My first study at HelloFresh was launched on top of a dumpster.
With no lab, no budget, and no mandate, I found the one quiet corner of the office, set a laptop on the nearest flat surface, and ran the session. The flat surface was a dumpster lid. It worked, not because the setup was good, but because the bar to start was finally low enough that we started.
I bring up this experience because it taught me the lesson I keep relearning at every scale of research: research rarely fails because the methods are wrong. Research fails because the team that needs the answer can't get it fast enough, so they decide without it. Every research leader I know is fighting the same fight: more questions than the team can possibly answer and a growing pile of decisions being made on instinct because the queue is too long.
For most of my career, the obvious response was to do more: cook faster, run the kitchen harder or ask for headcount. While those instincts are natural, they are wrong. But they have taught me vital lessons that have reshaped the way I view research teams.
A research team is a restaurant
Picture your research function as a restaurant. The researchers are the chefs, and the stakeholders are the customers. Every project is a dish, and the quality of the dish is determined by your ordering customers. The food has to be good, and it has to be right for the person eating it.
A great restaurant also has a hard ceiling. One kitchen can only seat so many people, and great chefs are expensive and rare. I spent a long time believing the answer was a bigger kitchen: hire more chefs, take more orders, and push out 100+ projects per year. It scaled until it didn't, and when it stopped scaling, the bottleneck wasn't the tooling or the talent. It was me, insisting that every dish come out of the one kitchen I controlled.
Franchising, not gatekeeping
The move that actually changed this pattern was franchising.
Take the dishes you've made a hundred times, the studies with tried-and-true recipes, and let other teams cook those themselves using your templates and your guardrails. A franchise will never match the flagship. McDonald's isn't a Michelin star and never pretends to be. But when someone needs a fast, reliable, decent meal, the franchise serves a job the flagship never could in 1,000 places at once.
I resisted delegating for years. My background is in research, and rigor is near and dear to me, so the idea of handing a survey to someone untrained felt like inviting food poisoning into my own kitchen.
What changed my mind was a simple observation: gatekeeping research doesn't protect the craft so much as it protects the bottleneck. When you're the only one who can cook, every order routes through you, and you become a service desk instead of a strategic partner. The work that actually needs your expertise sits in the queue behind a hundred requests that don’t.
So we franchised. But the order of operations matters more than anything else I can tell you here. Democratization without guardrails is lower-quality research with more authors, which is worse than no research at all because now bad data carries the credibility of a study. The guardrails come first: a vetted question bank, templates for the common studies, researcher review before anything goes live, and office hours for anyone who gets stuck.
One discipline matters above all the others: document your menu. Write down what dishes you've served, when you served them, and what you found as a catalog of what you ran and where the results live. In my experience, that catalog turned out to be the single most valuable asset we built.
Research started showing up in three times as many product decisions, and risk on customer-facing decisions dropped. We advised on hundreds of studies and generated thousands of insights across dozens of markets, without proportionally growing the team. The chefs stopped running every order and started working the line that actually needed them: the hard, ambiguous, high-stakes studies that no template can cover.
Scaling research is not dependent on hiring more chefs. It's about deciding which dishes can be franchised and protecting the few that can't.
The AI layer: from in the loop to on the loop
AI changes where the researcher stands in the kitchen. My graduate training was in human factors, and there's a distinction in that field we used to obsess over: being in the loop versus on it. When you are in the loop, you perform every step by hand. When you are on the loop, you supervise the process and step in when it drifts. It's the difference between driving the car and watching the car drive while keeping your hands near the wheel.
For a quant researcher, the prep work used to consume an enormous portion of the research workflow. Clustering open-ended responses, running factor analysis, and hunting for the relationships hiding in a dataset; all of it meant writing scripts and losing evenings. A lot of that now runs in minutes. That doesn't make the researcher obsolete; it moves the researcher from line cook to head chef. You taste, you check, you send back what isn't right.
This is the only way franchising survives at scale: the recipes get smart enough to catch the common mistakes before a dish leaves the kitchen, and the expert oversees quality control instead of making every chop by hand. AI won't replace the chef.; it replaces the chopping. The judgment over whether a finding is good enough to put in front of leadership is still yours, and it always will be.
Being honest about the limits, AI is genuinely strong at the first pass or grunt work and genuinely clumsy at judgment. It will confidently surface a theme and quietly drop three others. Used carelessly, AI produces summaries that sound authoritative and don't survive a close read. The researcher on the loop is not a formality. It's the entire reason the output is trustworthy.
The Sprig layer: the franchise kitchen as software
I spent years building the franchise kitchen by hand. I built question banks in a wiki, templates in a folder, and a review process held together by Slack and goodwill. It worked, but it was held together with tape.
Now, I work at the company building that kitchen as software, and the parallel is impossible to ignore.
The thing that always made franchising risky was consistency. A recipe card only gets you so far when people skip steps, swap ingredients, and write the exact leading question you warned them about. What you actually want is a kitchen that holds the standard for them.
That's what the agent model does. At Sprig, one agent writes the recipe with the rigor built in, handling neutral phrasing, logic, branching, and bias checks, so a non-expert doesn't ship a broken study. Another agent runs the study and adapts it to each diner in real time, asking natural follow-ups instead of pushing everyone through the same static form. A third agent plates the raw results into a report you'd actually put in front of leadership with the evidence attached.
It’s true that an agent will still occasionally draft a question I'd send straight back to the kitchen, which is precisely why, by design, a researcher always stays on the loop. But the agent model is the closest I've seen to successfully franchising quality research. The guardrails I used to assemble by hand are becoming integrated into the kitchen.
The dream has never been fewer researchers; it is letting more people eat well, while the chefs spend their time on the dishes that actually deserve them. The dumpster experiment taught me to lower the bar to start, franchising taught me to scale without losing the recipe, and the agent model is how we finally do both at once.
.png)
.png)