The Groupon Experiment

Remember Groupon? In 2010 it was the fastest-growing company in history, was treated as a new model for local commerce, and even declined a $6 billion acquisition offer from Google.

Sixteen years later, Groupon’s market capitalisation is roughly $700 million. The company is now testing whether a listed legacy marketplace can revive itself by moving operational work into AI agents, behaving as though human execution has become the constraint.

The architect of the experiment is Dušan Šenkypl, chief executive of Groupon since 2023. He arrived from Pale Fire Capital, the Czech investment firm that is Groupon’s largest shareholder. Before Groupon, he co-founded Slevomat, a Czech marketplace he repositioned away from deep-discount daily deals towards higher-value local experiences. That was the operating idea he carried into Groupon, and between 2023 and 2025 some Central and Eastern European markets began to show traction under it. In late 2025, the strategy changed shape when Šenkypl announced Project Foundry.

Project Foundry is Groupon’s name for a company-wide move to embed AI agents into the work of every function. Moved work includes merchant outreach, deal creation, review analysis, internal prioritisation, and parts of the operating rhythm by which the company decides what needs attention. In plain terms, AI is moving close to the parts of Groupon that find supply, shape offers, and direct effort.

This places the technology closer to the core business than starting with customer support, marketing copy, or back-office triage. Šenkypl has put the revenue side of the business inside the AI programme, and he has made much of the philosophy visible.

Why Project Foundry Became Plausible

Project Foundry is interesting because Groupon is placing still-contested LLM capability inside the revenue operation, rather than limiting it to productivity support around the edges. Šenkypl does not give the impression of a chief executive stumbling into AI theatre, which means the choice warrants an organisational explanation.

Project Foundry implies a diagnosis: the constraint was no longer merely tooling, process, or cost. It was the pace at which the organisation could form, test, and execute commercial judgement.

Šenkypl had already spent two years signalling a harder performance culture and publicly recruiting for a narrower kind of operator, the 'full-stack entrepreneur'. The continued emphasis on this profile suggested the organisation had not yet acquired the operating cadence his strategy required. If Groupon cannot hire or retain enough of these operators at the required pace, substitution becomes easier to justify.

The Marketplace Test

The cost case is straightforward. Automated execution appears to lower the marginal cost of outreach and analysis, and the share price suggests investors may be pricing some version of the cost-reduction story. But merchant economics test that logic. Voice agents booking meetings are useful only when those meetings convert into deals where the merchant makes money and the customer returns at full price. Volume in the funnel without conversion at the bottom risks replicating the original Groupon problem, just more efficiently.

Demand conditions reinforce the point. International markets, particularly in Central and Eastern Europe where Slevomat-derived experiences have taken root, are growing. North America appears more exposed to a decade of inbox-fatigue and predatory deal economics. The operating model is the same in both regions, but the underlying brand conditions differ in ways that AI agents do not automatically resolve.

The Boundary of Delegation

The first consideration is what Groupon is asking the system to carry. In December 2025, two months before Project Foundry's most public communications, Anthropic published an internal study on how its own engineers and researchers use Claude. While Anthropic’s internal use of Claude by elite technical staff operates in a different setting than Groupon’s commercial-agent deployment, it provides a useful boundary test.

The headline finding: most employees report being able to 'fully delegate' only 0 to 20% of their work to the tool. Claude is described as a 'constant collaborator', and using it 'generally involves active supervision and validation, especially in high-stakes work.'

The study describes a clear boundary. Engineers delegate work that is 'easily verifiable, where they can relatively easily sniff-check on correctness', low-stakes, or boring. They keep 'tasks involving high-level or strategic thinking, or design decisions that require organisational context or taste'. The survey showed the smallest productivity gains in design and planning tasks, because 'these are tasks people tend to keep in human hands.'

Outbound merchant prospecting falls safely within this boundary, as does review classification. Deal generation probably fits too, depending on supervision. However, other parts of Project Foundry are harder to place. Strategic thinking about which merchants to pursue involves judgement about brand fit, regional dynamics, and the long-term shape of the platform that historical data is unlikely to capture fully. It is difficult for an outside reader to assess this, as Groupon has not disclosed how the work is actually supervised.

What Disclosure Can See

Groupon is not hiding AI from its formal materials. The 2025 10-K describes AI-ready search and checkout, internal tools, and an AI-native operating culture. It notes that AI fluency factors into performance evaluation, and in March 2026, Groupon announced a board-level Artificial Intelligence Committee.

Although this exceeds the disclosure many companies offer, it leaves the movement of judgement largely unaddressed.

Filings can describe AI as infrastructure, risk, oversight, productivity, and customer discovery, along with committee structures and risk factors. They are not designed to reveal whether the leadership team is forming the view or receiving a view whose first formation has already happened outside the meeting.

Naming the tool in a filing is different from capturing how thinking might be shifting from the organisation towards the person most able to use it at speed.

My own recent testing points to the same tension from another angle. In the Subtraction Study, I tested the same model with and without a simulated enterprise governance prompt. The governed version became safer, highly qualified, and less willing to take a definitive strategic position. This is a common effect of formal deployment: it makes the tool more acceptable to the institution, while reducing some of the force that made the tool useful to the individual operator. Groupon may encounter a similar split, where the chief executive’s direct use of AI, the board’s governed view, and the workforce’s operational experience pull in different directions.

Concentration Without the Founder Mechanism

A public company typically distributes judgement across a chief executive, a leadership team, a board, reporting lines, and friction. When the people around a CEO cannot move at the required speed, the CEO will naturally reach past the leadership team to test ideas, pressure-check decisions, and draft strategy. If the alternative is a tool that produces useful work at any hour and does not introduce the same negotiation costs as human colleagues, the reach towards the tool becomes rational.

Fewer people are involved in forming the view, while more execute it once formed. The CEO acquires a thinking partner outside the leadership team, if that is how the system is actually being used, one that lacks human institutional memory and does not have to live with the workforce after the decision is made.

The same move has been made before by founder-led private companies, where one dominant individual makes consequential decisions with minimal input. Founder structures rely on equity concentration, narrow accountability, and the founder's reputation carrying the workforce. A turnaround CEO operating inside the disclosure regime of a listed company, with institutional shareholders and an independent board, lacks those specific mechanisms. The analogy is imperfect because listed-company accountability imposes formal constraints that private equity concentration bypasses.

The mechanism by which concentration becomes a problem would show up as governance failure (the board unable to challenge decisions taken before the meeting), workforce disengagement (people executing conclusions they did not help reach), or error amplification (the absence of dissent allowing one wrong reading to propagate unchecked). None of these outcomes are confirmed, but they remain plausible.

The safer corporate language would describe this as augmentation, even when the boundary between augmentation and substitution is becoming harder to maintain. Šenkypl has published his operating philosophy instead. This honesty is unusual, but it limits the option to retreat into the language of augmentation should substitution begin to strain the institution.

The market can price cost reduction and operating leverage, but it is less equipped to price what happens when judgement moves away from the organisation and towards one person with a tool.