AI plans a route: a similar address can mean the other side of town
Check an AI route from address and map point to visit order, distance measure and time conditions. A similar result does not confirm a place.
Article language: English

In this article
The car reached the right number, but on the other side of the fence. Everything looks excellent on the map: the marker is almost on the building. The employee carrying a box of samples is less delighted because the entrance is on the neighbouring street. A few metres on screen can mean a substantial walk with company identification tucked under one arm.
In the fictional Nurt studio, AI helps prepare the order for delivering demonstration materials. This is ordinary company-sample logistics. We are not assessing route safety or guaranteeing deadlines. We check whether a proposal preserves addresses, identifies the right places and clearly explains the basis of travel times. The situations below are examples for planning a trial, without actual map calls or journeys.
The address in a message must survive the search for a place
Start with information from the visit's contact. Company name, locality, postal code, street, building number and entrance description have different roles. “The studio by the station” may help in conversation but is not an unambiguous address. Retain the original beside the AI-prepared version so changes before the place search can be identified.
A model can arrange components or expand an obvious abbreviation, but should not add a missing city without confirmation. A street with the same name may exist in several localities. If the source does not resolve it, send the address for completion. Adding the best-known city does not turn a gap into confirmed data.
Prepare easily confused entries for the trial: a building number with a letter, a unit number, a similar street name and an address missing a locality. Record the expected behaviour at each transition. The point is not that AI must always return some point. With ambiguous input, a useful result may be a request for confirmation before further route planning.
If addresses pass to another tool as a structure, check the meaning of JSON fields in AI output (in Polish). Valid format will not detect a unit number swapped with a building number. Retain a visit identifier in the final list so that after reordering, the address does not become detached from the contact and the correct package.
A found point is not yet a confirmed entrance
Converting an address into a geographic point is called geocoding. Results can describe different kinds of place at different levels of precision. Google's comparison of geocoding and address validation describes, among other things, partial matches and location types. Returning coordinates does not itself confirm that every part of the input was recognised precisely.
In the described service, a partial-match signal requires looking at what was actually matched. Do not automatically reject it, but do not hide it under a general “address valid” either. For Nurt, the decisive step is comparing locality, street and number with a confirmed source. A response indicating only the street is insufficient to establish the sample handover location.
Absence of a partial-match signal does not itself confirm that an address is correct. Compare the result with the input. In the trial show the received and returned addresses separately. The person confirming should see the change, not just an application's green label.
Check what a map point means. An area centre, an approximation near a road and a place corresponding to an address are not interchangeable. Even a precise point does not explain which entrance reaches the recipient. In our example, separately confirmed information about the side-street entrance is needed. AI should not invent it from the building's shape.
A facility record should connect name, address and how to reach it. That relationship is described in a map of service points and location list (in Polish). When planning visits to several branches of the same company, check their identifiers too. Choosing the first similar result can preserve the correct brand while moving the visit to another district.
If an application passes coordinates, check that latitude and longitude are assigned to the right fields and in the expected format. Test this on a known approved point before applying the check to the whole list. A model should not correct the order from the numbers' appearance. What matters is the particular tool's field and format description, not an assumption that every pair has the same order.
Distinguish a correspondence address from a meeting place. A company may have its registered office in one building and accept materials at a studio elsewhere. Record each address's role. If sources conflict, clarify the discrepancy with the right person. A form more convenient for a tool is not sufficient reason to change the agreed place.
Keep a missing response from a mapping service as a separate state. Do not replace it with coordinates generated in text by a model. If the tool is unavailable, the team can use an agreed manual process. That plan should state who confirms the location and how a visit awaiting clarification is marked. Do not omit a visit from the full list merely because its location field is empty.
Distance and order answer different questions
“Closest” needs a qualifier: straight-line distance, road distance or predicted travel time? A model may sort by one measure and describe the result as another. In acceptance, record the source and calculation method of each measure. A number of kilometres alone cannot reconstruct the time between visits.

At Nurt we do not infer driving from the closeness of two markers. The route can follow a different road from the line joining points on a map. If a tool returns a route for a specific mode of travel, retain that condition in the report. Do not put a walking time into a car plan just because both results use minutes.
Google Routes API documentation describes waypoint-order optimisation as a separate option. By default the supplied order is retained; optimisation considers travel time and other factors such as distance and turns. The words “optimal route” in an AI answer do not establish that the relevant function was used at all.
Keep input and returned orders in the trial record. After a change, check that all visits remain, none appears twice and the right materials are assigned to each place. A result with a new point order must be correctly joined to input records. A shorter calculated drive does not compensate for swapping two packages between partners.
Also retain visits excluded from calculation. If one location needs clarification, show it beside the route with a status and the person responsible for the next step. The list should account for all input cases: included, rejected under an agreed rule and pending. Otherwise a short route may simply mean the application dropped the hardest point and an AI comment called that an improvement.
Clarify start and finish too. A route starting at the studio, including three visits and returning to the warehouse is a different problem from a journey ending at the last recipient. If the model omits the return, the list may look better. Do not present that difference as an improvement without identifying the changed assumptions.
Do not extend an example from one tool to every mapping service. The team should know which calculation the provider performed, what the application arranged and what AI added in its commentary. That separation helps locate the right stage for a correction. A fluent story about a shorter route can be harder to check than a simple list of points with their bases retained.
Travel time does not automatically cover the whole visit
For each time value, record the conditions: mode of travel, assumed departure time and date the result was prepared. Also establish whether traffic data was considered, if the chosen service offers it and the relevant mode was used. Missing information limits interpretation. A model should not call an ordinary estimate a guaranteed arrival time.
Establish the activity at the point: delivery, collection or a meeting with the contact. Delivery and collection rules (in Polish) help distinguish the handover method from location itself. At Nurt the package must reach a specified person, so ending a route at a building does not yet confirm a visit was completed. Record the completion condition before starting the trial.
Driving alone does not cover every team activity. A plan may need package collection, finding an entrance, handing over materials and a conversation. Enter those assumptions separately rather than hiding them in a general route margin. Omitting three meetings of 10 to 19 minutes each can affect a plan more than the choice of street order.
Visit times are conditions that must survive reordering. If one recipient accepts visitors only in the afternoon, an order favourable for driving alone may not suffice. Calendar planning with constraints (in Polish) helps separate available time from a proposal arranged without that data. Do not assume every route-sorting feature solves all meeting conditions too.
The date must also be unambiguous. “Friday at ten” must refer to a particular day; different time zones need further clarification. Date choice and representation (in Polish) matter for data passed to planning too. A correct map point will not rescue a visit scheduled for another week.
Walk through a few decisive cases
Prepare a correct address, an ambiguous entry, two similar branches and a changed order. The reviewer should follow the source to the point and then to the visit list. For each discrepancy identify the stage: transcription, location selection, route calculation or interpretation. Do not try to fix every problem with one instruction to “be more accurate”.
Also check the link opened by the person travelling with the materials. Does it lead to the same branch and point approved in the list? Establish whether the application passes the confirmed location or searches again from a text name. In the latter case the result needs its own check. A matching link label does not establish which place actually opens in the recipient's tool.
Keep the proposal version and later corrections. If the contact confirms a different entrance, check the effect on the relevant segment and visit arrangements. If the time or list of points changes, earlier approval concerned the old conditions. The person using the plan should see the current version without comparing several employees' private messages.
Aeneia Audit presents audits concerning AI use and information flows. In the fictional Nurt studio, the team can separately establish its own acceptance criteria: a list of confirmed locations, explicit time conditions and unresolved addresses for clarification. The samples must reach the right studio, not merely a point that looks close enough on screen.







