Viking Cruises is known for premium service, including a promise that guests can reach a real person without jumping through hoops. Behind that promise, contact center agents work across several systems on calls that run a median of 45 minutes, so Viking set out to bring AI tools into a single pane of glass that gets agents to the right answer faster.

Lovelytics designed one governed Databricks platform and used it to build four agent-facing tools inside Viking’s existing CRM: shore excursion answers, email categorization, call summaries and knowledge base management. Each new tool reused the same foundation, which is what made an eight-week timeline possible.

When you book a Viking cruise, the ship is only part of the experience. The first call to a Viking agent is part of it too. Viking built its reputation on being a premium line, and that includes being easy to reach and getting premium help fast.

For agents, that standard comes down to a few moments that repeat all day. A guest in Barcelona asks whether a shore excursion works with a mobility aid or a pacemaker, and the answer sits in vendor details. A care email asks to change a reservation and needs to reach the right person quickly. A call runs 45 minutes (on average) and its notes have to be clear enough that the next agent never starts from scratch. Each of these sent agents searching across several tools for a single detail

Viking picked four places where AI could take that weight off agents, and asked a direct question: what can be delivered in eight weeks?

Viking and Lovelytics met at the Databricks Data + AI Summit earlier this year, and Viking brought us in to put its four contact center priorities on Databricks and deliver them on a tight timeline. What set the engagement apart was the combination of deep Databricks expertise and a willingness to commit to speed: when Viking asked what could be delivered in eight weeks, the team signed up for all four.

Four Tools, One Platform

Each initiative could have filled a quarter on its own. Instead of building four separate solutions, the team built one platform and made each tool a thin layer on top of it:

Shore Excursion Assist. An AI chat that helps agents check whether an excursion fits a guest’s needs, from mobility to medical considerations. Answers come from curated excursion records and cite their source. When vendor data is missing, the tool says so instead of guessing, and accessibility details are shown exactly as the vendor reported them. The first release covers 200 standardized excursions.

Email Categorization. Recommends a category and subcategory for each inbound care email so agents can get to the request faster. Viking had no labeled email history, so the team worked with business stakeholders to build the category list from the emails themselves. Agents stay in control and can override any suggestion.

Call Summarization. Turns long calls into structured summaries for the agent to review and post to the case. A Whisper speech model served on Databricks handles transcription, and sensitive details are redacted before any summary is generated.

Knowledge Base Manager. Finds topics callers ask about that the knowledge base does not cover yet, flags duplicate or outdated articles, and drafts new content for a human to approve. Only reviewed, published articles are ever used to answer a question. The first release covers about 20 articles migrated from SharePoint.

The tools sit inside the CRM agents already use, so there is nothing new to learn. Databricks supplies the intelligence while the CRM remains the system of record for every case.

Built for Trust and Speed

Contact center calls include names, addresses and payment details, so privacy and accuracy shaped the design from day one. Databricks Unity Catalog governs data access and lineage once for all four tools. Personal and payment information is masked before it reaches a model, using pattern-based rules plus Databricks’ native AI masking.

Every model change has to pass a tested evaluation set, with a 98% accuracy gate for shore excursions and 95% for email categorization.

Two Databricks capabilities did a lot of the heavy lifting. Lakebase, Databricks’ managed Postgres service, handles both conversation history and the structured excursion lookups, which is a new use for it and runs faster than the Unity Catalog tools typically used for that job. MLflow tracing and DSPy prompt optimization let the team tune prompts systematically instead of editing them by hand.

That setup paid off quickly. When Viking asked to try a different language model, the team swapped the model, re-optimized the prompts and reached higher accuracy without rebuilding anything.

Raising the Engineering Bar

The engagement also moved Viking’s AI work onto a production footing. Code that once lived in notebooks now sits in a Git repository, deploys through Databricks Asset Bundles, goes through formal code review and runs on dedicated user acceptance testing endpoints. Those habits carry over to every project that follows.

Clear boundaries make fast delivery possible. Fitting four builds into eight weeks meant scoping each first release carefully, such as 200 excursions and about 20 articles. During testing, agents sometimes asked about excursions outside the 200 because the tool was useful enough to want more of it. The teams walked through the agreed boundary together, and we would now make that boundary visible inside the tool itself, not only in project documents.

No labeled data means building it together. None of the four tools started with a labeled evaluation set, and the call summary had no approved template. Viking’s subject matter experts helped write the test questions, review the email categories and shape the summary format. Open questions, like where the category list should live and which case fields each tool writes to, went into a shared decision log so work could continue while decisions were made.

A short strategy phase pays off. Viking moved quickly from idea to build, which is what got the tools in front of agents this fast. A brief up-front step to rank the four use cases by expected value would give both teams a shared yardstick for judging results, and we now recommend pairing it with builds like this one.

In eight weeks, three AI tools reached user acceptance testing on a shared, governed platform, and the fourth, Knowledge Base Manager, continues under an extension through December 11.

Because testing is still underway, impact on call handling time will be measured once the tools roll out to agents.

• 3 of 4 contact center AI tools delivered to user testing in 8 weeks, with the fourth in progress
• 1.45 sec measured time to first answer for shore excursion questions, against a target of under 3 seconds
• 98% and 95% accuracy gates in place for shore excursions and email categorization before any model is promoted
• Faster model changes: swapped the language model and improved accuracy without rebuilding the solution
• Production-ready practices: code moved from notebooks to version control, automated deployment, code review and dedicated testing endpoints

Viking and Lovelytics are already working on the next phase. A machine learning project for price optimization kicked off this week, and a second ML project focused on river depth for river cruises is under discussion. Over time, Viking has an opportunity to bring these efforts together with the contact center tools into a customer 360 for its agents.