DATA-DRIVEN MARKETING · AI STRATEGY
Volare
An AI plan for a travel agency, built on deciding what not to automate
Client
Role
Strategy, analysis and deck
Year
2026
Team
4
Format
Data-Driven Marketing · POLI.design
Volare is competing in a predictive market with pre-predictive capabilities.
SEVEN PROBLEMS, ONE GAP
The brief described a mid-size Italian travel agency — 800,000 registered users, a HubSpot CRM, a fifteen-person service team — and listed seven phases of the customer journey, each with a weak metric. Open rate 12%. Cart abandonment 68%. Repeat bookings 22%. The task: propose an AI improvement for each.
Seven problems invite seven fixes. I went looking for what they had in common.
Six of the seven symptoms trace back to one gap: Volare records who its customers are, and nothing about what they do.
Its CRM is a well-kept database run on manual rules — the logic of marketing twenty years ago. The market it competes in, Booking and Expedia, runs on predicting individual intent. That gap comes first, because every model downstream would have to learn from behaviour that was never recorded.
WHAT VOLARE CANNOT SEE
Marta signed up in March after clicking a Google ad. Volare knows four things about her: name, email, source, date. She gets the same weekly deals email as 800,000 other people.
What she actually did: searched Lisbon three times in eight days, filtered every search under €600, looked at August for two, and never reached checkout.
Volare can tell she exists. It cannot tell she is planning a specific trip.
The same blindness applied to the analysis. Three of the seven phases have no data that could evidence a problem at all, so I marked them not observed rather than guessing. An early draft of the journey had customer quotes on it; there had been no interviews, so they came out.
The phases Volare cannot see turned out to be exactly the ones the plan has to instrument first.
A HYPOTHESIS, NOT A CONCLUSION
"No behavioural data" explains the pattern. It isn't the only thing that could.
- Deliverability. Low opens could be sender reputation or a stale list, not relevance.
- Price and proposition. Low retention could be the destinations or the prices, not the timing.
- Checkout mechanics. Abandonment could be payment failures, not missing personalisation.
Each rival came with the test that would settle it, and the first module of the plan runs all four in parallel.
If a rival explanation carries more of the problem, the sequence changes. That is what the stage gate is for.
I led with behavioural data for one reason: it is the only explanation that accounts for all four measured gaps at once, and the only one whose fix is a precondition for testing the others.
TWO PLACES NOT TO SPEND
Every phase was compared against a sourced industry benchmark before anything was proposed. Four were behind. Two were not.
- Cart abandonment, 68% — the brief's most alarming number. The online-travel norm is around 89%: customers price-compare across five sites by design. Volare is ahead.
- Response time, 14 hours — against a travel average of about 15. At par on average, though an average hides the 2am disruption case.
Recommending against investment is a finding, not a gap in the analysis.
That doesn't mean doing nothing. Cart recovery still ships, but only after the recoverable share is sized. It means the money goes first where the distance is real: retention, at 22% against a benchmark of 32%.
DOES IT HAVE TO BE AI?
The brief asked for AI in every phase. I tested every proposal against the simpler option first.
- A — a rule. Move the "we miss you" email from six months to ten.
- B — analytics. Segment customers by how often they actually rebook.
- C — a model. Predict each customer's next booking and contact them then.
For retention the answer was start at B, move to C: segments capture most of the gain in weeks, and the model earns its cost once the real variance is known. For service, a knowledge base comes before any chatbot — it is the bot's prerequisite, not its alternative.
A model that cannot beat the existing rule on a holdout should not ship.
Three of the seven phases stay hybrid by design. Two start with no AI at all.
A SCORE IS NOT A DECISION
The plan classed three phases as hybrid — AI ranks and a person sells, AI drafts and a person approves. Then it never designed the moment where the two meet.
That gap came from the service-design question an analytics framing tends to crowd out: everything in the plan was the customer's journey or the data systems. The people delivering the service were invisible.
So for each model the plan now states what it produces, where the person meets it, and the decision they still own. "Marta — 0.81 likely to lapse, value tier 2." The CRM marketer still decides which offer to send, or none.
After the first module, Volare doesn't have a data problem. It has a new decision to make every day, and nobody there has ever had to make it.
Three rules apply to every model: a named owner, a working override, and nothing auto-sends in the first three months.
WHAT IT COSTS
Three gated modules over twelve months. Volare can stop after any of them.
- Module 1 — Foundation, months 0–3. Behavioural tracking and consent, then the three models that need no new data: churn, segmentation, service triage. €155,000.
- Module 2 — Convert, months 3–6. Lead scoring and cart recovery, measured against a holdout. €105,000.
- Module 3 — Grow, months 6–12. Recommendation, value-based routing, referral. €187,000.
The brief never disclosed booking volume, so payback is bounded rather than pinned. Two independent estimates disagreed by three times — 100,000 to 330,000 bookings a year — and both ends are shown.
Module 1 pays back in under a month at the high end, and in three at the low end. The margin and booking value behind that are named assumptions, to be replaced before anyone signs.
MY ROLE
Team of four. I produced the research, the analysis, the deck and the quote.
- The reframe — seven problems into one structural gap, stated as a hypothesis with its rivals.
- The benchmark set — every phase measured against a named source, including the two that turned out not to need fixing.
- The method — map, benchmark, redesign, decide, assess: process first, technology at step four.
- The adoption layer — designing where each model meets the person who acts on it.
- The commercial plan — module sequencing, staffing, day rates and a payback modelled across a range.
- The deck — structure, writing and visual system, revised after mid-project feedback from Prof. Della Bella.
WHAT I'D DO DIFFERENTLY
- The deck ran long. The brief asked for fifteen slides; I submitted twenty-five. Around seven go deep into the models in ways the decision didn't need. The argument was strongest when it was shortest.
- The benchmarks are indicative. They come from industry aggregators, not peer-reviewed sources, and cart abandonment in particular is measured differently by everyone. I stated that on the slides; I'd want Volare's own definitions before relying on any of them.
- The real answer exists, and I haven't seen it. The case comes from a real engagement with a real agency, and that company has already made its decision. The most useful thing I could do now is compare the two.
Next Project
The Switcher →