Skip to content
AlbaNova

Practical Guide · 4 min read

How to automate repetitive client questions without losing the human touch

How to draft answers to the questions clients ask every week, using your own material, and when a person still has to send the reply.

By Chiraag Mutreja · Updated

AI AgentsClient ServiceKnowledge ManagementWorkflow Automation

The questions are not hard. They are the same ones. Where is my document? Can you resend the invoice? What happens next, and who is doing it? They arrive by email, and sometimes through a portal, and a capable person answers them from memory or by searching a folder they have searched before.

The cost is interruptions, not drama. The person who should be doing the work that needs judgement spends the morning on retrieval. Clients do not experience that as efficiency. They experience it as waiting.

It is tempting to put a model on the front of that queue and let it reply. That is the version that goes wrong. The video is the itch. The design below is the constraint: draft from material you trust, in sight of the person who is allowed to answer, and stop when the question is no longer repetitive.

Stop answering the same client questions by hand

Why teams keep searching for answers they have already given, and what should happen before a reply goes out.

Separate retrieval from advice

Most of the repeat questions are not advice. They are status, process and documents. "Have you received the statements?" is a lookup. "Should I claim this?" is not, and should never be answered by a draft that a tired person clicks send on.

Write down the questions the team can recite. Ten is usually enough. For each one, note where the true answer lives: a policy page, the matter file, the practice management system, a template. If the true answer does not live anywhere, the first job is to write it down. Automating a rumour just sends the rumour faster.

A workflow with a person at the end

Classification is where a model earns its place. Free-text questions do not arrive with a dropdown. A model can propose what is being asked, which client it relates to, and whether it looks like one of the known questions or like something else. A rule is better when the client already filled in a form. Use the rule.

The reply, if one is drafted, should be built from the source you named, not from the model's general knowledge of "how firms usually do this". Show the person the source alongside the draft. If the source is thin, or the question mentions money, a complaint, a deadline or a professional judgement, do not draft a final answer. Escalate it, with the classification and the excerpts attached, to the person who owns that client.

Permissions are the part that gets skipped in demos. The workflow may only retrieve what the person handling that client is allowed to see. A shared bot with a folder full of every client is how one client's answer leaks into another's reply. If you cannot express that limit, you are not ready to automate the replies. The readiness questions on data and confidentiality are the right test.

Then keep an audit trail: what was asked, what was retrieved, what the model suggested, what the person changed, what was sent. You will need it the first time a client quotes a sentence back to you, and you will need it to improve the ten answers.

What this is not

This is not a licence for an agent to handle the inbox. The distinction is the one in AI agents versus workflow automation. Understanding a messy question is a model. Fetching the file, creating the task and recording the outcome are workflow. Sending is a person, except for the rare reply that is genuinely a template and has been approved as such.

It is also not the same project as fulfilling a complex inbound request. For one global fund data provider, triage and fulfilment of requests fell from six weeks to four days. That case study is about reading a request, costing it and running a defined fulfilment path. It is neighbouring work. It is not evidence that a chatbot can answer an accountancy client.

Use it where the downside is small

Good candidates: where a document is, whether something has been received, what the next step in a published process is, how to book time, which form to use.

Poor candidates: fees, tax or legal positions, complaints, anything that changes a filing, anything you would not let a new joiner answer on their first day.

Confidence thresholds should be boring and explicit. Above the line, and only for the good candidates, a person gets a draft. Below the line, they get the question, the guessed classification, and no suggested answer to be lazy with. Review a sample every week. The failures are the design feedback, not an embarrassment to hide.

A practical test

  • List the questions that arrived more than once last month.
  • For each, name the source a competent person would open.
  • Mark which replies must be sent by a named role.
  • Confirm the workflow cannot see other clients' files.
  • Decide what gets stored, and for how long.

The service line this sits under is email and request triage. If the team can already recite the questions, bring that list to a free Operations Audit. We will tell you which of them are retrieval, and which only look repetitive until you read the second sentence.

Free Operations Audit

List the ten questions the team can already recite.

Bring those to a free 45-minute Operations Audit. We will separate the ones that can be drafted from source material from the ones that need a person from the first reading.

Book a free Operations Audit

If there's nothing worth automating, we'll tell you.

What happens next

  1. 1Pick a time that suits you
  2. 2Tell us briefly which process hurts most
  3. 3We map it together in 45 minutes
  4. 4Your written summary arrives within 3 working days