AI Client Query Handling

Only for a narrow, factual class of query, and only if you switch that on — the default is that a person approves everything. The queries that can safely go straight out are the ones where the answer is a lookup, not a judgement: has invoice 4471 been paid, what date was the payment run, what is the outstanding balance on this account, have you received the receipts I sent. There is no professional opinion in any of those, and the answer either matches the file or it does not. Everything else — anything touching deductibility, structure, tax position, an amount a client is disputing, or a client who sounds unhappy — is drafted and queued for a person to review, edit and send. Plenty of firms run the whole thing in review-everything mode for the first few months and then relax specific categories once they have watched it work. That is a sensible way to do it.

A hard classification boundary, plus the design decision that when it is uncertain it escalates rather than attempts. Advice queries — is this deductible, should I buy the vehicle before June, what happens if I take money out of the company, will this trigger GST — are routed to a human, always, without a draft attached, because a well-written draft of a wrong answer is genuinely dangerous: it is the fastest way to get a busy person to approve something they would never have said themselves. This is not just liability management. Advice depends on facts about the client that are not in the ledger and on judgement that is the actual product your firm sells. An AI that answered those questions would be doing the one part of the job you cannot afford to have done badly.

It works from the client’s own file — the ledger, invoices, payments, documents and the correspondence history for that client. When a client asks whether an invoice has been paid, it looks in their file and answers from what is there, citing the transaction. It does not know things that are not recorded, and it does not infer: if a client asks about a conversation someone had on the phone last Tuesday, that is not in the file and the query goes to a person. Each client is a strictly isolated context — one client’s information cannot surface in another client’s response, which is not a nice-to-have when you are handling other people’s financial affairs under the Privacy Act 1988.

They should be able to, and we think you should let them. Trying to pass automated responses off as personal correspondence from a partner is a bad trade: the upside is cosmetic, the downside is a client who finds out and now wonders what else the firm is quietly doing with their affairs. The firms that handle this well are simply matter-of-fact about it — routine status queries get an immediate accurate answer, anything of substance gets a person, and clients are told which is which. In practice most clients care far more about getting a correct answer today than about who typed it. The ones who want a conversation want it about their tax position, and that was never going to be automated anyway.

It escalates immediately, and it does not draft. Tone is a signal the AI is deliberately conservative about: a client who is upset, disputing a fee, threatening to leave, or raising something that sounds like a mistake by the firm goes straight to a named human with the history attached and a flag on it. The failure mode we are avoiding is the automated system that responds to a serious complaint with a cheerful acknowledgement — which converts a solvable problem into a lost client. The same applies to anything that sounds like it might be a professional issue: those need a partner’s eyes within the hour, not a queue position.

The cost is rarely the volume — it is the interruption pattern and the deferral. Most client emails to a firm are short, factual and completely routine, and each one costs far more than the thirty seconds it takes to answer: it arrives while someone is mid-way through a reconciliation, it needs the file opened to check, and it competes with billable work, so it gets left. Then it gets left again. Then the client follows up, slightly less warmly, and now it costs an apology as well. The value is that the routine ones get an accurate answer immediately without breaking anyone’s concentration, and the ones that need a person arrive already sorted and with context attached — instead of buried in an inbox between fourteen status queries.