Research or Data Agent? 6 of 11 Return Structured Data

Our category is called Research & data agents, and that name bundles two different buyer questions into one page. “Find and summarise what is known about X” is a research job, and the output is a paragraph with citations. “Turn this pile of numbers into an answer I can act on” is a data job, and the output is a table, a chart, or a query result. A tool that is excellent at the first can be useless for the second.

We checked all 11 tools in the category against one test: does it return structured, machine-readable output as a named feature, something you could hand to a spreadsheet or a downstream system, rather than a written document? 6 of the 11 do. 5 do not. Measured 18 August 2026 against the 11 listings Published in that category on that date, out of 89 published listings across the index.

The split

Tool	Structured output	What it actually returns
Microsoft Fabric Data Agent	Yes	Live NL2SQL/NL2DAX/NL2KQL queries over your own warehouse or Power BI model, capped at 25 rows and 25 columns
Genspark	Yes	AI Sheets: a live spreadsheet built from a prompt, with auto-generated Python visualizations
Webhound	Yes	A structured Dataset, a table with a source trail on every individual value, built by running Python for calculation and parsing
Elicit	Yes	A data-extraction table pulled out of papers (sample sizes, outcomes, methods), up to 40 columns on Enterprise
Julius AI	Yes	Tables and charts generated from your uploaded spreadsheet or connected warehouse, held in a Notebook alongside the code that produced them
TinyFish	Yes	Browser-rendered JSON from live web pages, scoped by CSS selector, returned through an API rather than a screen
Perplexity	No	A synthesized, cited answer or a longer Deep Research report
GPT Researcher	No	A long-form cited report, typically a couple of thousand words
Manus	No	A finished deliverable (report, deck, small app); it can open a spreadsheet but does not produce one as a feature
GC AI	No	A reconciled, cited legal analysis across case law and statutes
Undermind	No	An organized report with match scores and a citation network

The six are three different products, not one

Structured output turns out to be a weak family resemblance. The six do very different things with a table, and the differences decide which one a buyer should be looking at.

Two of the six query data you already own. Microsoft Fabric Data Agent is the purest case: a natural-language layer that translates a plain-English question into SQL, DAX or KQL against your lakehouse, warehouse or Power BI model, read-only and scoped to your own permissions, per Microsoft’s own documentation. It requires an existing Fabric or Power BI Premium capacity to run on at all. Julius AI starts from the same place but a rung down the org chart: you upload a CSV or connect Snowflake, BigQuery or Postgres, ask questions in chat, and get back tables and charts with the Python that produced them kept in a Notebook. Fabric assumes a data platform team; Julius assumes a spreadsheet and a deadline.

Three of the six build a table that did not exist yet. Genspark’s AI Sheets turn a prompt into a live spreadsheet with generated Python charts, inside a wider workspace that also makes slide decks. Webhound builds a Dataset the way a research assistant builds a market map, browsing and verifying and running Python, then handing back a table where every cell traces to a source, priced per dollar of compute from $1 to $25 and up. Elicit does the same trick against academic papers, pulling sample sizes, outcomes and methods into an extraction table of up to 40 columns on Enterprise.

One of the six does not analyse anything. TinyFish is an API you call, not an analyst you ask. Its Search and Fetch endpoints return browser-rendered JSON from live pages, its Agent runs multi-step extraction billed at $0.016 per step, and nothing in the product writes a sentence for a human to read. It passes our test more cleanly than anything else in the category and answers a completely different question, which is the finding worth carrying away from this census.

What the five that return a document are for

Perplexity, GPT Researcher, GC AI and Undermind all return a written, cited document, and none of them has a table-producing feature on its own listing. Manus is the fifth, and it is the one genuinely close call in the group, so it gets its own section below. Returning prose is not a weakness. Undermind and GC AI in particular are scoped to a single deep corpus, academic literature and US case law, where a narrative answer with a full citation trail is the deliverable a researcher or a lawyer actually wants.

The distinction that matters is what the tool hands back when it finishes, not how sophisticated it is on the way there. GPT Researcher runs a planner-executor pipeline across twenty or more sources and returns two thousand words. Undermind reads hundreds of full papers and estimates statistically how much of the relevant literature it found. Both are doing more work than a spreadsheet generator; both leave you with prose.

Where we drew the line, and the two cases that test it

Manus is the edge case the first version of this census flagged, and it stays where it was. Its own product description says it processes “spreadsheets and documents” as part of a broader task, and independent testing cited on its listing notes it does well at data collection. That is general file handling shared by most autonomous agents that touch a filesystem, not a named structured-output feature the way AI Sheets or a Dataset is.

Perplexity is the case that TinyFish’s arrival created, and it is the sharper of the two. Perplexity ships a Sonar API returning web-grounded answers and raw ranked search results, so a reader could fairly ask why an API earns TinyFish a Yes and not Perplexity. Our line is about what the answer is made of rather than how you fetch it: TinyFish returns extracted field values as records, while Sonar returns the same cited prose the web product returns, with links attached. Someone who weighs a developer endpoint more heavily than we did would move Perplexity across and make the count 7 of 11, and we would rather write that sentence than pick the line that makes the split look tidy.

Why the split matters before you buy

If the job is “get me a sourced answer to read,” any of the 11 will do it, and the household names are the more mature choices. If the job is “get me a table I can put in front of someone else,” 6 of the 11 are built for it and they are not interchangeable. Fabric only works if you already have a Fabric or Power BI capacity, Julius wants your CSV, Genspark and Webhound start from nothing, Elicit only makes sense if the source material is papers, and TinyFish will hand you JSON and no opinion at all.

Searching “AI agent for data analysis” and landing on a citation-and-paragraph tool is the mismatch we see most often in this category, and it is a naming problem more than a vendor-quality one. The category grew from 9 tools to 11 between 14 and 18 August 2026, and both arrivals landed in the structured column, which moved the ratio without changing the advice.

Method

Checked 18 August 2026 against the 11 listings Published in our Research & data agents category on that date, from an index of 89 published listings. This supersedes our 14 August 2026 measurement of the same category, which found 4 of 9 on an identical test; Julius AI and TinyFish were published on 16 and 17 August and both qualify, and none of the original 9 changed classification on re-check.

A tool counted as “structured output: yes” only where its own listing describes a named feature returning tabular, per-value or queryable data, rather than a general claim of data analysis or file handling. The six qualifying features are Fabric’s NL2SQL/DAX/KQL routing, Genspark’s AI Sheets, Webhound’s Dataset, Elicit’s extraction table, Julius AI’s generated tables and charts, and TinyFish’s Search and Fetch JSON, each documented on that tool’s own listing with a primary source. Manus’s file processing and Perplexity’s Sonar API were both weighed and excluded for the reasons given above, and a different defensible line on either would move the count to 7 of 11.

This is a census of our index, not a survey of the market. There are data-analysis agents in the world we have not listed, and this reports on the 11 we have.

Corrections and additions: write to us.