A trip to Tokyo with an AI agent over SMS: real-time translation, personal context and the limits of latency

The text is a first-person personal account, published in the "Originals" section of an AI-focused outlet, signed by a guest contributor who holds executive roles tied to the narrative around agentic AI and open source and who was previously a technology editor at a business magazine. It is not a report with verified third-party data or an independent product analysis, but the chronicle of a week-long trip to Tokyo using, throughout the journey, a beta-stage AI agent called Instinct.

The author describes Instinct as an assistant accessed via SMS (to avoid burning mobile data), able to receive photos and videos and to use the phone's shared location in real time. He compares it to "having a local tour guide", though he explicitly notes that it suffers from latency, meaning responses do not always arrive instantly.

Among the concrete examples he offers: finding an unusual shochu on the list at a restaurant with no menu and no English-speaking staff; comparing Mikimoto pearl prices between the Ginza district and New York (although the agent itself made an initial error by confusing the pearls' size with prices in US dollars, something the author only caught by pressing with more questions); and, at the station, receiving highly specific visual directions —such as "look for the orange and blue doors"— along with warnings about common traveler mistakes, something a conventional two-dimensional map does not provide.

A central thread of the article is the idea that the language barrier is becoming the real battleground of user experience, alongside latency. The author recounts that, all over Tokyo, people reach for translation apps as soon as someone speaks English, but he considers that interaction —holding up the phone and passing it back and forth in front of the microphone— still clumsy. He also mentions Apple's translation feature for AirPods, which he found interesting but slow to set up and dependent on a network connection. His conclusion is that real-time translation is no longer a problem of technological capability but of response speed (latency), and that solving it —ideally in both directions— would radically change the experience of getting around a city like Tokyo.

Another point he highlights is the combination of translation and visual elements: sending a photo of a street or a sign and receiving contextualized instructions, with visual landmarks to orient yourself, instead of relying on a dotted blue line on a map. From this experience he speculates about a near future of "ambient AI" —mentioning new Apple AirPods with a camera— with latencies below 50 milliseconds, which would make visual and spoken information available at all times without taking out your phone.

The account also underlines how the agent gradually "got to know" the author over the course of the trip: with access to his calendar and email, it began setting alarms automatically, and once he shared his location it was able to correct his route on the fly (again, with some delay). This progressive learning capability is what, according to the author, let him venture into small izakayas under the train tracks near Ginza, places without a single word of English, and still receive specific recommendations along the lines of "order this, not that" when faced with an unusual dish or a bottle of rare Japanese whisky.

The text is not purely laudatory: the author devotes an explicit section to the failures. He acknowledges that Instinct pointed him to the wrong trains on several occasions, trains that did not stop where he needed, and he attributes these errors to the uneven quality of public data on the internet, which remains the information layer these tools rest on. He also flags the "recency" problem: the agent was unable to recommend a café that already existed but had not yet been sufficiently reviewed or surfaced on social media. His practical conclusion is that, for critical, time-sensitive decisions —like confirming whether the train on the platform is the right one— the agent cannot yet be trusted one hundred percent, and he recommends always asking it to double-check important decisions.

On the other side, he also documents a case where the agent clearly outperformed conventional map apps: when asked for somewhere to eat nearby with reviews only in Japanese, Instinct found a fish donburi restaurant that had opened just four days earlier, set up by one of Tokyo's major tuna wholesalers, something the author describes as practically impossible to locate with Google Maps.

The article places this experience within a broader trend: it notes that tools such as ChatGPT and Google Maps itself are adding conversational interaction, which reduces the friction of changing plans on the fly and exploring unplanned routes without losing route optimization. As an additional detail, he points out that after the trip ended, Instinct added a feature to turn the entire itinerary into a travel guide published as a simple web page, designed to preserve and share the experience.

One metatextual detail relevant to judging the nature of the piece: the author himself says he wrote the whole article by voice dictation while riding the train to Narita, and then used another generative AI to polish the text for about ten minutes. This reinforces the article's character as personal, informal testimony, in keeping with its enthusiastic tone and its lack of usage figures, market comparisons or external validation of the Instinct product, which was in beta at the time of publication.

Taken as a whole, the article's value lies not in providing hard data on the state of agentic AI applied to tourism, but in serving as first-hand testimony about where these tools are heading: multimodal, personalized and conversational assistants that reduce the friction of language and navigation in unfamiliar environments, with the author openly acknowledging their current limits —latency, dependence on the quality and freshness of internet data, and the need for human verification on critical decisions.