Conversational AI agent for excursion sales at Southern Spirit

Southern Spirit Lago Argentino is a concessionaire of Los Glaciares National Park and operates cruises to the Perito Moreno glacier. Its audience comes from all over the world, asks at any hour and in several languages, and decides weeks in advance.

In travel an automated agent cannot improvise: a booking taken wrong is not a data error, it is a passenger stranded 80 kilometers from El Calafate.

The business has conditions that shift and do not tolerate approximations.

• Prices that move: the fare depends on the season and the passenger category, with discounts that require documentation.
• A calendar with gaps: there are months with no operation and specific dates with no departure, so offering the wrong date means selling something that does not exist.
• Physical restrictions: the port has accessibility conditions that must be flagged before payment, not after.

An assistant that fills gaps with intuition, in this business, is not an annoyance: it is a promise the operation cannot keep.

We designed, trained and deployed an agent that behaves as a sales advisor, on a closed knowledge architecture where anything not written down does not get answered.

The difference between a chatbot that replies and an advisor that sells without getting it wrong is not the model but what it is forbidden to do.

A closed knowledge base

• One general company document and one document per product, plus a prompt that governs behavior.
• The agent can only use information in that base, and is forbidden from inventing availability, prices or conditions.
• Faced with a gap it answers that the matter must be checked internally, rather than filling in with intuition.
• It only shares links from the client's domain.

Commercial behavior, not just conversation

• It recommends actively based on traveler context and treats someone in a hurry differently from someone hesitating.
• It validates date availability before asking for personal data, and never offers past dates or months without operation.
• It captures the data the operator needs per product and generates the booking link to the site.

When it stops answering

• It hands off to a human advisor in four cases: when it cannot answer with certainty, when the user asks, on conflict or frustration, and on changes, cancellations, refunds or discounts.
• It declines activities the company does not operate, does not opine on matters outside the business and makes no promises.

Architecture and technology

• Conversational agent with per conversation memory, on WhatsApp and the website.
• Service in Spanish, English and Portuguese, with scripted answers for edge cases.
• Integration with the site booking engine to generate the purchase link.

How generative AI is integrated and governed

• The solution integrates generative artificial intelligence services: language models that write, converse, extract and structure inside the workflow, rather than a rules engine with prewritten answers.
• Consumption is measured and billed in tokens, separating input, output and cache write and read, which are different line items with different prices and are accounted for separately.
• Telemetry attributes consumption per agent and per client, so the token budget is bounded and controlled per project instead of being read as a single invoice.
• Limits live in the record, not in the model: a record declares roles and caps, and which provider each role maps to is configuration that changes without touching the agent.
• The architecture is multi provider: the provider is fixed per agent and per session, never per request, and automatic retry against a different provider is deliberately switched off. If the provider fails, the agent stops and says so instead of silently continuing on another one.
• Context caching is the main cost lever, because rereading context accounts for most of the token volume in any agent operation. Sustaining the cache is what keeps the same task from costing several times more.
• And the minimum context window is declared per agent: a provider that does not cover it is ruled out by architecture, not by price.

Learn more about artificial intelligence agents.

The agent went live after six iterations and two validation rounds in which the commercial team tested it as if they were passengers.

Starting point

• Commercial service concentrated in people, with inquiry volume split across channels.
• No coverage outside office hours for an audience in other time zones.

Results

• Six successive iterations of the knowledge base and prompt before going live.
• Two formal validation rounds with the commercial team, with every deviation documented and resolved.
• Three service languages and two simultaneous channels, WhatsApp and website.
• Four explicit handoff cases to a human advisor.

Strategic impact

• The company serves three languages around the clock without adding people to the commercial team.
• The risk of a mistaken booking is bounded by design rather than by supervision.