FlowsBook a callSign inCreate free accountSign up
Site selection and expansion

AI agent for site selection: what AI tools can check

What AI tools for site selection check today: an AI agent looks up places, residents and roads, compares candidates and writes the brief. Where people decide.

Part of the guide AI employee for site selection

By the xMap team · · 5 min read

An AI agent for site selection is an assistant that, when someone on your team asks about a location, looks up the facts itself with live location data (competitors nearby, residents in the area, how the roads move) and answers with figures it can show. It saves the hours spent pulling the same counts for every address an agent or a franchisee sends. It does not sign leases: it gathers and compares, and people decide after a visit.

This is for expansion teams, franchise developers and consultants who are asked "what about this spot?" several times a week. It covers the two ways to use an agent (answering questions in a chat, and working through a long brief on its own), the tools to give it, a made-up week for a restaurant group in Paris, the guard rails worth adding, and what an agent should never be trusted with.

Two ways to use an agent

  • A site advisor in a chat. Colleagues write "how many sushi restaurants within 1 km of Rue de Rivoli 120, and how are they rated?" in a chat page or on WhatsApp. The agent picks the tool it needs, runs it, and answers in a few lines with the figures. It remembers the conversation, so the next question ("and around the other address?") needs no repeating.
  • A brief it works through on its own. Set to work through a big task, the agent takes a goal such as "rank the eight districts on our list for a second branch and write a one-page memo", makes a to-do list, works each step with its tools and ends with a report. It can run on a schedule, for example every month for the expansion committee.

The tools to give it

An agent in xMap Flows can use any block linked to its Tools port, and only those. For site work, a useful set is:

  1. Nearby places: competitors and anchors around any point, with ratings, chains and opening dates.
  2. Demographics: residents and age groups around a point, in the countries that have population data (France, Canada, India, Thailand, Egypt and Australia among them; not the United Kingdom or the United States).
  3. Traffic Flow: live road speed and the jam factor next to a site. It measures speed and congestion now, not counts of cars.
  4. Mobility Footfall, where it covers the city: how busy the area is, as people-hours of presence, arriving about ten days late.
  5. Look things up in your own files and tables: your store list, past committee memos, the rent ranges your property team keeps.

In each tool you decide what the agent may change. Writing {{agent}} in a setting, such as the point to search around, lets the agent fill it on each call; settings you fill yourself, such as the types of competitor, stay as you set them. That keeps it searching the right kinds of places even when a question is vague.

A made-up week in Paris

This example is made up. A restaurant group with four sushi restaurants in Paris gives its expansion manager a site advisor on a shared chat page. On Monday a broker sends three addresses. The manager pastes them into the chat and asks for rivals within 1 km, their average rating, residents within the same distance and the road speed outside at lunchtime. The agent answers with a small table and notes that the third address has eleven sushi restaurants nearby, four rated 4.5 or more.

On Wednesday the founder asks in the same chat which of the three is most like their best restaurant. The agent looks up the best restaurant's own figures from the store table, compares, and says the second address matches on residents and rivals but sits on a road that moves slowly at lunch. The manager visits the second address on Thursday and asks the broker for terms. Separately, on the first of the month the agent's brief mode writes the committee memo ranking eight districts, which the manager edits before it goes out.

Guard rails worth adding

  • Check the work. A Check the work block after the agent can refuse answers that quote a number its tools did not return, and send them back to be redone.
  • Ask before anything goes out. If the agent can email brokers or update a sheet, put Approval after it so a person says yes first.
  • Protection on a shared chat page. Keep it to site questions, so it does not wander into topics you did not set it up for.
  • Test cases. Save a few questions with known answers and run them after every change to its instructions.

What an agent should not be trusted with

  • The final choice. It cannot see the unit, the frontage, the landlord or the rent. Treat its answer as a brief for the visit.
  • Figures it was not given. If a country has no population data, the honest answer is "none here", and you should expect it to say so rather than guess.
  • Forecasting sales. It compares places on what is around them. Revenue forecasts need your own sales history and a model built for that.
  • Long drive times. Its reach is what its tools measure: travel areas go up to 15 minutes or 10 km by road for now.

Setting it up

Start from the template site advisor chat for the expansion team, which links a chat, an AI agent with place and population tools, and Reply. Add your store table under Look things up in, write the agent's instructions in your own words (what your formats need, which rivals count), and share the chat page with the team. For the monthly memo, switch a second agent to work through a big task and give it Make a report. The site selection guide shows the other site jobs an AI employee can take on, and where to open next is a plain flow version without the chat.

Questions

Can an AI agent choose the best site for me?

It can rank candidates on the figures its tools return and explain why. Choosing still needs a visit and the commercial terms, which it cannot see.

Does the agent make up numbers?

It is told to use only what its tools return, and a Check the work block can refuse answers that quote figures it was not given. Test it with questions you already know the answer to.

Can colleagues ask it on WhatsApp?

Yes. The same agent can answer on a chat page, on WhatsApp or in Slack, with Reply sending the answer back on the channel the question came from.

What does it remember between questions?

The last messages of each conversation, so follow-ups work, and, if switched on, short notes about each person it talks to, such as the formats they ask about.

Templates that do this

Read next

More on site selection and expansion · Every article · Templates

Put it to work

An AI employee in xMap Flows does this on live data, on its own timetable. Start from a template or describe it in plain words.