Case study · Travel & tourism
Enquiries answered in seconds, not minutes
A cruise and travel retailer’s consultants now answer brand-specific enquiries from a governed knowledge assistant their own team runs.
Headline result
Seconds
to answer a brand-specific enquiry, down from minutes, with the source cited
- Client
- Cruise and travel retailer, two consumer brands, about a dozen consultants
- Engagement
- Knowledge management, with replies drafted in two brand voices
- Stages
- Discover, Design, Deliver
- Where it got to
- Pilot with real consultants on real enquiries. Full rollout is paced to content volume
The situation
The retailer runs two consumer brands, with a retail team of about a dozen consultants answering guest enquiries by phone and email. The knowledge behind those answers lived in a sprawling wiki, in shared documents, and in the heads of the most experienced people.
The same fee rule sat in two places with different wording. Retired procedures looked identical to live ones. Consultants hunted through browser tabs while a guest waited, and newer staff escalated to the senior people the business could least afford to interrupt. Every slow or inconsistent answer put a booking at risk.
The approach
We started by spending almost nothing to find out what would be hard. A throwaway prototype, built in days on an export of the wiki and a sample of catalogue data, spoke fluently in both brand voices and invented facts. That priced the real work, grounding and live data, before any real spend.
About seventy per cent of the effort that followed went into the knowledge itself. The AI was the smaller job.
Discover
A throwaway prototype built in days, then a full audit of the knowledge estate
What the client keepsA priced view of the two hard problems before committing to the build
Design
One-topic documents, each with an owner, a status and a review date. Grounding rules set before content volume
What the client keepsA knowledge structure and governance model that will outlast any single assistant
Deliver
Assistant live in Microsoft Teams for a pilot group. Every answer cited, and an evaluation suite run on every change
What the client keepsA working assistant their team operates, with runbooks that train a new operator in a day
What’s next
Full-team rollout as content grows. Customer-facing chat and voice on the same foundation
What the client keepsThe choice of when to extend, at their own pace, without a rebuild
What it exposed
The first build showed how deep the problem went
The prototype failed in a useful way: perfect tone, invented facts. Retrieval over one monolithic manual can’t be trusted. The first grounded build found more. Rules duplicated with drifted wording, processes nobody had written down, and supplier passwords sitting inside searchable procedures.
So we restructured the knowledge into single-topic documents, made only approved content answerable, and taught the assistant to say “I couldn’t find that” instead of guessing.
Where we drew the line
Written down before anything was built: what the system may own, and what stays with a person.
The system may own
An answer from approved content
Nothing outside the governed library is answerable
The source behind every answer
Cited, so a consultant can check it
A draft reply in either brand voice
Ready for the consultant to send or change
“I couldn’t find that”
Said plainly when the knowledge isn’t there
Only a person may own
What counts as approved
Content is published by the team, at their pace
Each document’s accuracy
A named owner, a status and a review date
What a guest is told
The consultant answers; the assistant informs
When to extend it
Chat and voice for customers, when the content is ready
Results
Consultants answer brand-specific enquiries in seconds, not minutes, with the source cited on every answer.
Their team runs and extends the assistant without us. Content, instructions and even the model are theirs to change.
The same governed foundation is built to serve customer-facing chat and voice without a rebuild.
How we measured this
Adoption rates and before-and-after handling times weren’t baselined, so we don’t report them. The evaluation suite measures answer quality on every change, and this study claims only what that evidence carries.
What transfers
Refuse to scale the assistant past its knowledge
The unglamorous work came first: one topic per document, an owner and a review date on each. It’s why the answers can be trusted. At handover the team had a pilot batch live from a planned library of about seventy documents, and the runbooks to keep publishing on their own.
Contact
Something worth changing?
Tell us what’s slow, expensive or stuck. If we’re not the right people, we’ll say so and point you somewhere better.