---

Partner note

Vinyl is an accounting-native option of the kind described above: an AI meeting assistant for accounting and bookkeeping firms that produces file notes, follow-up email drafts, engagement letters and minutes, builds pre-meeting briefs from prior conversations and the calendar invite, and deploys them into practice management. Its integrations are shaped around FYI, Xero Practice Manager (XPM) and Karbon, so a CRM-centric or US firm should check coverage first. Firm plans price pooled meeting hours with per-minute overage.

Is an AI meeting assistant worth it for a small bookkeeping practice?

Usually, if a meaningful share of your client contact is spoken rather than emailed. The gain is not a prettier transcript, it is having every client approval written down and sent back for the client to confirm. Trial it on a month of check-ins and count the follow-ups you no longer type.

Who should own an AI meeting notes rollout inside the firm?

One partner as sponsor, one operations or practice manager as owner. Not IT alone. The owner needs authority over the two things software cannot decide: the house file-note template, and the review step that runs until output is trusted unread.

What do we do when a client says no to being recorded?

Take the no without negotiating, and have a fallback: no bot, notes typed live, a short summary afterwards to correct. Record the refusal on the file so nobody re-asks. Many who decline recording accept a written summary, so ask.

Are AI meeting notes discoverable in a dispute?

Not legal advice, your lawyer should answer for your jurisdiction, but assume transcripts and summaries are business records producible in a dispute. Accountant-client communications generally do not attract the privilege lawyer-client ones do, with narrow exceptions such as the US federally authorized tax practitioner privilege under IRC section 7525, which reaches tax advice in non-criminal federal matters and little else.

How do we run a trial that actually tells us something?

Pick meetings you already hold good file notes for, and grade the tool against the note a reviewer accepted at the time: a baseline rather than an impression. Run it internally for a fortnight before any client sees a bot, and set the decision date up front.

Do we still need to keep the recording once the note is written?

Often not. Once a note is reviewed and filed, the recording is a second copy of client information with its own retention, access and breach exposure. Most firms should keep the note and delete the audio on a short schedule.