Illustrative training scenario. Names, cities, experience and figures in this story illustrate the method. This is not a verified case study or a confirmed result of an XO network member.

AI can speed up tour research, but it doesn't turn an unverified answer into fact. For a travel agent, the value is not in getting a ready-made recommendation in a minute, but in quickly parsing a request, compiling a list of questions, structuring sources and preparing clear comparisons - while maintaining responsibility for every promise to the client.

Without a working protocol, speed creates new risks: the model mixes dates, confidently fills in spaces, does not see tariff changes and can reproduce personal data from an overly detailed request. Therefore, AI should occupy a limited space between the brief and human review, and not replace the supplier, official source or agent's professional judgment.

Illustrative scenario: Olga: a quick draft almost became a ready-made answer

Consider this illustrative training scenario: Olga, a travel advisor from Dnipro, has worked in tourism for four years and likes turning messy messages into clear shortlists. A client asked for a complex family itinerary with a flight connection, late arrival and special room requirements. Olga asked AI for a verification plan and received a convincing answer containing several specific claims.

Previously, she could have moved them into the presentation. But one statement did not match the actual description of the supplier, and another did not contain a source at all. Olga stopped her preparation, divided the answer into questions and facts, and re-checked the tariff, transfer time and room configuration using the available primary data.

After that, she created her own protocol: an anonymised brief, a separate package of sources, a verification table and a final review by a colleague for complex routes. AI still saves her time, but the client only receives information that has a source, a review date, and an understandable level of confidence.

AI helps me quickly see what else needs to be asked. But the phrase reaches the client only after I can show where it came from and what date it refers to.

Olga — character in the illustrative training scenario

Divide tasks into three zones

  • Green zone: structure of the brief, list of questions, wording options, brief summary of your own notes.
  • Yellow zone: comparison of options and identification of contradictions - only on the data transmitted by the agent and with mandatory double-checking.
  • Red zone: visa decisions, medical findings, guarantee of entry, real-time availability and price, legal interpretation of the contract, payments and final promise to the client.
  • Any task goes into the red zone if an error could disrupt a trip, lead to the loss of money or the disclosure of sensitive data.

This division is more useful than the general rule of “check everything.” The agent understands in advance where the AI ​​creates only a draft, where a source is needed, and where the tool should not make a decision at all. Boundaries should be established for the entire team so that quality does not depend on the personal caution of a particular person.

Start with an anonymised working brief

  • The purpose of the trip and the desired result of the client are described without the full name.
  • Dates, duration, group composition and age categories are indicated only to the extent necessary.
  • The budget is given as a range if the exact amount is not needed for the task.
  • Preferences and deal-breakers are separated from already confirmed conditions.
  • Passport numbers, telephone numbers, addresses, payment information and documents have been deleted.
  • No medical details are shared; instead, a functional need is formulated if appropriate and agreed upon.
  • The question to the AI ​​is formulated as assistance in analysis, and not as a request to guarantee the result.

The principle of minimization is simple: transfer only those things without which the tool cannot perform a specific task. If you need to make a list of questions for a hotel, the tourist’s name and a copy of the passport do not help. If you want to compare three rate descriptions, insert the descriptions themselves and the dates they were received, not the entire customer chat.

Collect a package of sources before requesting

  1. Determine what information needs to be relevant specifically for your travel dates.
  2. Save exact excerpts from the contract, tariff, supplier account or official page.
  3. For each fragment, indicate the source and time of review.
  4. Separate the primary source from the review, old publication, and tourist review.
  5. Mark the gaps: what needs to be requested from the supplier or confirmed manually.

AI works better not with the request “tell me everything about the hotel,” but with a limited set of materials and a clear task: highlight differences, find contradictions, formulate questions. Even then, the result is not a new source. It only helps to more carefully process the data that the agent has already collected.

Query formula that's easy to check

  • Context: what client choice is being prepared and at what stage.
  • Input data: only checked fragments with source labels.
  • Task: a specific action - to group, compare, find gaps or simplify the language.
  • Restrictions: do not add facts, do not draw conclusions if there is insufficient data, mark contradictions.
  • Format: table, list of questions or draft message with a separate “requires review” block.
  • Readiness criterion: each significant conclusion is associated with a transmitted source.

Use only materials A, B and C. Do not supplement them with knowledge from outside. For each difference, indicate the source; if there is no data or they contradict each other, write “need to confirm” and formulate a question to the supplier.

Example of a safe instruction

Maintain a fact-checking matrix

  • The statement is worded as the client will see it.
  • The primary or most authoritative source available is cited.
  • The date and time of the inspection are recorded.
  • Compliance with the exact product, tariff, dates and composition of tourists has been verified.
  • There is a status: confirmed, requires clarification, contradictory or outdated.
  • A person is assigned to follow up the question with the supplier.
  • After the answer, the entire linked recommendation was updated, not just one line.

Plausible details are especially dangerous: distance “five minutes on foot”, guaranteed view from the window, round-the-clock service, free option or transfer rule. If this influences the customer's decision, it must be confirmed for the specific booking. The absence of a clause in a beautiful answer does not make it unimportant.

Red flags in AI response

  • An exact number or rule without connection with the transmitted source.
  • Categorical words “guaranteed”, “always”, “no problem”.
  • A mix of similar hotels, airports, terminals or fares.
  • A link that cannot be opened or found in the original material.
  • Conclusion about entry, health, insurance coverage or return without competent confirmation.
  • Price or availability presented as current without live verification.
  • Too smooth an answer where the sources clearly differ.

Check recommendations in two passes

  1. Walk through the facts: dates, names, tariffs, included services, restrictions, deadlines and amounts.
  2. Decision walk: does the conclusion correspond to the original brief, are there any hidden trade-offs and deal-breakers.
  3. Passing through the language: are there any promises stronger than confirmation of the source.
  4. For a complex or expensive route, a second person checks key dependencies.

Human checking is not the final spelling test. The agent must retrace the customer's journey: understand why the option was chosen, see the alternative, assess the risk, and ensure that the unknown is called the unknown. If a recommendation can't be explained without reference to "the AI ​​said so," it's not ready.

How to use AI in a message to a client

The tool can make the explanation shorter, calmer and clearer, but the final text is checked and sent by an agent. Useful format: client need, two or three comparable options, confirmed differences, open questions, date of relevance of data and next step. Don't mask uncertainty with persuasive style.

  • All facts coincide with the verification matrix.
  • Price and availability are marked with time of relevance.
  • The trade-offs are shown next to the benefits.
  • Unknown parameters are listed separately.
  • There are no internal notes, service instructions or personal data of other clients.
  • The next step is realistic and does not create false urgency.

What to save in CRM

  • Initial anonymized brief or link to a secure client card.
  • List of sources used and dates of review.
  • Key confirmations and open questions.
  • The version of the final recommendation that the client received.
  • Who performed human verification.
  • What changed after the supplier responded or the tariff was updated.

There is no need to store every trial formulation. What is important is the auditable path from the need to the confirmed conclusion. It allows you to quickly update the selection, transfer the case to a colleague, and explain the decision if the supplier’s data has changed.

Metrics that show benefit rather than novelty effect

  • Average time from brief to first verified selection.
  • The share of statements returned for clarification due to the lack of a source.
  • Number of corrections after final review.
  • Number of customer issues arising from unclear comparisons.
  • Proportion of complex routes that were verified by a second person.
  • Incidents involving outdated or unconfirmed information.

Implementation plan for seven working days

  1. Day 1: Choose two safe tasks from the green zone.
  2. Day 2: Create an anonymized brief template.
  3. Day 3: Set up the source packet and fact matrix format.
  4. Day 4: Test the protocol on a closed old case.
  5. Day 5: Compare the result with real documents and find errors.
  6. Day 6: Agree on red zones and second check rule.
  7. Day 7: Apply the process to one current query and parse the result.

Final check before sending the recommendation

  • The request to the AI ​​did not contain unnecessary personal or payment data.
  • Every material fact has a source and a date.
  • Live price and availability checked in the working system.
  • Contradictions are not hidden and are submitted to the person responsible for clarification.
  • The final conclusion is made by the agent and corresponds to the brief.
  • The client sees the restrictions, expiration date and next step.

Safe work with AI is not built on prohibition or blind trust. Limit the task, minimize the data, give the tool verifiable sources, separate the draft from the fact, and leave the decision to the agent. Then the speed really helps the service, and does not create a new hidden place for an error.