Gary Ellis has spent much of his career near rooms full of opinions. Some belonged to scientists working through data. Some belonged to voters, donors, campaign staff, customers, employees, and research participants. The furniture changed. The recurring problem did not: how do you hear enough people to make a decision without sanding every interesting answer into a checkbox?
At Remesh, the company he co-founded and now leads, that problem became software. A moderator asks an open-ended question to a large online group. People type their own answers. Then they see samples of other answers and vote on what resonates. The platform uses those reactions to expose patterns, points of disagreement, and segments while the discussion is still alive. A researcher can notice the odd remark that suddenly finds a constituency and ask the next question before everyone leaves.
It is a curious hybrid: part focus group, part survey, part town hall with a very diligent clerk. The focus group preserves language and surprise but struggles with scale. The survey handles scale but can count only the possibilities its author thought to list. Remesh occupies the middle, where hundreds of people can speak freely and still produce something more structured than a transcript the size of a paving stone.
Before software, there were constituencies
Ellis studied political science at the University of Rochester. His early professional record crosses two fields that look unrelated until one notices their common appetite for evidence. At Cleveland Clinic, he worked in scientific research and with large datasets, contributing to published studies. In politics, he served in organizing, finance, and campaign-management roles, including work for Lee Fisher in Ohio and for council and mayoral campaigns in Washington.
A lab and a campaign office use different nouns, but both can punish a lazy reading of the room. A correlation is not an explanation. A loud supporter is not an electorate. The useful question is often buried beneath the first answer, and the cost of missing it appears later, when a result fails to reproduce or a voter declines to behave like the spreadsheet.
Remesh itself began with a civic ambition. Co-founder Andrew Konya, then working through problems in theoretical physics and complex systems, wondered whether software could help a large group express a collective voice. The idea grew through a prototype and Cleveland's Flashstarts accelerator. Aaron Slodov, a former Google engineer, became the second full-time co-founder. The team had mathematics, code, and a very large mission. What it needed next was a commercial use narrow enough to pay an invoice.
Ellis arrived as the third full-time co-founder during the 2015 Barclays Accelerator powered by Techstars. He was the former political operative entering a company born from a question about representation. The fit was not subtle. Remesh pivoted toward market research, where businesses already spent money trying to understand groups and regularly confronted the same tradeoff between conversational depth and numerical scale.
The operator's long apprenticeship
For nearly a decade, Ellis oversaw Remesh's revenue and operations. This is the quieter half of a startup origin story. The founding concept gets the charming anecdote; the operator gets a forecast, a hiring plan, and a customer who would like the security questionnaire returned by Friday. Turning a collective-dialogue experiment into enterprise software meant finding repeatable buyers, proving the method inside existing research workflows, and making the strange mechanism easy to explain.
The company graduated from the accelerator, raised institutional capital, and expanded from live conversations into asynchronous studies, video, recruiting, multilingual work, and research services. Its clients have included large consumer brands, financial institutions, consultancies, and research firms. The platform's appeal was concrete: collect open language from many people, quantify which comments resonate, and let a moderator follow the signal without waiting weeks for coding and analysis.
Ellis also gave the company's argument a name: “representative intelligence.” In a 2018 essay, he contrasted active conversation with the passive trail of clicks, purchases, and visits that companies routinely treat as a portrait of a person. Behavior matters, but it does not volunteer its motive. The neglected material is everything someone might say about an experience, relationship, fear, or ambition if a business bothered to ask. The phrase was grander than a feature list, yet the product lesson underneath it was plain. Observation can show what happened. Conversation creates a chance to learn why.
Ellis became CEO after leading the operating side of the business, while Konya moved into the role of Chief Science Officer. In announcing the change, Ellis called the appointment “the honor of a lifetime” and credited his co-founder for preparing the company for its next phase. The division of labor suited Remesh's nature. One founder would keep pressing on the mathematics and technology. The other would be accountable for making the method useful, durable, and legible to customers.
“People are not made of a series of clicks, purchases and website visits.”Gary Ellis, writing on representative intelligence
A very qualified yes to AI
Ellis is an AI executive with a conspicuous fondness for caveats. His argument is not that researchers should refuse automation. It is that the value of automation depends on keeping the evidence and the decision-maker in view. Large language models can summarize open responses, suggest classifications, translate material, and clear hours of clerical work from a researcher's desk. They can also produce fluent errors, inherit bias, and encourage a hurried reader to mistake a coherent paragraph for a verified finding.
“Summaries are not insights,” he has said. The sentence is short enough to survive a product launch. A summary compresses material. An insight connects evidence to context, judgment, and a decision. Remesh's approach has been to use its participant-voting data to identify representative responses before an LLM prepares a summary, while still giving researchers access to the underlying material. The machine supplies a first pass. The professional remains responsible for the last mile and, crucially, for deciding whether the road was sensible in the first place.
Where the conversation sits
That distinction has become sharper with synthetic respondents. Generated participants promise quick, inexpensive answers without recruiting anyone. Ellis accepts that synthetic data can be useful, but rejects the leap from useful simulation to replacement for observation. A model confronted with something new does not raise its hand and confess bewilderment. It completes a plausible pattern. The prose may be orderly. The opinion was still not held by a person.
His warning lands because it comes from someone selling AI-assisted research, not defending a stack of paper questionnaires. He wants the machines in the workflow. He also wants real people at the origin of claims about what real people think. “Synthetic data is powerful,” he wrote in 2026. “But it's not a substitute for reality.” The aspiration is amplification: give researchers enough leverage to support far more decisions while preserving methodological rigor.
“The future is AI amplifying our ability to understand people, not replacing the need to actually listen to them.”Gary Ellis on synthetic participants
The useful thing to steal
Remesh contains a product principle more portable than market research. When users create a flood of unstructured material, invite them to perform one small act of organization. A participant does not need to code a transcript or write an analyst's memo. They only need to say whether another response resonates. Thousands of modest judgments can reveal an outline that no single participant had to draw.
This works because the structure arrives after expression. A conventional multiple-choice form edits the world in advance by deciding which answers are available. Remesh lets people speak first, then adds measurement around what they actually said. For community tools, review platforms, workplace systems, and knowledge products, the sequence is worth borrowing: collect freely, ask for lightweight judgments, aggregate carefully, and keep an operator close enough to investigate the surprise.
There are boundaries. Agreement is not truth. A badly recruited audience remains a badly recruited audience, no matter how elegant the visualization. Popular responses can bury an important minority view. A weak discussion guide can industrialize a weak question. In employee research, anonymity and the employer's willingness to act matter as much as the interface. Ellis's public emphasis on transparency and expert oversight is, in part, an admission that software cannot make those choices disappear.
His career has now looped back to the old political question in a new technical costume: who was heard, how were their words interpreted, and what will the people with power do next? The answer cannot be outsourced entirely to a model. Nor does it need to be assembled by hand at the speed of a committee reading transcripts.
Between those extremes sits Ellis's wager. Let the machine handle abundance. Let the crowd help reveal its own shape. Let the researcher inspect, question, and decide. It is less theatrical than announcing the end of human judgment. It is also a workable reason to invite everyone into the room.