I asked my 40-person research and data science team what a year of AI changed.
Everyone has an opinion about what AI is doing to knowledge work. Research is becoming obsolete. Data scientists are no longer needed now that anyone can query data in plain language. Everything is collapsing into one big AI insights role.
A few months ago I wrote about this, arguing AI commoditizes execution but not judgment, and that teams like mine had to move out of the comfortable middle before it disappeared underneath us. That wasn’t just my view. My leadership team and I had spent real time aligning on it and building a strategy around it. What we hadn’t done was have it affirmed by the people doing the work day-to-day. So this time, instead of adding another opinion, I asked the whole team.
I lead a 40-person research and data science (RAD) team at Fin, where we build an AI Customer Agent. A few weeks ago I sent them a short survey:
What are you spending time on now that you weren’t a year ago?
What’s actually changed in how you work?
When you hear “our version of 2x,” what does that mean to you?
Is your role evolving into something else, and if so, into what?
I expected stories about efficiency and moving faster. Instead, three themes surfaced again and again:
The routine work has largely disappeared.
The lines between roles are blurring.
Impact, rather than output, was the greatest opportunity AI afforded us.
And underneath all three was something I hadn’t fully expected. We’d set the direction, but no one planned how it would play out. There was no reorg and no new job titles. People changed how they worked one by one, each following wherever AI made them most useful, and together that added up to a transformation faster than any formal restructure could have delivered.
The routine work fell away
The first change was the biggest. Many of our most common day-to-day tasks have moved to the background.
SQL, and the work that surrounds it, used to be central to our jobs. Significant time was spent on data investigation: finding the right tables, figuring out where all the data you needed lived, getting the grain right, building intermediate tables, getting the correct calculations, and debugging incorrect ones. All of this was necessary to get useful output from a SQL query. But today, this effort is largely owned by AI.
As one data scientist put it: “SQL used to be the majority of the job, and now it’s near zero.” That’s not a small change. SQL was the main thing we interviewed and tested for on the data science side of the team, and now it barely rates a mention.
Instead, we spend time building infrastructure around the SQL: the automated pipelines that gate, classify, analyze, and deliver insights. This helps teams outside of data science readily and reliably access them.
Our data visualization workflows have also changed. Creating Tableau dashboards and reports used to be a core deliverable and now they’re being replaced by HTML reports from internal tools like Claude4data and Cockpit.
Sample report from Cockpit.
The researchers feel this work from the other side. They pull their own data now, no longer waiting on a data scientist to unblock them. As one of them put it: “Exploring and analyzing customer data is now a given for any project. It’s at our fingertips, and we have every opportunity to interrogate, strengthen, and stress-test research findings without relying on data science. It would be crazy not to check whether your insights are strengthened or challenged by larger-scale customer data.” They also draw on far more evidence than they used to. A year ago, triangulating a finding mostly meant asking a colleague what they’d recently heard. Now a researcher can cross-check a hunch against real data from across the business: churn surveys, product satisfaction scores, win/loss records, and sales-call transcripts.
The most impactful part, though, is the speed of synthesis. One researcher described reading hundreds of sales calls, win/loss records, and interviews as a single body of evidence, in one working session. A year ago that was either weeks of manual reading or it simply didn’t get done. Starting from evidence that already exists, rather than running something new, used to be the fallback option. Now it’s often the smartest first step. You can build deep context on a question from existing studies, reports, and customer data before deciding whether you even need to commission fresh primary research.
So data scientists aren’t spending their days writing SQL, and researchers aren’t waiting on them for data. The slow work of gathering and synthesizing information is disappearing. The question is what rushes in to fill the space it opened, and that’s where the survey surprised me.
The roles are collapsing into each other
The most interesting takeaway from the survey was that people no longer describe themselves the way their job titles do. And the blurring runs in multiple directions.
Researchers are drifting toward strategy. “I feel less like a researcher and more like a strategist,” one wrote, because the manual work of gathering and reading evidence has diminished and the hours now go to framing the question and stress-testing the answer. Another was even more direct: “I don’t see myself primarily as a researcher anymore, if I’m honest. I’m more of a product and market strategist who focuses on analysis. I’m not sure what the title should be.” A third described research and data analysis as simply “collapsing into one.”
Meanwhile, data scientists are drifting toward engineering. They’re shipping production front-end, wiring language models into pipelines, and building internal tools that used to need a dedicated engineering team. One of them described forward-deployed data science and forward-deployed engineering as “overlapping from both sides.” Engineers can do the basic analysis that used to sit with data science, and data scientists are starting to make changes directly in our product rather than handing off. Another listed the roles he’s gradually becoming as: “part sales engineer, part value consultant, and part researcher at times.”
These changes weren’t entirely unguided. A few months earlier we’d asked the team to exit the comfortable middle, the routine work that no longer needed us, and move toward the harder, higher-judgment work. What we hadn’t done was script how each person would get there. There was no reorg, new titles, updated competency framework, or career ladder. People followed the work to where the leverage was, and the disciplines bent to meet it.
Building systems, not reports
The survey included dozens of examples, but they all shared one idea. People have stopped producing one-off work and started building things that keep working after they walk away.
One data scientist built an AI Agent that generates customer ROI decks. It started as a single pitch he wrote for one customer during a forward-deployed session, got good feedback, carried into the next deal, and then became something any account team could trigger with a click. It has now produced well over 450 decks, more than a hundred a month, and changed how the sales team pitches Fin. One system, built once, adopted far beyond the person who built it, doing commercial work every week.
Another data scientist built an entire customer feedback product inside Cockpit, our internal platform, more or less from scratch in about a month without an engineering background.This included: the backend, the React front-end, and a LLM job that clusters and summarizes customer feedback. It has fundamentally changed how the team does research, because instead of reading fragmented feedback across a dozen sources, there’s now one place to find it, group it and act on it. Interestingly, before its creation, we had been evaluating external tools for this skill but ended up building our own customized solution in less time than it usually takes to run initial evaluations on vendors.
On the commercial side, a data scientist and a researcher built a win/loss and live-deal intelligence system. It turns unstructured deal text, Gong calls, Salesforce notes, emails, and Slack threads into structured insight on why we win and lose, how competitors show up, and how live deals are progressing. The process used to be manual and infrequent. Now it runs daily, at a scale no single person could manage by hand, and it feeds a fortnightly deal briefing to our senior leadership team.
Our researchers are building similar tools. One of them ran a deep customer discovery effort, sitting down with more than 60 customers over different points in time. That part doesn’t get faster because AI exists, and it shouldn’t. Taking time to talk to the right people about the right thing is valuable. What changed is what he did with the material afterwards. Instead of writing it up once, through a single lens, as a report that starts to age the day it ships, he turned the whole set of interviews into a searchable repository that he adds to and built a skill he can query whenever a new question lands. Same high-quality research, but it now answers questions for months, from angles he never planned for when he ran it.
And some of the most important building is invisible. A few people noticed that different AI skills were each defining the same metric slightly differently, so the same question could come back with different answers depending on which one ran. Creating shared, verified infrastructure for our AI tools is essential to their accuracy, so the team is building out a semantic layer.
What 2x means for RAD
It would be easy to read all this as “the team builds systems and software now,” but it’s more than that.
Take the researcher who stood up audience messaging cohorts for marketing leadership. She pulled the raw data herself, including Gong calls, win/loss records, and recent studies, into a single picture of who the buyer is and what moves them. From there, she put an initial cohort portrait and messaging decision in front of the stakeholder to react to, and reworked it quickly in a matter of days. The value, as she put it, “was a faster decision.”
This is an example of what 2x means for my team. Or as one researcher said: “Our 2x isn’t twice the output. It’s the same hours spent a level up.“ And that “level up” looks different for different people. For some, it’s spending more time socializing insights, getting the right evidence in front of the leaders and into the rooms where decisions actually get made. For others, it’s rolling out an internal tool they’ve built and driving its adoption across the business so it genuinely improves how our customers succeed with Fin. Intelligently reinvesting our capacity is how we’ll have the greatest impact.
I find this genuinely encouraging, as the team got there on their own. The trap with any productivity gain is that the time you free up quietly fills back up with more of the same, new habits don’t get encoded. The fact that the team can see that trap, and name it, is worth more than any number of skills shipped.
The honest part
I’d be misrepresenting the survey if I made it sound like pure upside.
The velocity has a cost. One data scientist shared how it’s now harder to find time for deep focused analysis because the constant context-switching and the raised expectation of speed pull against it. She’s also spending more time verifying findings – hers and her stakeholders’ – because AI will confidently over-interpret data if you let it. Faster is not automatically better, and she’s right to say so.
There’s an identity cost too. When someone with years of craft behind them tells you they no longer know what their title should be, it’s a real and unsettling thing to sit with, and pretending otherwise helps no one.
And there’s a quieter counter-current I didn’t anticipate. One researcher described research as having “a grounding moment, coming back to basics” where the most valuable thing is cutting through the noise and helping the business focus on what matters. In a year that’s been all about acceleration and output, the instinct to slow down and ask “what’s worth knowing” might be the most strategic move of all.
Where this leaves us
These jobs aren’t going away, not from where I’m sitting. But they’re being rewritten in real time by the people doing them.
When I wrote about all this a few months ago, it was pure conviction. A thesis I believed and a vision I’d set for the team, but hadn’t yet watched play out. The survey is the closest thing I have to proof. It’s exciting to watch my team adapt to these new ways of working, designing and executing valuable projects we could have never accomplished before.
I won’t pretend we’ve arrived. We’re still working out how to keep self-serve trustworthy as more people across the company query the data directly, how to protect time for the deep questions when everything rewards speed, and how to say clearly what we’re going to stop doing. And we’re still pulled into the middle more than I’d like, because reactive work is genuinely hard to quit.
I asked the team one last thing: what they’d want people outside it to understand about this moment. I’ll leave the final word to one of them, because it’s better than mine. The next wave of value, he said, is “making what we create compound.”