What 1388 Teams Want from
AI Business Research
“I wanted to know whether my app idea would actually work.”
“I needed a few seasonal street-drink concepts, judged by taste, appearance, and novelty.”
AI agents write code better than they do almost anything else. That is where AI labs have concentrated their research. Coding is the one place agents have moved from demos to daily use.
Most businesses do not run on code. They run on research. They size markets, study audiences, read competitors, and test ideas before they build anything. What people want from AI in that work is far less clear, and no one has mapped it.
So we asked. We invited people who use our platform to describe, in their own words, what they were trying to accomplish, and drew a first map from 1,338 of the research needs they shared.
How we did this
We invited Atypica's users to tell us, in their own words, what they were trying to accomplish, and worked from the 1,338 research needs they described. AI read and classified each one, which let us capture texture at a scale a human-coded study could not reach.
To make sense of a qualitative study at this volume, we drew on the approach Anthropic used to learn what 81,000 of its users want from AI, the largest qualitative study of AI users to date. We built classifiers that labeled each need across a few dimensions: what deliverable someone wants, how the research should be done, and what business decision it serves. This piece is about the last of those three. Because a single need often serves more than one decision, we let each one carry several labels rather than forcing it into a single category.
Before taking part, users were informed their inputs might be used for research and product improvement. All records were de-identified before analysis, and the examples quoted here underwent review for removal of any potentially identifying details. They are paraphrased from users' own stated goals, with identifying details removed.
A few limitations are worth naming. This is a self-selected sample of people who already reach for AI to do research. The labels were assigned by a model, not a panel of humans. And each record captures a stated need at the moment it was asked, not what the user did next.
What decisions are people trying to make?
We asked our classifier to name the primary business decision behind each need. These were the most common.
“I wanted to test whether an AI meal-planning app had real appeal.”
These decisions share one thing: timing. Almost all of them are questions people ask before a product exists. The one exception, creative testing, is also the least common decision in the data. It's the question you can only ask once a product is finished and already in front of people.
Concept validation, audience profiling, product development, positioning, and the rest — decisions made while a product is still taking shape.
Creative testing and performance diagnosis — the questions you can only ask once something is live.
Share of primary decisions, N = 1,338.
The chart above shows the shape: most of what people bring us happens before a product exists, and almost none of it happens after. The traditional research industry looks nothing like that. ESOMAR, which tracks global commercial research spend, puts customer satisfaction and CRM tracking at the top: the largest category, and work that happens entirely after launch. The two aren't measuring quite the same thing: ESOMAR counts industry dollars, and our sample already skews toward people who chose an AI platform for exploratory work. Even so, that category never appears anywhere in our data. The work people bring to Atypica starts earlier than where the traditional industry spends its money.
This pattern has a familiar explanation. Product development now moves in days and weeks: a new apparel line, a new snack, a new bubble tea flavor arrives every month. But the research meant to feed those decisions still moves in quarters and years, because it rests on traditional user interviews, and interviews are slow and expensive to run. So the pre-launch questions get asked far less often than they are needed. And even when a study does run, a second cost appears. To weigh what people said against the rest of the data, their answers get quantified so they can be weighed against the rest of the data. That conversion discards most of what made them worth gathering in the first place: the reasons, the psychology, the values behind a choice. What survives is a number. What mattered was the why.
These are the two costs Atypica removes. Synthetic users replace the slow, expensive work of traditional user interviews. Instead of flattening what people say into numbers, a language model reads it in full and keeps the reasons behind each answer. The pre-launch questions can now be asked as fast as products are built, without losing the why.
This is the research Atypica's users reach for first.
One request, several jobs
Most of the needs above did not stop at one decision. When we asked what people wanted to accomplish, 67% named two or more. On average, each one touched just over two.
"I wanted to understand the size and growth of a category, who the main players are, where the market is heading, and what customer pain points still matter."
This one request contains three different kinds of research: sizing the market, mapping competitors, and figuring out what's bothering customers.
| Decisions that arrive together | Share of all needs |
|---|---|
| Audience profiling+Product development | 8.5% |
| Audience profiling+Competitive analysis | 6.8% |
| Audience profiling+Brand positioning | 6.3% |
| Concept validation+Product development | 6.0% |
| Brand positioning+Competitive analysis | 5.9% |
When competitive analysis appears, audience profiling sits next to it nearly 30% of the time.
The table above shows which decisions tend to arrive together. The most common pair is audience profiling and competitive analysis: they appear together in close to 7% of all needs. And whenever competitive analysis shows up, there is nearly a 30% chance audience profiling is right next to it.
"I wanted to understand the target customers for a new personal-care brand, how they make purchase decisions, and how existing brands already meet or miss those needs."
That pairing is hard for an AI agent to serve well. Audience profiling and competitive analysis call for completely different work. Audience profiling means identifying demographics, behaviors, and pain points, usually through interviews or behavioral observation. Competitive analysis means market positioning, feature benchmarking, and tracking share of voice, usually through desk research. A research team would staff these two jobs differently and judge the results by different standards.
So the agent has to split the request into its parts, and treat each part with its own method. That is already hard on its own. What makes it harder is that the parts are rarely independent. The output of one is often what the other needs to even start. A first pass at who the audience is can reshape which competitors matter. A look at the competitive field can reshape who counts as the audience.
The agent has to run two different methodologies and manage how they feed into each other, all inside a single request.
Most AI agents today were not built for this. A coding agent's methodology stays the same no matter how many features a user asks for: it writes the code, runs it, checks the result, and repeats. A research agent doesn't get that shortcut. Different goals call for different methods, not more of the same one.
A research agent is routinely handed several jobs at once, has to work out what they are, and has to do each of them to the standard a specialist in that discipline would expect, all inside a single conversation.
What the map is for
Every product on a shelf started as a decision someone had to make first. So did every campaign, every price, every market a company chose to enter. That is what business research actually is: the work behind nearly every commercial decision of consequence. Globally, it is worth $142 billion a year. Nobody had mapped that work from the inside before: from what people actually ask for, rather than what the industry assumes they need. That only became possible once AI could read this much of it closely.
Two things stood out. Most of this work happens earlier than where the traditional industry spends its money. And it rarely arrives as one job. It arrives as several, each needing its own kind of expertise, and often depending on each other just to start. That is the harder problem no one has built an agent for. It is the one this map tells us to build.
We plan to keep asking. Every question above came from someone with a real decision to make. There will be more of them, in categories this first map hasn't found yet. We are not interested in building for whatever is convenient. We are interested in what business research actually requires. That is what a map is for.