Using AI for Customer Enquiries Without Losing the Human Touch

AI, enquiries and guest service

Customer enquiries are one of the clearest places where AI can assist a tourism business—and one of the easiest places to damage trust. The sensible model is not an autonomous chatbot making promises. It is a well-controlled assistant helping a knowledgeable person respond more clearly and consistently.

Why tourism enquiries are not ordinary messages

An enquiry that looks straightforward may contain several decisions. Dates and party size are only the beginning. The customer may be signalling pace, budget, confidence, accessibility needs, family dynamics or a particular reason for travelling. A strong response notices those signals without inventing certainty.

Tourism information also changes. Availability, opening arrangements, journey times and supplier terms cannot safely be assumed from a language model’s general knowledge. The answer must be grounded in current, approved information.

A practical human-in-the-loop workflow

1. Capture the enquiry without losing the original

Keep the customer’s message as the source record. AI can extract a structured brief, but the original should remain visible to the person responding.

A useful brief might include:

• dates, duration and party size;
• starting point, end point and transport assumptions;
• interests and stated priorities;
• budget signals;
• dietary, mobility or accessibility information;
• questions already answered and questions still outstanding;
• anything ambiguous that requires clarification.

2. Retrieve approved information

The drafting process should use material you control: current product descriptions, operating notes, pricing rules, FAQs and supplier information. If the answer is not in an approved source, the tool should flag the gap rather than guess.

3. Prepare a draft—not a final reply

AI can combine the structured brief with approved information and produce a first response. The instruction should be explicit about uncertainty: do not invent availability, prices, inclusions or policies; clearly mark anything requiring confirmation.

4. Review against a fixed checklist

The reviewer should check more than spelling:

• Are names, dates, numbers and locations correct?
• Has the reply answered the customer’s actual question?
• Are availability and price statements properly qualified?
• Does the proposed pace make geographical and operational sense?
• Has sensitive information been handled appropriately?
• Does the reply sound like your business rather than a generic travel company?
• Is there a clear and proportionate next step?

5. Send through the normal business channel

The approved reply should be sent from the usual inbox or customer system, leaving a record that the team can see. Avoid creating a parallel AI process that makes correspondence harder to trace.

6. Learn from corrections

Record recurring changes. If staff repeatedly correct the same tone, fact or formatting problem, update the instructions or source material. If the tool continues to create more checking than value, stop using it for that task.

What can reasonably be automated?

Protecting the human quality

The goal is not to make every response sound identical. Good service often includes a recommendation based on experience, a small piece of reassurance or a sensible challenge to what the customer first requested. Those are signs of judgement.

Give the tool examples of your preferred tone, but do not ask it to imitate friendliness by adding empty enthusiasm. Specificity is more convincing than superlatives. “This avoids a second hotel change” is more useful than “This unforgettable option is perfect for you.”

It is also reasonable to tell customers when they are interacting with an automated system, particularly if it directly handles their messages. Transparency becomes more important as the system takes on more of the interaction.

Data protection and security still apply

Customer emails may contain personal or sensitive information. The NCSC advises organisations not to enter sensitive information into public large-language-model services and to understand how a provider handles queries. The ICO’s AI guidance covers accountability, transparency, lawfulness, accuracy, fairness, security and data minimisation.

Before using live enquiries, decide which tool and account are approved, what information may be entered, whether data should be removed or anonymised, who can access the output and how long it is retained. “We only used it to draft an email” does not remove those responsibilities.

A sensible first trial

1. Select 20–30 past enquiries with no live customer consequence.
2. Remove personal information that is not needed for the test.
3. Define an extraction template and a drafting instruction.
4. Compare the draft with the response your team actually sent.
5. Record factual errors, missed nuances, tone problems and time saved.
6. Only move to live work if the checks are clear and the benefit is real.

This creates evidence from your own business. It is more valuable than a demonstration based on an artificially neat example.