What should you automate—and what should stay human?

Start with one ordinary request

Imagine a community workshop receiving a message: ‘Can three of us come on Saturday? One person has not tried this before. What should we bring?’ Someone needs to check the session, understand the question and reply. This is a fictional workflow for thinking about automation; no messages are being sent or processed here.

Calling the whole job ‘answer enquiries’ hides several different decisions. Some are routine, such as noticing a missing date. Others involve interpretation, such as understanding what the writer is worried about. Still others create a commitment, such as promising three places.

Splitting the job into steps makes the opportunity clearer. You can remove repetitive preparation while keeping the important choices understandable.

1. 01

   Check

   Use rules to flag missing or inconsistent details.

2. 02

   Prepare

   Organize the question and suggest a reply draft.

3. 03

   Review

   Check availability, wording and any promise.

4. 04

   Act

   Send or update only within an agreed authority.

Use rules when the answer is well defined

If a request form has no date, ordinary software can flag that field. If a number of attendees must be positive, a clear rule can check it. Those steps do not need an AI model to guess what is acceptable.

A rule is useful because its condition and result can be stated plainly: when the date is missing, ask for the date. Before automating it, agree what should happen if the information is incomplete or contradictory. An exception needs a route to someone who can resolve it.

Simple arithmetic may also help prepare the work. It cannot confirm that the workshop actually has three places unless the availability record is current and the booking process prevents competing reservations.

Use AI to help with varied information

The same question can arrive in many forms. AI might suggest that a message concerns a booking, summarize the questions or prepare wording from approved session information. Here the useful role is assistance with language, rather than authority to make up an answer.

Microsoft’s human-AI interaction research discusses the uncertainty of automated inferences and the importance of interfaces people can understand and control. A preparation tool should make clear what it used, what it is proposing and how the person can correct it. [1]

For the workshop, showing the original message beside the summary lets a reviewer catch an omitted question. Showing the session details beside the reply helps them check the date and instructions. A polished draft should not conceal the evidence behind it.

Keep the promise in view

Sending ‘you have three confirmed places’ changes what the recipient expects. In this example, someone must check the actual booking situation and decide whether that promise is justified. Preparing the sentence and authorizing it are different steps.

The person reviewing also needs practical choices: edit the draft, ask another question, hand the request to someone else or stop. A review button is not useful if the reviewer cannot see what they are approving or lacks the information to judge it.

Some routine actions may later be suitable for carefully bounded automation. Define the allowed cases, the record being used, the exceptions and the way to recover from mistakes first. Responsibility should be explicit even when a person does not click through every instance.

Walk through the workflow before connecting anything

Try an illustrative paper exercise with three invented messages: a clear booking question, a request with no date, and a question the session notes do not answer. For each, write what software could prepare, what a person must check and what should happen next.

GOV.UK’s research guidance emphasizes understanding how people currently do the task and treating unsupported suggestions as assumptions to test. Apply that approach here: check whether your proposed preparation actually helps a willing reviewer, rather than assuming every automated step saves effort. [2]

Include corrections, interruptions and review time when judging the idea. A small improvement that reliably removes an awkward step can be useful. Automating a confusing process from start to finish may simply move the confusion somewhere harder to see.