The most expensive habit a firm can adopt as it rolls out AI is the assumption that every question can now be answered without picking up the phone. The opposite is closer to the truth in many situations. A well-timed five-minute client conversation often does more, faster and more cheaply, than any AI tool - and the firms getting the most out of AI are the ones that have got more deliberate about when to call the client, not less.
This piece is a working partner's view of where the conversation still wins. It is not anti-AI. It is an argument that the productivity gain from AI should be reinvested into more human contact, not used to justify less.
When the question is "what did you actually mean"
A surprising share of accounting problems trace back to a client instruction that was ambiguous when it was first given. The team works for a week on one interpretation, the partner reviews, the conclusion is wrong, and the work has to be redone. The five-minute call up front to confirm "when you said you wanted to extract value from the company, did you mean a dividend, a salary, a capital reduction, or a buy-back?" saves the week of misdirected effort.
No AI tool can perform this disambiguation, because the AI only sees what is on the page. The unspoken context - the client's actual intent, their family situation, the timing pressures they have not mentioned - is in the conversation, not in the file.
When the answer depends on something the client has not told you
Most non-trivial tax and accounting questions depend on facts that are not in the documents on file. Whether a transaction was at arm's length, whether the client has a connected party in another jurisdiction, whether there is a planned exit in the next 18 months, whether the spouse has used their CGT allowance - these are facts the AI cannot infer because they are not in the data.
A junior who asks the AI for the answer in isolation gets a confidently wrong answer based on the assumption that the missing facts are not relevant. A junior who picks up the phone and asks the client gets the facts and then either asks the AI a better question or works through the answer themselves. The second is faster and right; the first is fast and wrong.
When the client is unhappy and does not want to say so in writing
A client who is starting to feel poorly served often gives the first warning in a slightly cooler email, a slower response to a request, or a chasing message they would not normally send. A junior who reads the email at face value misses the signal; a partner who reads it and picks up the phone catches the relationship before it deteriorates. AI tools can be trained to flag tone, but they cannot have the conversation, and the conversation is where the recovery happens.
This is one of the highest-leverage uses of senior time in a practice. The cost of replacing a £20,000-fee client is several times the cost of a five-minute call. Firms that automate away the relationship contact in pursuit of efficiency end up paying for it in churn.
When the client is about to do something they will regret
Clients sometimes share a plan in an email that is, on closer questioning, going to cause them a tax, regulatory, or commercial problem they have not seen. A response that addresses the email at face value gives them the comfort to proceed. A response that picks up the phone and walks them through the implications often stops the problem before it happens.
AI can surface the issue in a draft response, but the conversation is the medium that gets the client to actually re-think. A written response - especially one that arrives quickly and looks comprehensive - can have the perverse effect of confirming the client's plan rather than challenging it.
When you need to negotiate a deadline
A regulator-imposed deadline cannot be negotiated; a client-imposed deadline often can. The five-minute call to the client to explain that the year-end pack is incomplete and to agree a realistic revised timetable saves the team a fortnight of working flat out against a date that was always going to slip. No email negotiation works as well as the conversation.
When the AI is not refusing but should be
A subtle case is when the AI gives a confident answer to a question that should really have been answered "I don't know, ask the client". The model has filled in the blanks reasonably from the documents it has, but it does not know what it does not know. The reviewer who notices the answer feels too smooth, picks up the phone, and asks the client the underlying question, often finds that the AI's smooth answer was wrong in a specific way.
The discipline is to treat a too-smooth AI answer as a prompt for a call, not as a confirmation that the question is settled. The model's confidence is not a measure of its accuracy.
When the client has multiple advisors
When a client has a solicitor, a corporate finance advisor, and an IFA in addition to the accountant, the coordination work cannot be done by AI. The 30-minute call with the solicitor to agree the structure of a transaction, the 10-minute call with the IFA to confirm the pension position, the five-minute call with the corporate finance advisor to align on the timetable, are interactions that do not happen in writing without losing nuance and speed.
A firm that has invested in AI without investing in the relationship time with the other advisors finds itself doing more in-firm work for less coordinated outcomes. The advisor network is a competitive asset and it has to be maintained by conversation.
How to make sure the calls still happen
The risk in any AI rollout is that the convenience of the chat box replaces the friction of the phone call. A few practical disciplines help:
- A weekly review of which clients have not had a partner-grade conversation in the last quarter, with a target to close the gap
- A workflow rule that any client query taking more than 30 minutes of team time is escalated to a five-minute call before further work
- A standing item in the partner board to review the relationship temperature of the top 20 clients by fee
- A training point with juniors that "ask the client" is a valid and often correct answer to a question they cannot resolve from the documents
- A culture where picking up the phone is praised, not treated as a sign of inefficiency
The reinvestment dividend
The right way to think about the time AI saves is that it is not money saved; it is time freed up for the work that AI cannot do. The firms that have got the most commercial value out of AI are not the ones that have used the saved time to cut costs; they are the ones that have used it to have more conversations with clients, advisors, and prospects. That is where the next year's growth comes from.
How Accupe helps
Accupe is built to surface the moments where the conversation is needed. The Smart Boards show which clients are stuck, the Compliance Radar flags the deadlines worth a partner call, the time tracking shows which engagements are over-running, and the AI document analysis with its docs-only mode and source citation does the routine work fast - leaving the partner's time for the calls that actually move the relationship. Per-firm pricing from £20/month.