The short version
The highest-value use in a practice is not client-facing advice, it is the preparation work that currently eats junior time: reconciliation, close checklists, statement consistency and getting messy client records into a usable state. Templates now cover the shape of that work. The adaptation — per-client conventions, your review thresholds, and an audit trail your professional body would accept — is what a practice actually has to build.
What can AI actually do in an accounting practice today?
Anthropic’s financial services agents include a general ledger reconciler, a month-end closer that runs close checklists and prepares journal entries, and a statement auditor that reviews for consistency and audit-readiness. Alongside these sit templates aimed at research and client coverage — meeting preparation, earnings review, model building.
They deploy in two ways: as plugins inside Claude Cowork or Claude Code, sitting alongside how your staff already work, or as managed agents running autonomously on a schedule with per-tool permissions, credential vaults and audit logs. For a practice, that distinction is the important one, and we come back to it.
On the smaller end, the small-business workflows connect directly to QuickBooks and cover monthly close, invoice tracking and margin analysis — relevant both for your own practice management and for what your clients will start asking about.
Where the value actually is
Not in advice. Partners give advice, and clients pay for a name on an opinion. The cost in a practice sits in preparation:
- Getting client records into a usable state. The bank feed nobody categorised, the receipts in a shoebox photographed at an angle, the sales ledger maintained in a spreadsheet.
- Reconciliation across sources that do not agree and were never meant to.
- Close checklists where the work is remembering the steps and chasing the missing item, not the arithmetic.
- Consistency checks across a set of statements — the kind of review that is tedious, mechanical, and where a human misses things precisely because it is tedious.
- Drafting the same explanatory notes for the fortieth client with a slightly different set of numbers.
Every one of these is high-volume, rule-governed, and currently done by someone whose time you bill at a rate that makes it painful.
Why can’t we just switch the template on?
Per-client conventions
Every client codes things differently, and much of that logic is undocumented. Which expenses are capitalised, how intercompany balances are treated, which accruals are standing and which are judgement calls. A reconciler that does not know your client’s conventions produces confident output that a senior then has to unpick, which is worse than no output.
Your review policy
What may be posted without review, and what must always be looked at? This is a professional standards question before it is a technical one. The output of an agent is preparation, not sign-off, and the practice needs an explicit position on where the line sits — documented, because you will be asked.
An audit trail your professional body would accept
If a journal was prepared by an agent, you need to be able to show what it saw, what it produced, who reviewed it and what they saw at the time. The managed-agent deployment mode provides audit logs and per-tool permissions; the plugin mode is closer to a member of staff using a tool. Which mode you choose is a controls decision, not a convenience one.
Client data handling
Client financial records carry confidentiality obligations that predate any of this. Where the data goes, how long it is retained, and whether anything is used for training are questions your engagement letters and your professional indemnity insurer will both have views on. Answer them before the pilot, not after.
Should we build this ourselves or hire someone?
Try the templates against your own practice first — not a client’s — on a month you have already closed, so you can compare output against a known-correct answer. That costs a few days and tells you more than any proposal.
Bring in help when you hit the three things that are not template work: reaching a practice management or ledger system that has no usable API, encoding per-client conventions at scale, and building the evidence trail your regulator would accept. Those are engineering and controls problems rather than accounting ones, and they are where practices stall.
The professional risk, stated plainly
The output carries your firm’s name. A misstatement prepared by an agent is a misstatement by your practice, and neither the model provider nor the person who configured it will be the one answering to your client or your professional body. That is not an argument against doing it — it is an argument for a real review threshold, a reconstructable audit trail, and running in parallel for a full cycle before anyone’s workload changes.
Practices that get this right will not have replaced juniors. They will have stopped paying juniors to do the seventy percent that was always mechanical, and started using them on the exceptions, which is where they learn anyway.
Will this replace our junior staff?
Not in the way the question implies, and practices treating it as a headcount exercise are the ones most likely to end up rehiring. The mechanical seventy percent of preparation work is what a template handles. The remaining thirty — the client whose records make no sense, the balance that will not reconcile, the judgement about whether something is capital or revenue — is both where the cost sits and where juniors learn the profession.
Practices getting real value are changing what a junior’s day looks like rather than how many juniors they have: less keying and ticking, more time on exceptions, earlier exposure to the work that actually requires an accountant. That is a better training pipeline than the one it replaces, which is worth something in a profession with a recruitment problem.
What does it cost to run, not just to build?
Three things, and only one is the licence. Platform cost, which is predictable and usually the smallest line. Review time, which is real ongoing work and should be designed down deliberately rather than discovered. And maintenance — when a model version changes or a client switches bookkeeping systems, someone has to notice and adjust.
Budget for a named owner. A practice that deploys this and assigns it to nobody will find, eight months later, that a partner quietly stopped trusting the output and went back to doing it by hand while the licence kept billing.
What about client confidentiality and our professional body?
Client financial records carry confidentiality obligations that predate all of this, and your engagement letters may not contemplate a third-party tool touching them. Settle two questions before a pilot rather than after: whether your engagement letters permit it, and whether your professional indemnity insurer has a view. Both are quick to ask and awkward to answer retrospectively.
Establish where data is processed and how long it is retained. Retention cuts both ways — you want enough to reconstruct how a figure was arrived at if it is ever queried, and not so much that client records sit indefinitely with a supplier.
Where should a practice start?
On your own numbers, not a client’s. Take a month you have already closed, run the ready-made workflow against it, and compare the output to the answer you know is correct. That costs a few days, risks nothing, and tells you more than any vendor demonstration.
Then pick one client with clean records and one with messy ones, because the gap between those two results is the actual scope of the project. Practices that pilot only on the tidy client consistently underestimate what the rollout involves.