How We Built the AI Concierge — and What It Can (and Can't) Do

We built an AI tool for date planning. This is an honest account of what it actually does, where it falls short, and why we built it the way we did.

Most AI product announcements are written to maximize excitement. This one is written to be accurate. We think that is more useful.

The AI Concierge is a set of four tools built around a single problem: the gap between "we should do something" and "we have a plan."

Date Ideas generates activity suggestions based on your mood, your city, your budget, and what kind of experience you are looking for. It is not a search engine — it synthesizes across a range of inputs and produces suggestions that are specific to your situation rather than generic lists.

Venue Finder helps you find specific places — restaurants, bars, activity venues — based on what you are looking for. It is better at this than a general search engine because it understands context: "somewhere quiet for a first date" produces different results than "somewhere with energy for a group outing."

Conversation Starters generates questions and prompts for specific situations — a first date, a long-term relationship that has gotten quiet, a date with someone you know well but want to know better. These are not generic icebreakers. They are calibrated to the situation you describe.

AI Chat is a free-form conversation interface for date planning questions. It is the most flexible of the four tools and the most useful for situations that do not fit neatly into the other categories.

This is the part most AI product announcements skip.

It cannot verify real-time information. The AI does not know whether a restaurant is open tonight, whether a venue has availability, or whether an event is still happening. It can suggest; you need to confirm. We are working on integrations that will close this gap, but they are not there yet.

It cannot know your partner. The AI can work with what you tell it about your partner's preferences, but it does not have access to your relationship history, your partner's actual preferences, or the context that makes a suggestion genuinely personal. The more you tell it, the better it does — but there is a ceiling.

It can be wrong. AI language models produce confident-sounding output that is sometimes incorrect. The AI Concierge will occasionally suggest a venue that has closed, a neighborhood detail that is outdated, or an activity that does not quite fit what you described. Treat its suggestions as a starting point, not a final answer.

It is not a replacement for knowing your partner. The best date planning comes from paying attention to the person you are with — what they mention, what they light up about, what they have been wanting to do. The AI Concierge is a tool for generating options and filling gaps. It is not a substitute for the knowledge that comes from actually being present.

Building an AI product involves a series of choices that most users never see. Here are the ones we think matter.

We do not train on your data. Your conversations with the AI Concierge are not used to train the underlying model. What you tell the AI about your relationship, your partner, and your preferences stays in your session and is not retained or used to improve the model. This is a deliberate choice that costs us something — training on user data would make the product better over time — but we think the privacy trade-off is not worth it.

We do not show you ads. The AI Concierge does not recommend venues because they paid to be recommended. It recommends venues because they are relevant to what you asked for. This sounds obvious, but it is not the default in most recommendation products. The business model of most "discovery" apps is built on paid placement. Ours is not.

We tell you when we do not know. The AI is prompted to acknowledge uncertainty rather than paper over it. When it does not have reliable information about a specific venue or event, it should say so. It does not always — language models are prone to confident incorrectness — but we have built the prompting to push toward honesty rather than away from it.

We built four focused tools rather than one general one. We could have built a single chat interface and called it done. We built four tools because the different use cases — idea generation, venue finding, conversation prompts, free-form planning — benefit from different interfaces and different prompting. The focused tools are more useful for the specific tasks they are designed for.

The AI Concierge is a first version. The things we know it needs:

Real-time venue data. The ability to check whether a restaurant has availability, whether an event is happening, whether a venue is currently open. This requires integrations we are building.

Partner preference memory. The ability to remember what you have told us about your partner across sessions, so you do not have to re-explain their preferences every time. This requires a careful approach to privacy — we are working on a model that gives you control over what is stored.

Better local knowledge. The AI's knowledge of specific cities and neighborhoods is uneven. It is better in major cities and worse in smaller ones. We are working on improving this through a combination of better prompting and local data sources.

The AI Concierge is useful for generating ideas and filling gaps. It is not a replacement for knowing your partner, and it is not a source of verified real-time information. It is a starting point — a way to move from "we should do something" to "here are some options" faster than you could on your own.

We think that is genuinely valuable. We also think being honest about its limitations is more useful than overselling it.

If you want to try it, the AI Concierge is here. If you have feedback on what it does well or where it falls short, we want to hear it.