The pitch for an AI-first research platform usually centers on the headline capability. The model can generate a survey from a brief. The model can analyze open-ends. The model can write insights. The demo is impressive and the deck is clean.
What the deck rarely covers is what happens after the data comes in. How does the cross-tab actually render? Can the base of each percentage be configured independently? Can the same question appear twice with different aggregations? Can rows be re-ordered? Can the categories be remapped on the fly without altering the underlying data? Can the report be exported into SPSS in a format a senior analyst will accept?
These questions are the difference between a tool a researcher can use and a tool a researcher will keep open in a second tab so they can finish their work in their old system.
The long tail that nobody pitches
Every research platform has the obvious capabilities. They all have cross-tabs. They all have frequency counts and means. They all have an Excel export. The differences live in the long tail of small affordances that experienced researchers depend on.
A senior researcher building a tracking report needs to set a different percent base on a screener question than on the main brand-funnel questions, because the populations are different. They need to show the same question once as a frequency and once as a top-two-box, side by side, because that's what the client expects in the format they've used for ten years. They need to drop a row, recategorize three options into one, sort the rest in the order that matters for the deck, and have all of this persist across waves so next quarter's report comes out the same way.
None of those are AI features. None of them show up in a launch announcement. All of them are the difference between "I can ship this report from this tool" and "I'm going to export the raw data and finish it somewhere else."
The fastest way to lose a senior researcher is to fail at one of these. AI features that wow them in the demo do not buy you any forgiveness. The reporting craft is what they were doing yesterday, and yesterday's tool does it. If yours doesn't, they switch back.
Why AI-first platforms tend to underinvest here
I think this happens for two reasons.
The first is that the people building AI-first platforms are mostly engineers and AI researchers, not market researchers. The features they find obvious are the ones that map to the technology they're excited about. The features they find boring — the ones that exist because a working senior analyst spent fifteen years finding out exactly why they needed them — get classified as polish and pushed to "later." Later rarely arrives, because by the time you have enough users to feel the gap, your roadmap has been captured by your AI investors who want the next model story, not the next configurable percent base.
The second is that the long tail is genuinely long. Configurable percent bases, multi-aggregation, re-orderable rows, fine-grained recategorization, persistent settings across waves, SPSS export that matches the analyst's preferences, statistical tests that match the standards their organization has used since 1995. There's no headline feature here. There's a hundred of these, each of which takes a week to do right, and none of which buys you a slide in the next funding pitch.
The combination is fatal. The platform looks great in a demo and falls apart in production, because the production work depends on a thousand small things the platform never bothered to ship.
The bar for being usable
Here's the bar a research platform needs to clear, before any AI capability matters.
Reports need configurable bases per question. Not just per report. Per question, because the same report often combines screener percentages, brand-funnel percentages, and message-test percentages, each of which has a different base.
Reports need multi-aggregation. The same question, in the same report, shown as frequency and top-two-box at once, because the client wants both numbers next to each other and exporting them as two separate reports is not a workflow.
Reports need row-level control. Rows in the order the analyst chose, not the order the platform's default sort produced. Persistent across runs.
Reports need recategorization. On the fly, in the report editor, without modifying the underlying response data. So that the next wave can change the buckets without invalidating prior comparisons.
Reports need stat testing that matches the analyst's standards. Residual t-tests, row t-tests, column t-tests. Configurable confidence levels. Configurable minimum-respondent thresholds.
Reports need export fidelity. SPSS files that pass the analyst's eye. Excel exports that don't lose formatting on the third tab. Raw exports with stable respondent identifiers across formats.
This is the table-stakes list. None of it is exciting. All of it is necessary. A platform that ships all of this still has to deliver on AI. A platform that ships AI but fails on this list is unusable for serious work.
Why we keep pushing on this
Every release we ship adds an AI capability or two, and the announcement copy talks about those. What the announcement copy doesn't talk about is the four or five reporting affordances we added in the same release, because those don't make for an exciting headline.
Last release added configurable percent base per question, multi-aggregation on the same question within a single report, re-orderable rows, derived questions that combine two demographic variables into a single cross-tab dimension without writing a recode script, and stable question identifiers that survive a wording change so trackers don't break between waves. These are not the headline. They're the work that makes the headline matter.
The pitch for AI-first research is real. The agent stack changes what a small team can ship. But it changes it on top of a platform that still has to do the unglamorous reporting work that senior analysts have been doing for decades. AI-first doesn't mean skipping the craft. It means earning the right to call yourself research software, and then doing something more interesting on top.
A platform that takes the reporting craft seriously, and also has the agent stack, is the platform a senior researcher can actually move to. A platform that has the agent stack but treats the reporting layer as polish loses them on the third report.
Researchers will tell you which side of this you're on. They just won't tell you nicely.