AI assistants

What an AI assistant actually does in a firm

Strip the marketing and there are four jobs. Listening: a notetaker sits in the client meeting, transcribes it, and produces a summary with the actions and the numbers that were mentioned. Drafting: correspondence, engagement notes, the explanation of a variance, written from the file rather than from a blank page. Coding: the bank transactions and receipts that a rules engine leaves in the exceptions queue. Answering: a question about a client, put to the file in plain words and answered from it.

Most tools in this category do one of those well and gesture at the others. A practice choosing one should decide which job it is buying for, and judge the tool on that job alone. A notetaker that also drafts emails is a notetaker; the drafting is a bonus until it is tested on real correspondence.

The meeting notetaker, and where Vinyl fits

The meeting is where most of a firm's unbilled knowledge is created and lost. A partner spends an hour with a client, remembers the three decisions that matter, and files none of the context. A notetaker built for accounting firms captures the meeting, writes it up in the shape a practice uses (decisions, actions, figures, who owns what), and files it against the client record in the practice-management system, so the next person to open the file sees what was said.

Vinyl (usevinyl.com) is built for that job specifically: it is a notetaker and meeting assistant for accountants and bookkeepers, and it writes the meeting back to the client record in the practice tools a firm already runs. Vinyl is a commercial partner of The Firm and shares its founders, which is why it holds the labelled Category Partner placement in this category. That is a disclosure, not a verdict: if a firm does not run client meetings as a service, a notetaker of any kind is the wrong first purchase, and a firm that does should trial it on a fortnight of real meetings before deciding.

What to check on any notetaker, Vinyl included: whether it joins the platforms the firm actually meets on, where the recording and transcript are stored and for how long, what a client is told and asked to consent to, whether the summary can be edited before it is filed, and which practice-management or document systems it writes back to. The last one decides whether the notes are found again or lost in a second silo.

Data, confidentiality and the professional obligation

Every tool in this category reads client data. The questions a practice must be able to answer, to a client or to a professional body, are where the data is processed, where it is stored, whether it is used to train the vendor's models, who at the vendor can see it, and how it is deleted when the engagement ends. A vendor who cannot answer those in writing is not ready for a regulated practice, whatever the demo looks like.

Treat the answers as part of the firm's own data register. If the practice already documents where client data lives for its security policy or its professional indemnity insurer, an AI assistant is a new line in that document, not an exception to it.

Judging the output

The failure mode is not that the tool is wrong; it is that it is confidently wrong in a way nobody checks. Before adopting one, run it beside the current process for a few weeks and compare: does the summary miss the number that mattered, does the draft assert something the file does not support, does the coding put a director loan somewhere a reviewer would never look. Decide in advance who reviews the output and how visibly the tool marks what it is unsure about.

The good tools make the review easy: they show their sources, they flag low confidence, they hand back a draft rather than a decision. That is a feature to shortlist on, and it is a bigger differentiator than the model underneath.

What it costs a practice, and how firms recover it

Pricing in this category is per user per month almost everywhere, sometimes per meeting or per document processed. The directory records what each vendor publishes; it does not tell a firm what to charge its clients, and no figure here should be read that way. The practical exercise is to name the hours the tool is expected to recover per client per month, and to decide whether that saving stays in overhead or becomes part of a fixed-fee engagement. A firm that cannot name the hours should not buy yet.

Recovery models firms describe: absorbing the cost as overhead and pricing the engagement on the outcome; itemising a "meeting notes and actions" line inside an advisory package; or pricing the reviewed output (a written summary the client keeps) as a deliverable. Which one fits depends on how the firm already prices advisory work, which is the real question behind every purchase in this category.

What to avoid

A tool with no written answer on data residency and model training. A tool that cannot write back to the systems the firm runs, which guarantees a second copy of every note. A trial judged on the demo meeting rather than a fortnight of real ones. And buying for the drafting when the firm's problem is the meetings, or the reverse: the four jobs are different products wearing one label.