An AI agent for market research: what it can look up, and what it cannot know
How an AI agent for market research plans a question, fetches live counts of businesses, residents, hiring and news, and writes an answer you can check line by line.
Part of the guide Location intelligence automation
An AI agent for market research is a language model that is given tools: it breaks a research question into steps, fetches live data for each step (how many businesses of a kind are in an area, who lives there, who is hiring, what the local news says), and writes an answer from what it fetched rather than from memory. The difference from asking a chatbot is that every figure in its answer came from a lookup you can see, so you can check it.
This article is for founders, strategy teams and consultants who want first-pass market answers in minutes instead of days. It covers which questions suit an agent, how it works through a bigger task, what it cannot tell you, a made-up example from Bangkok, the traps to avoid, and how to set one up so it answers in a chat or delivers a report every month.
Questions an agent handles well
An agent is strongest on questions that break down into countable parts, where each part maps to a data source it can query:
- "How many gyms are within 3 km of these two addresses, and how many are chains?"
- "In which of these four districts are there the most residents per pharmacy?"
- "Which companies in this city are hiring store managers this month?"
- "What opened in our category near the old town in the past year, where the data has opening dates?"
It is weaker on questions whose answer is not in any data it can reach: willingness to pay, how a brand is perceived, what a landlord will accept. For those it should say it does not know. A well set up agent does; a badly set up one guesses.
How it works through a bigger task
For a short question the agent decides which tool to call, reads the result and answers, sometimes calling two or three tools in turn. For a bigger brief ("a first look at the bubble tea market in three districts") it can work in a planning mode: it writes a to-do list, works each step with its tools, keeps notes as it goes and finishes with a report. If a step depends on another AI employee's work, it can hand that step over and use what comes back.
Every call it makes is recorded. In xMap Flows the run panel's See how it answered shows the plan, each tool call with what it returned, and how long and how much each step took. That record is what turns an agent's answer from "sounds right" into "checked".
A worked example, made up
A made-up drinks start-up asks: "We want to open our first bubble tea shop in Bangkok. Compare the areas around three shopping streets: competitors within 1 km, residents within 1 km, and whether rival brands are hiring there." The agent's plan, in its own steps:
- Find the three streets and use a point on each.
- Count tea and dessert shops within 1 km of each point, noting the chains among them.
- Count residents within 1 km of each point, since Thailand has population figures.
- Search job postings in Bangkok for tea shop and cafe staff, and note which brands post near each street.
- Work out residents per competitor for each street and rank them.
- Write a short comparison and say what the data does not cover.
Its answer, in this invented case, ranks the second street first on residents per competitor, notes that one large chain is hiring near the first street (a sign it may open there), and ends with what it could not check: rents and how many people walk past at night. Each number links back to the step that produced it in the trace.
Mistakes to avoid
- Giving it no tools. An agent with no data blocks linked can only answer from its training, which is out of date and has no local counts. Link the data it needs.
- Not telling it to admit gaps. Write in its instructions that it must say when the data does not answer a part of the question. Then check that it does.
- Asking for a forecast. Counting what is there now is reliable; predicting next year's sales is not something these tools measure.
- Skipping the trace. Read how it answered the first few times. If it counted the wrong kind of place, fix the instructions once rather than correcting every answer.
- Expecting coverage everywhere. Population figures are not available for every country (not the United Kingdom or the United States, for example). An agent should say so rather than leave the column blank without comment.
Setting it up in xMap Flows
There are two useful shapes. The first answers questions in a chat; the second runs a fixed brief on a timetable.
A research assistant you can ask
- Chat: a chat window in the editor, or a page you share with your team.
- AI agent: instructions that say who it researches for, which kinds of question it takes, and to state what the data does not cover.
- Its tools: Nearby places for counts and lists of businesses, Demographics for residents, Job postings for hiring, News for local stories.
- Reply: sends the answer back in the same chat.
A monthly brief
Put a Schedule in front of the agent, set it to work through a big task, and end with Make a report. Each month the same questions are answered again with fresh data and the PDF arrives by email. For a question that needs a calculation over many rows, the Data analyst block writes and runs a small program instead of estimating.
The templates Research assistant over live location data and Market entry study, refreshed monthly are both shapes ready to adapt. The Learn page on market sizing by area explains the counting behind most answers, and the guide to location intelligence automation covers the wider set of jobs.
Questions
Does the agent make up figures?
It is built to answer from what its tools return, and every call is recorded so you can check. Instructions that tell it to say what it does not know, and a Check the work step, reduce the risk further.
Can it research any country?
Business counts work in most countries. Population figures, footfall and property data cover only some, and the agent should say when a layer is missing.
How long does a research answer take?
A short question usually takes under a minute. A planned brief with several steps can take several minutes and is better run on a schedule.
Can my team ask it questions on WhatsApp or Slack?
Yes. The agent can answer messages from apps such as WhatsApp or Slack, and Reply sends the answer back in the same conversation.
Templates that do this
Read next
More on business and places data · Every article · Templates