The consensus AI trade in market research is disintermediation.
Startups pitch brands on synthetic respondents, AI-moderated interviews, and DIY insight platforms, built on the assumption that research agencies are legacy middlemen waiting to be routed around. Listen Labs raised $69 million in January 2026 on a version of that thesis.
Metaforms raised $11.3 million on the opposite bet: the agency survives, and whoever runs its back office captures the value.
The platform now runs nine agent modules across the research chain: questionnaire design and review, programming, QA, change management, data validation, open-end coding, tabulations, and reporting.
Where an agency’s existing platform is GUI-based, Metaforms operates through a Chrome extension using the programmer’s own login, another set of hands inside software the team already trusts. Where the system is code-based, it replaces the IDE directly.
The market rewarded the contrarian position quickly.
Within six months of commercial launch, four of the world’s top twenty research players had signed, including Dynata, GLG, Konovo and Purespectrum. The platform now processes more than 1,000 surveys per month for some of the largest consumer brands in the world.
That last number is the one worth sitting with, because CEO and founder Akshat Tyagi refuses to let it be sold as proof of anything permanent.
“At our stage, 100% expansion is a fact about the customers we have today, not a law of nature,” he says.
A founder who corrects his own best statistic before anyone else can trust the mechanism underneath it more than the headline.
The mechanism is what this piece is actually about.
The sexy product was the wrong product
Metaforms did not start here.
The company launched as WorkHack, an AI-native form builder competing in Typeform’s territory, a product with everything venture logic celebrates: a horizontal market, product-led growth, Product Hunt launches, a total addressable market measured in every business that collects data online.
It also had no wedge.
A horizontal form builder competes with every survey tool, every no-code platform, and eventually every LLM that can generate a form on request.
The bigger problem was structural rather than competitive: no single user of a horizontal tool has enough pain to anchor a company, because the product is shallow by design. It has to be, to serve everyone.
Shallow products lose to depth the moment a better-funded competitor or a smarter model shows up, because nothing about the relationship is sticky.
Market research looked comically narrow by comparison, but that narrowness was the entire point.
A questionnaire is not a form. It is the opening gambit in a long, revenue-critical workflow: survey programming, validation, sample coordination, quota management, data processing, reporting.
Insert yourself inside that workflow and ripping the product back out costs the agency more than keeping it does.
Tyagi’s own summary of the trade cuts against conventional startup wisdom about market sizing: “TAM is almost irrelevant before you have earned control over a painful workflow. We stopped asking how many people could theoretically use the product and started asking who would be genuinely worse off without it.”
Demand Was Never the Problem
Every AI pitch in this category starts from the same premise: agencies are dying middlemen, starved for clients, ripe for replacement.
The premise doesn’t survive contact with the real-world in terms of how agencies actually operate.
A firm turning away paying work is not a firm without demand.
It is a firm without throughput, and while it might not seem obvious, those are different diseases with different cures.
A single study can consume days converting hundreds of questions into survey code, coordinating five to ten panel vendors by email, and writing validation scripts by hand.
None of that work requires judgment.
All of it requires a skilled researcher’s time, which means the agency’s real constraint was capacity, not client acquisition.
Tyagi’s own framing of what he found once he looked past the industry’s polished front end: “The industry presents itself through the language of insights and strategy, but behind every insight is a fairly complicated factory.”
Attack the acquisition problem and a startup has to rebuild trust, methodology, and client relationships from nothing.
Attack the throughput problem and the incumbent’s trust becomes the startup’s distribution, free of charge.
The fast car in slow traffic
Speeding up one step in a slow chain does not speed up the chain.
A typical research project runs roughly 120 hours from questionnaire to final report, and survey programming, the wedge Metaforms entered through, accounts for only 8 to 16 of those hours.
Make that single step instant and the project barely moves, because the bottleneck was not programming in isolation.
One agency head described the effect as driving a Ferrari through gridlocked traffic in Times Square.
The car works. The street doesn’t.
That mismatch is the real argument for owning nine modules instead of one.
Programming speed exposes QA as the next constraint. QA speed exposes change management as the constraint after that.
A wedge earns entry into the workflow. It does not, by itself, earn the outcome.
The same mismatch cuts the other way, and this is the detail that separates a company paying attention to its own product from one selling a demo.
During an early pilot, a junior programmer’s Metaforms-generated code needed enough correction that a senior programmer had to step in and explain the fix.
The programmer running that pilot did the arithmetic out loud: “Now it’s two people working and AI working. And we could have done it all with just one person. No AI. If it’s partially correct, somebody still has to review it. Reviewing might take as long, and now it costs us twice as much.”
She was right, and the fix that followed says more about the company’s discipline than any number in a pitch deck would.
Metaforms now generates code in batches of three to five questions instead of the whole survey at once, so a programmer corrects each batch before the agent continues, and the agent adjusts before generating the next one.
An error in an early question stops spreading before it reaches the tenth. By the end of a complex survey, the output has converged close enough to that specific programmer’s style that review becomes skimming instead of rewriting.
The rule this produced is unforgiving in the right direction: if reviewing the agent’s output takes less time than doing the work by hand, the product earns its place.
If review takes just as long, the company has more work to do, not the customer.
Selling to the disrupted
Incumbent research agencies own three assets no model output can generate: client relationships built over years, methodological credibility that survives an audit, and accountability when a study informs a nine-figure decision.
Selling around the agency means building all three from zero before a single enterprise contract closes. Selling to the agency means inheriting them on day one, fully formed.
The economics compound in the agency’s favor, too.
Compress turnaround time and cut operational cost, and the agency starts taking on clients it previously turned away, from startups testing a first idea to global brands running multi-country trackers.
That makes the buyer an enthusiastic adopter rather than a threatened one, the opposite of what the synthetic-respondent pitch asks of a brand’s insights team, which has to bet its own credibility on data with no human in the loop.
Tyagi’s version of the bet: “We would rather give the strongest agencies an unfair advantage than spend a decade trying to recreate their distribution and client trust from scratch.”
What expansion actually rewards
Most workflow automation sell on cost.
Metaforms’ newest move sells on category instead: AI-moderated voice interviews, offered to agency customers as a way to compete directly with Listen Labs and the wave of AI-research startups going after the same budget line.
That is a narrower bet to build toward and a much harder one to prove, because a cost argument has an obvious ceiling – hours removed, headcount avoided – while a category argument has no ceiling and a much higher bar for evidence.
Metaforms is not just automating a step an agency already does. It is handing agencies a way to sell a research method they don’t currently offer, using a budget that would otherwise go to a standalone vendor.
Tyagi is careful not to oversell the substitution: “We’re not telling agencies to replace their moderated research with this. We’re giving them a way to say yes to the client who was never going to pay for a moderator in the first place. That is new revenue, not cannibalized revenue, and we will know we are right when we see it show up as a new line item, not a swapped one.”
The restraint is in itself informative. It suggests a company confident enough in its metrics to let them arrive on schedule instead of front-loading the pitch.
The expansion pattern follows the same logic.
Each agent, once trusted, sits next to the workflow that comes right after it: survey programming feeds into data validation, which feeds into quota tracking, which feeds into reporting.
Voice moderation is the first agent that doesn’t sit next to an existing workflow – it opens a new one, and a new revenue line for customers who adopt it.
The product architecture created the expansion before; here it’s creating the category. The sales team still just shows up for it.
No self-serve, on purpose
Metaforms cannot be trialed from its website.
Every customer goes through a demo, an onboarding process, and configuration against the agency’s existing stack, a full renunciation of the product-led motion the company was originally built on.
Part of the reason is straightforward liability: a broken survey link or a corrupted dataset damages the agency’s relationship with its own client, immediately and specifically, and buyers facing that much downside demand verification before commitment.
But a sharper reason sits underneath the liability argument.
Every research company runs on a house style that lives mostly nowhere documented: habits, reviewer preferences, prior corrections, conventions and tribal knowledge an experienced programmer follows without being able to fully explain them.
A generic model can produce output that is technically correct and still gets rejected on sight, because it doesn’t match standards nobody wrote down.
Internally the company calls this an immune response, and no accuracy score talks a rejecting programmer out of it.
That is why onboarding now runs for days before a pilot even starts.
Real questionnaires, template files, prior projects, and standard-practice documents get converted into customer-specific rules before a single agent output reaches a live workflow.
Every correction a programmer makes afterward becomes a captured rule instead of a one-off fix, so the next programmer on the next project inherits judgment that used to live only in one person’s head.
The company reinforced the posture with a Mumbai office staffed entirely by research domain experts, treating industry knowledge as core infrastructure rather than something bolted on after the models were built.
As Tyagi puts it, “In a production research workflow, those edge cases are the product,” which reframes what most AI teams write off as noise into the actual substance of the job.
Betting on the outcome, not the method
Listen Labs’ entire thesis is a direct threat to the premise Metaforms has spent two years building on: that a human-run agency layer is worth arming rather than replacing.
If synthetic respondents and AI moderators keep improving, some of the workflow Metaforms automates disappears regardless of how well it’s automated.
Most founders facing that question reach for a confident dismissal.
Tyagi doesn’t. “It might shrink in certain areas,” he says. “We should have enough epistemic humility to admit that nobody knows the exact endpoint.”
What he holds fixed and constant instead is the outcome rather than the method: agencies remain the accountability layer for research quality even if the underlying technique shifts entirely, which means Metaforms’ job is to help them deploy whatever method wins, not to defend the one currently in place.
It’s a harder position to hold than a clean rebuttal would be, because it requires rooting for the method most likely to obsolete part of the company’s own roadmap.
What killing a company actually costs
Pivot stories get told, almost by default, as a single decisive moment: the founders saw the data and made the hard call.
Tyagi’s account of killing WorkHack refuses that shape entirely, and the refusal is more instructive than the tidy version would have been.
“There was no single cinematic week in which we bravely killed the old company,” he says. “We probably continued with the horizontal product for longer than we should have because the team, our users and our own identities were attached to it. The cost was code, momentum, morale and ego. The hardest cost was admitting that our original insight had been directionally wrong even though the product had users.”
What he draws from that experience is a sharper distinction than most founders bother to make: killing a product with no evidence behind it is easy, and killing one with partial evidence, some users, some traction, a team that believes in it, requires an entirely different order of judgment.
It wasn’t hard to spot the exit. The real challenge was telling the difference between evidence that a product could survive and evidence that it should define the company.
The efficiency you don’t want to name
Every vertical AI pitch in this genre reaches for the same reassurance: automation reallocates human effort, it doesn’t eliminate it.
Metaforms leans on that framing too, and most of it holds up.
But Tyagi is unusually willing to admit, on the record, where the reassurance runs out.
On one rollout call, a customer executive told him this would be the first tangible example his employees could point to and say, “That tool came in, and those people went out.”
A colleague added that it would be the first time the company had to stand in front of its own staff and admit AI adoption had directly cost jobs.
Tyagi’s account of that moment carries zero spin: “I do not have a clean answer to this. But I do have a position.”
The position holds that margin pressure in research agencies is real and will keep pushing toward automation regardless of how any one vendor frames it, and that the honest response is to take responsibility for the transition rather than narrate around it.
When survey programmers started growing afraid of the company’s own product, the response was not a messaging exercise. Metaforms built a training and certification program aimed at the programmers most exposed to the change, designed to make them employable in the workflow the industry was moving toward rather than the one disappearing under them.
It does not resolve the underlying tension.
A company selling automation to agencies under margin pressure is, by definition, part of what applies that pressure. But refusing to look away from that fact, while most of the category recites the reallocation line without examining it, is its own kind of differentiation.
Tyagi calls the underlying principle “not looking away,” and says it’s the one lesson he would keep if he had to give up every other one the company has learned.
Leverage defeats replacement
Metaforms’ trajectory suggests a rule for any founder staring at an industry the AI narrative has already marked for death.
Before building the replacement, check whether the incumbent’s real constraint is operational or existential.
Agencies turning away demand are throughput-limited businesses, and throughput problems are software problems, not extinction events.
Secondly, check whether automation grows the incumbent’s revenue or threatens it.
Buyers who get richer from a product expand at rates buyers who feel hunted never do.
Tyagi built the company by running exactly that test in reverse against a category everyone else had already declared dead, and the diagnostic he reaches for now is more precise than most outside frameworks.
“Ask what exactly is dying. Is customer demand disappearing? Is distribution changing? Are margins compressing? Or is an old production method being replaced? Quite often, the old way of producing an outcome is dying while demand for the outcome remains strong. That can be an exceptional place to build.”
The framework has a second half Metaforms is careful not to drop: this is not an argument for defending incumbents on principle. Some agencies will refuse to modernize and will lose their share to the ones that do.
The goal though was never to preserve every existing player. It was to make the best operators radically more productive and let them take share from the rest.
The roadmap from here, voice-based research, automated reporting, a target of 100,000 surveys processed per year, doesn’t read like expansion for its own sake, it is clearly about the same bet compounding.
In a market where every competitor’s pitch deck promises elimination, arming the incumbent turned out to be the harder position to hold and the more durable one to have taken.
