Start with one recognisable point in the journey. An enquiry has arrived; what must a member of staff decide before useful work can begin?

They may need to identify the type of request, check whether enough information is present, find an existing customer or project record, choose an owner and agree the appropriate response time. These are related decisions, but they do not all carry the same risk. A suggestion that a message concerns “support” is different from a decision that a complaint is low priority.

Write down the current route before choosing any AI tool. Include website forms, shared inboxes, direct email, phone notes and referrals from colleagues. Note where people copy information, which judgement needs experience and how unusual or sensitive enquiries are escalated.

Axisware’s approach to workflow automation starts with these responsibilities and hand-offs. The aim is to improve the journey people already need to complete, not to add an impressive feature without a clear owner.

“Understand this enquiry” is too vague for a dependable trial. Give the AI assistant a limited output that a person can inspect quickly.

For example, a draft review card could contain:

• the apparent subject of the enquiry;
• a short summary using only the supplied information;
• the product, service or existing work it may relate to;
• details that appear to be missing;
• any date or requested next step stated by the enquirer;
• a suggested team or role for review;
• the source message and attachments for comparison; and
• a clear indication when the assistant is uncertain.

This output does not become the business record automatically. It is a proposed starting point. The reviewer should be able to correct the subject, change the owner, reject the suggestion and see the original enquiry without searching another system.

The structure matters more than polished prose. A concise card that exposes uncertainty is more useful than a confident paragraph that hides where it has guessed.

Incoming enquiries can contain names, contact details, commercial information and sometimes material the organisation did not expect to receive. Do not copy a live inbox into a public AI service simply because a demonstration is easy to arrange.

Choose the information boundary before the trial. Use fictional, redacted or otherwise approved examples where possible. Decide which fields are required for the proposed output and which should remain outside the AI step. Check the provider’s terms, privacy arrangements, retention controls and access settings with the people responsible for information governance.

The National Cyber Security Centre advises organisations not to include sensitive information in queries to public large language models and to understand how a provider manages submitted data. Its guidance on ChatGPT and large language models is a useful reminder that convenience does not remove the need for an information-handling decision.

The Information Commissioner’s Office also recommends a risk-based approach when AI processes personal information. Its AI and data protection guidance was marked as under review when checked on 12 October 2026, so organisations should consult the current guidance and their own advisers rather than treating a blog article as compliance advice.

AI can propose that an enquiry looks like a sales question, a support request or a supplier message. A person should decide whether that label is correct and what the organisation will do with it.

Make the boundary visible in the design. Use wording such as “suggested route” rather than “assigned team” until a reviewer confirms it. Record who accepted or changed the suggestion. If a response target, commercial commitment or safeguarding concern is involved, keep the relevant authorised person in control.

The same principle applies to a draft reply. AI may prepare wording from approved information, but a named member of staff should check the facts, tone, recipient and commitment before anything is sent. The review should be a real step in the workflow, not a small button people are expected to press without reading.

Axisware can add practical AI assistance where it supports a real workflow and appropriate human review remains in place. That does not make AI suitable for every enquiry or every decision within the journey.

The tidy enquiry is not the best test of the workflow. Try examples with a vague subject line, an attachment, several questions, a request that belongs to two teams and a message that refers to earlier contact without including the reference.

Ask what the assistant should do when it cannot find enough evidence. A safe response may be to mark the item as uncertain, show the missing detail and place it in a general review queue. It should not invent a customer record, infer a sensitive characteristic or choose an answer merely to complete every field.

Keep a simple exception list during the trial. Record:

1. what made the enquiry difficult;
2. what the assistant suggested;
3. what the reviewer changed;
4. whether the source information was sufficient; and
5. what should happen if the same situation appears again.

This turns corrections into design evidence. A repeated uncertainty may show that the intake form needs a clearer question, that the routing rules are incomplete or that the AI step should be removed from that case altogether.

A good summary still creates extra work if somebody must copy it into another application. Map what happens after review. The accepted subject, owner and next action may need to update a customer relationship management record, create a task, attach the source message or notify the responsible team.

A well-designed application can help reduce duplicate entry and unnecessary hand-offs. That improvement may come from a straightforward connection, a shared status or a clearer form; AI is only one possible part of the design.

Avoid creating a second inbox that people have to monitor. The reviewer should work in a place that fits the organisation’s responsibilities and permissions. Colleagues who receive the item should be able to see the original source, the reviewed summary and the agreed next action.

This is where bespoke software development can be relevant. A tailored application can connect an approved AI-assisted step to the existing records and roles, while keeping the review point explicit. It can also leave the AI out where a normal rule or human conversation is the better choice.

Some enquiries need only an acknowledgement and a clear next step. Others require product knowledge, a quotation, investigation or a conversation. Do not ask an AI assistant to fill gaps from general knowledge when the organisation has not approved the source.

If draft replies are included in the trial, give the assistant a small set of checked information: the organisation’s preferred tone, current contact routes, standard questions and any wording that has been approved for the relevant service. Require it to distinguish supplied facts from proposed phrasing.

The reviewer then checks whether the draft answers the actual message, avoids unsupported promises and asks only for information that is necessary. They should also be able to discard the draft and write a different response without breaking the process.

Measure correction effort as well as drafting speed. A fluent reply that takes longer to verify than to write is not a useful result. The trial should reveal where assistance genuinely helps and where direct human handling remains simpler.

Every enquiry needs an owner even if the AI service is unavailable. Define the manual route before the trial begins. Staff should know where new messages appear, how to recognise an unprocessed item and how to continue when a suggestion cannot be produced.

Keep proportionate evidence of the review: the source enquiry, the suggestion shown to the reviewer, the accepted or corrected outcome and the next action. Decide how long those records are needed and who may see them. Avoid collecting extra information merely because the technology can analyse it.

Review access when roles change. Test what happens when an enquiry is forwarded, duplicated, withdrawn or linked to the wrong record. A dependable workflow deals with correction as a normal activity rather than assuming every first match will be right.

These controls make improvement possible. The team can identify common corrections, missing routing rules and training needs without treating the AI output as an unquestionable account of what happened.

Choose a bounded enquiry type with a manageable volume and a team that understands the current work. Use an agreed set of examples first, including ordinary and difficult cases. Do not begin by switching every channel or allowing automatic replies.

For each example, compare the proposed review card with the source and record the reviewer’s changes. Then judge the trial using practical measures:

• Did the summary preserve the important facts?
• Did it show missing information and uncertainty clearly?
• Did the suggested route help the reviewer reach the right owner?
• How often did staff need to correct the subject, priority or next step?
• Did the process reduce copying without creating another queue?
• Could staff recover easily when the assistant was unavailable?
• Did enquirers receive an accurate, considerate response through the normal approved route?

Include the time spent checking and correcting the output. If the trial appears helpful, improve the intake, rules and review screen before widening it. If it does not, keep the lessons and simplify the process instead of forcing AI into the workflow.

AI for incoming enquiries is most useful when it helps a person see what has arrived and what needs attention. It should make the source easier to review, not hide it; suggest a route, not declare a decision; and prepare wording, not speak for the organisation without approval.

Start with one enquiry type, one useful review card and one named owner. Define the information boundary, include awkward examples and decide how the work continues when the suggestion is wrong or unavailable. That creates evidence for a sensible decision about whether to integrate, change or stop the trial.

If you would like to discuss a human-reviewed enquiry workflow, contact the Axisware team. We can help map the journey, identify a proportionate AI-assisted step and connect the reviewed result to the work that follows.