Structured so that we can abandon it cheaply at the end of any phase. Every gate is a number we already measure, not a judgement call.
Georgia only, because we sell it well and know the ground product. Runs inside the WhatsApp number customers already message. Staff see every reply before it goes out and can override it. No website changes, no app release, no new infrastructure.
GATE: beats manual first-response conversionSTOP IF: cost per enquiry exceeds AED 4Add Georgia, Armenia, Sri Lanka, Salalah and Azerbaijan. Staff stop pre-approving every message and instead review a sample. Introduce the three-option format and the change-my-mind loop, which is where we expect most of the value to sit.
GATE: sampled reply quality above 90%STOP IF: complaint rate rises at allThe planner becomes the primary entry point on holidaymakers.com, replacing the destination-first search box. Hands the composed trip to the existing booking flow. This is the first phase that touches the live consumer product, and the first that carries real brand risk.
GATE: planner sessions convert above current searchSTOP IF: overall site conversion dropsThe same tool, white-labelled, sold to agencies who compose itineraries by hand today. This is where it stops being a cost centre. Note the conflict: those agencies compete with us, so this needs a decision at board level about which business we are actually in.
GATE: three paying pilotsDECISION REQUIRED: do we arm competitors?Listing these now is cheaper than discovering them in month five.
The single worst outcome. Mitigation is architectural: the model never generates a number. Prices come from our systems and are injected into the reply. If pricing is unavailable, it says so rather than guessing.
Each conversation costs real money in model calls, and the customers who chat longest are often the ones who never book. Phase 1 exists mainly to measure cost per booked trip, not to prove the thing works.
Visa rules change without notice and getting one wrong strands a customer at an airport. Rules must come from a maintained table with a review date, never from the model's memory. Where the table is stale, it must refuse to answer.
Travel is emotional and expensive. A bad automated reply gets screenshotted. Phase 1 keeps a human in front of every message precisely so we learn the failure modes privately.
They will, unless it is framed and measured otherwise. The metric that matters is enquiries handled per agent, and it should go up. Say that in advance, not after the rumour starts.
A real possibility. The defence is not the technology — it is our own inventory, our contracted rates and our customer base, none of which a general-purpose assistant has.
Approval for Phase 1 only. Six weeks, one destination, existing channel, reversible at any point, and measured against numbers we already track.
Show me the demo again