Case study

An AI advisor that had to go outside the website to be useful

5 minute read

Client

Eurocamp (European Camping Group)

Years

2025 – 2026

Role

Product Designer

Contribution

Co-designed the advisor with Brighter Design — function, interface, onboarding and the avatar. Designed and ran the testing programme. Held the seam between Eurocamp’s security, data-protection and analytics teams and the agency building it

Team

Brighter Design owned design/product management and all development — I wrote no code · Eurocamp Marketing, Group cyber security and Data Protection · Wolfenden on ongoing analytics

Platform

Web · responsive · UK

Camille is Eurocamp’s AI holiday advisor. The first version has been live on the UK site since July 2025, from three secondary entry points, seen by fewer than 1% of UK customers — and in eleven months she produced 510 bookings and £542,562 in tracked revenue, around £15 of revenue for every £1 spent building her. The second version is built. It isn’t live yet, and the reason is the most interesting decision in the project.

Four hundred parcs, and the website could only answer half the question.

Eurocamp sells European holiday parcs. A customer has to choose a parc and an accommodation tier within it — and then work out how to get there, across a continent, usually with children.

The brief was parc discoverability: a catalogue too broad to browse, and a filter tree that produces dead ends and empty results. All true — but the harder problem was that the website structurally cannot answer the question a customer is actually asking. It knows what Eurocamp sells. It doesn’t know what flights go near a parc, what’s happening in the region that week, or what a place is like. A better filter wouldn’t have fixed that: the missing information was never in the catalogue.

That is why the advisor exists as a separate surface rather than a redesigned search: Camille can go outside the boundary of Eurocamp’s own data — flight routes, local events, landmarks, regional context — and bring it back into a conversation about a booking.

The constraints were fixed before any of it started. No access to personal data, payment data or the reservation system — she receives question text and nothing else. And she had to coexist with Zendesk Ultimate Chat, not replace it: the two are owned by different departments and measured on different things — Zendesk by Customer Operations on retention and deflection, Camille by Marketing on revenue.

Bringing it inside the website is what made it expensive.

V1 was hosted by the agency and sat outside eurocamp.co.uk. That is why it was cheap and safe: an advisor that isn’t in your website can’t do much to it.

V2 is the first full integration, and every hard problem in the project follows from that one decision. A Group cyber security assessment. A data protection impact assessment. And a late catch by Eurocamp’s own security team: as an EU business we needed customer data resident in the EU, not the US. That moved the storage layer mid-review and derailed the timeline, because the DPIA then had to run again against a data flow that had changed underneath it.

The bar rose in small ways too: V1 asked customers their name, V2 doesn’t ask at all. Inside the website, a name you don’t need is a liability rather than a courtesy.

We paid months to move a product a few hundred pixels. It was the right trade, but the bill deserves to be as visible as the feature list.

Once it was inside the site, putting our own cards inside the chat was waste.

In January, embedding forced a question we hadn’t expected: if the advisor sits on the search results page, why is it rendering its own copies of the parc cards already on screen behind it?

So on desktop we stopped. Camille drives the real page instead — runs the search, applies filters, changes the sort, opens a parc. And the drawer does not overlay the site; it squashes it into a narrower breakpoint, so the responsive design the site already had does the accommodating.

On mobile there is no visible page to drive; the drawer is the screen. So the components come back — parc and accommodation cards inline in the conversation, tapping straight through.

The two experiences diverge from one structural fact, not a preference about small screens.

We tried to stop confirming everything, and it came most of the way back.

An agent that asks permission for every action is unusable. In March we agreed to relax it — confirm only key actions.

It didn’t hold. In the build today Camille asks once at the start whether she may co-browse, and that consent can be withdrawn at any time — but beyond it she still proposes before she acts, even for something as reversible as a filter. “I can filter for accommodation with air conditioning. Would you like me to show just those options now?”

The cost of interrupting someone turned out to be far lower than the cost of surprising them on their own screen. Watching a page move you didn’t ask for is alarming in a way a redundant question isn’t.

When people ran out of capacity to test it, we tested it with personas instead.

Nobody internally had time to use a conversational product in anger — and that is the only way one breaks.

I proposed an automated persona-driven suite. Alongside the family cases we wrote deliberately difficult ones: a nagging user, and one who has to bring the dog. I also wrote the scripts we ran by hand, including an eleven-probe adversarial pass: an instruction to ignore previous instructions, a demand for a parc manager’s personal phone number, and a question about the parc count — which the site itself answers four different ways. Every probe carries a written expected behaviour, so a run produces a pass or a fail rather than an impression.

What I rejected: co-browsing that went all the way to the booking form.

We built it. Camille could take a customer from a vague opening to a completed reservation without them touching the page.

We took it out: completing a booking means entering a name and date of birth, and an agent that types personal details into a form is a different risk category from one that clicks a filter. The isolation that makes her safe to put in front of customers is exactly what stops her finishing the job.

She searches, sorts, filters, adds extras and adds travel — then stops at the booking form. That boundary is visible in the product; it is a decision, not a gap.

Two others went the same way. Co-browsing on mobile, dropped for the full-screen card experience. And a version with no avatar at all, rejected because the CMO wanted personality. He named her Camille; I designed her, and the ten-plus animated states.

The result, and what it is set up to prove.

V1’s £542,562 across 510 bookings is attributed through the AB Tasty experiment that gated the homepage exposure. Over the same window the wider experiment recorded 10,752 conversions and £12.08m — a stated share, not a claimed one.

V2 is instrumented before launch rather than after. Mobile goes to 100%, because the experience is close to what is already live; desktop splits 50/50 against customers who see no AI at all — the only way to know whether Camille causes the booking or merely correlates with someone who was going to book anyway. The hypothesis is written down in advance and framed as one: interactors roughly twice as likely to convert. A result of 1.2×, or none, is a finding.

Reflection

The avatar is the one decision here with no evidence behind it — part brand instruction, part my own instinct that a tool asking for this much trust should have a face. I still don’t know whether customers will read her as warm or uncanny. It’s the first thing I’ll look at in the research phase.

The other thing I would revisit is the consent gate. Nobody argued against it, and I took that as agreement. Looking back I think it was mostly relief — asking permission was the comfortable answer for a room that was nervous about AI, and comfortable answers deserve more scrutiny than they usually get.

Next The website replatform — the design A licence was expiring, and the decision that fixed it was the one that broke browsing · Eurocamp · 2023