Skip to content

AI, used carefully

AI picks the places. It never writes a price.

Outrouted uses a language model for the parts that need judgement: what kind of trip you want, which places belong together, the order and the nights. Distances, prices and travel constraints are not creative writing, so they are computed.

1Vilnius3 nights2Riga2 nights3Tallinn3 nights4Helsinki3 nights
  1. Vilnius3 nights
  2. Riga2 nights
  3. Tallinn3 nights
  4. Helsinki3 nights
Vilnius, Riga, Tallinn, Helsinki: eleven nights. The model chose the places, the order and the nights. The prices and the distances came from somewhere else. Illustrative route. The lines are straight on purpose: they show the order, not the road.
  • It never writes a price.

    A language model proposes destinations, order and nights per stop. Code checks the total nights and calculates dates, distances, travel estimates and the budget. Hotel search prices come from the provider. The model supplies no prices and no budget totals. It does pick how long each place gets from a short list of choices, and the card calls that time an estimate.

  • It only chooses from a real pool.

    The candidate places are assembled before the model is asked, from the countries and cities you picked. Anything it returns that is not in that pool is dropped, so a plan cannot contain a town that does not exist.

  • A night split that does not add up gets fixed.

    If the model returns nights that do not match the length of the trip, they are re-spread before you see them. The trip is as long as you said it was, whatever the model thought.

  • Your note is a constraint, not an instruction.

    The free-text note you write goes into the prompt in a block of its own, marked as the traveller's constraints, and the system prompt says that block carries no instructions. A note can tell the planner your mother cannot manage stairs. It cannot tell it to ignore your budget.

  • It still works when the model does not.

    A deterministic planner sits behind the model one, with the same contract. If the model times out or answers with nonsense, you still get a route.

Photographs of the route

Vilnius
Riga
Tallinn
Helsinki
Curonian Spit
VilniusThe old town from Gediminas Tower. A baroque skyline with almost no high-rise in it. Three nights suits a city that rewards walking slowly.
1 of 5 · Photo: Radosław Botev · Openverse / wikimedia

Questions

How the AI part works

Yes, for four jobs: choosing places, their order and the nights in each; writing the short note on why a place is there; estimating how long each place takes; and asking up to three questions before a search. Not for prices, distances, travel times or availability.

The trip inputs and the pool of candidate places. Your name, email and account are not part of the request.

The route is the answer and all of it is editable: add and drop stops, reorder them, change nights. Rebuilding the route afterwards will not undo those choices.

No. Planning is a job you ask for. It reads the catalogue, searches flights and stays, asks the model, then prices what comes back. What you do with the result is yours.

Ask for something specific and see what comes back.

Free to plan. Routes with prices, no account.