"In the age of AI, the future belongs to small teams." The line is well on its way from bold claim to unquestioned consensus. The logic is seductive: if a tool can do the work of ten specialists, you no longer need the apparatus that used to coordinate those specialists. The small, fast team beats the lumbering corporation. Five people with the right models do what took a department yesterday.
It is a good story. It just is not what the data shows, and even less what it shows unambiguously.
The honest answer is less comfortable and far more useful: whether small wins is context-dependent, and the conditions are more specific than the Linked consensus lets on. In several places the answer is even "we don't know yet." Draw the wrong lesson from the right observation, and you build your organization on an exception while treating it as a rule. This piece takes the thesis seriously, because there is something to it, and then takes it apart, because that is what decides the right call.
What the German evidence complicates immediately
If the thesis were broadly true, small companies would use AI more widely and deeply than large ones. The opposite holds. In Germany, large firms deploy AI across more functions. In customer service, for example, 43.9% of large AI users apply the technology, against 36.4% of small AI users[1]. The same pattern runs through marketing, sales, and management: more resources, more data, and more processes mean more places where AI pays off at all. Size is not the obstacle to AI use; it is often the precondition for rolling it out broadly.
The starting point is also soberer than the debate suggests. Only about 22.5% of German firms describe themselves as established AI users[2]. AI is not a universal reality yet but an unevenly distributed capability, and capability is expensive to acquire. The large majority is at the start line or outside it, and that is true for small and large firms alike, just not equally.
At the macro level, the expected productivity gains are real but unevenly spread and heavily dependent on execution skill[3]. An IW expert report makes the caveat explicit: whether a technical possibility becomes a macroeconomic effect is decided not by the model but by an organization's ability to embed it[4]. That is the first crack in the thesis. It assumes AI is a universal leveler. The German numbers show an amplifier that works hardest where capability already exists.
None of this kills the thesis. It reframes the question: not "do small teams win?" but "under what conditions do they win, and when do large firms keep winning?"
The productivity myth: content wins, decisions lose
The most important finding in current research is the one the thesis leaves out: the AI effect is not uniform. It depends on the kind of task. The thesis treats "productivity" as a single quantity. The evidence treats it as at least two.
A systematic meta-analysis across many experiments reaches a sobering conclusion: combinations of humans and AI perform, on average, worse than the better of the two alone. Losses show up above all on decision tasks, gains on content creation[5]. That is not a footnote, it is the crux. Exactly where small teams benefit most visibly, in writing, designing, and producing, the effect is largest. And exactly where scaling gets hard, in complex, coordinated decisions, the effect turns negative. A human with AI does not, on average, make better decisions, but often worse ones than if you had let the more competent of the two work alone.
The much-cited productivity numbers deserve more caution than they usually get, too. Estimates land in the range of roughly 14.5% for ChatGPT-assisted tasks and about 17.3% for other tools, but with wide standard errors and many caveats[6]. "AI makes you 15% more productive" is therefore not a law of nature but an average with large spread, one that swings sharply up or down depending on task, context, and skill. A wide standard error means, concretely: for some tasks and some teams the effect is large, for others it vanishes, for others still it is negative. Build a strategy on the mean and you build it on a number that may not describe your team at all.
There is more: AI changes not only the task but the conditions under which a team works. Generative AI shifts how teams form, how they share knowledge, and how much experience they still build themselves, with long-term effects on skill and trust still open[7]. A productivity gain today can be a competence loss tomorrow, if the team unlearns what the AI takes over.
The fair reading, then: AI is a powerful lever for content-driven knowledge work with low capital needs. That is a genuine advantage, but a narrow one. It carries a content boutique, a design unit, a small engineering team. It does not automatically carry a company that has to coordinate capital, data, and regulated decisions, and least of all the decisions themselves.
Capital, data, and market access: why "small + AI" rarely suffices
The thesis systematically underrates what scale actually does. Scale advantages are not just cost advantages. They are access to data, to capital, to regulatory resilience, and to market power, and none of those vanish because a model writes better copy.
Start with capital and time. The road from small team to viable business is long. German AI startups take a median of more than four years to reach revenue, and employment growth correlates with technological breadth, not with leanness[8]. That is doubly awkward for the thesis. First: four years without meaningful revenue is survivable only for those with capital or the ability to raise it, precisely not the advantage of the lean bootstrapped team. Second: those who grow do so by adding capabilities and technologies, not by staying small. Breadth, not narrowness, is what growth is associated with.
Then data. Modern models work best where clean, extensive data and clear processes already exist. That is exactly what large organizations own and small ones often do not: mountains of data from years of operation, installed user bases, clean pipelines. In data-intensive industries the data stock itself is the moat, and AI tends to deepen it rather than level it. A small team with the same model but without the data meets a competitor who has the same AI plus the data lead.
Manufacturing SMEs show this especially clearly. AI-based business models there are still plainly limited by data availability, integration effort, and skills[9]. The technology is rarely the problem; the preconditions are, and you cannot download them. And SME productivity overall hinges on structural factors AI alone does not solve, from capitalization through process maturity to market position[10].
The thesis often escapes through platforms: small teams borrow infrastructure, compute, and market access from cloud and app platforms instead of building both themselves. That is real and it works, but it is not a victory of the small over the large, it is a dependency on the very largest. Building on someone else's infrastructure trades capital needs for platform risk: pricing, terms, and access sit with someone else. "Small + AI" then usually means "small + AI + dependency on a handful of platforms," which is a different thesis from the autonomous small team.
Talent and capability as the real bottleneck
If a single factor decides success or failure with AI, it is rarely the tool. It is the people who operate it.
AI skills and the talent to implement them remain a scarce bottleneck, especially in the mid-market[11]. That is the quiet irony of the debate: AI is sold as the answer to the skills shortage, while the skills shortage is at the same time the biggest barrier to using AI well. A tool you cannot operate competently produces activity, not advantage. It generates output that looks expensive and is rarely right.
This bottleneck hits small teams harder, not softer. A corporation can afford a center of excellence, dedicated data scientists, and a training structure. A five-person team cannot; it lives on happening to have the right people aboard. Where those people are missing, the lean structure is not a strength but an understaffed flank. This is borne out across SMEs: AI-based business models fail less often on the technology than on missing integration and implementation capability[9].
A popular counterargument says no-code and natural language make capability unnecessary: if you can talk, you can operate an AI. That is true on the surface and misleading at the core. No-code lowers the bar to building the first prototype. It does not lower the bar to judging whether the result is right, whether the data holds, whether the process is regulated, whether the model produces nonsense in the edge case. That judgment is the real expertise, and low-threshold tools do not replace it, they make it more important. The path of German AI startups says the same: growth comes with technological breadth and the people it requires, not with mere access to tools[8]. Capability is not a head start AI takes care of. It is the bottleneck that remains.
Regulation, security, and compliance
There is one domain where the thesis almost fully inverts: where decisions are regulated, safety-critical, or liability-bearing.
In health care, in finance, and in critical infrastructure, the problem is not producing a piece of content but bearing responsibility for a decision. And this is exactly where the evidence shows the biggest weakness of the human-AI combination: on decision tasks it trails the best single performer on average[5]. In a regulated environment, "on average a bit worse" is not an academic detail but a liability and approval risk.
Compliance also acts as a cost asymmetry favoring the large. Data protection, documentation duties, audits, and the demands of the emerging AI-regulation framework are largely fixed costs. A corporation spreads them across millions of transactions; a small team carries almost the same absolute burden on a tiny base. GDPR conformity, a clean record of processing activities, or a data protection impact assessment cost the small team many times more in relative terms. Regulation is thus one of the few areas where size does not just help but structurally protects.
Security compounds it. Models operating on sensitive data open new attack surfaces: prompt injection, data leakage through third-party services, unclear responsibility in platform chains. A small team building on someone else's infrastructure often has neither the contracts nor the control to manage these risks the way a regulated industry demands. And the labor-market and regulatory context in which all of this plays out is itself in motion, with effects on employment and qualification still open[2]. Where regulation is dense, "small and fast" is rarely the winner. There, the winner is whoever can carry the load.
Coordination and team size: what AI does not automatically solve
The thesis contains an unspoken assumption: that team size is mainly a coordination problem AI automates away. Fewer people, less alignment, more speed. The evidence does not cleanly support that equation.
Even the basic relationship between team size and performance on complex tasks is empirically unclear. An experimental study finds no simple "smaller is better" effect; the impact of size depends on the task and its structure[12]. Anyone claiming smaller teams are categorically more capable is leaning on an intuition the research does not confirm. Sometimes a small team helps, sometimes it simply lacks the capacity to solve a complex problem in a reasonable time.
AI changes this calculation, but not in one direction. It changes how teams form, how they communicate, and which roles they need at all, with effects that look like productivity in the short run and touch skill and trust in the long run[7]. A team can be smaller with AI and still have to coordinate more, because it now orchestrates humans and agents, checks outputs, and catches the errors automation introduced.
And scaling itself stays human. When the prototype becomes a product and the product becomes an operation, exactly those tasks arise where the human-AI combination falters: coordinated decisions under uncertainty, with responsibility and consequence[5]. The growth the thesis promises therefore leads into precisely the terrain where its tool is least reliable. Small is rarely a choice; it becomes a limit as soon as success arrives.
What the evidence base cannot support
Here, intellectual honesty matters more than a strong punchline. Much of the above is the best available evidence, and the best available evidence is thin in several places.
First, selection. Many findings on AI use come from firms that already use AI, or from startups that survived. The failures and the non-adopters are missing. That biases almost every comparison upward and makes it hard to move from "successful small teams use AI" to "AI makes small teams successful."
Second, measurement. Productivity in experiments is not productivity in operation. The available estimates carry wide standard errors and many caveats, and they often measure narrowly scoped tasks under lab conditions, not the messy whole of a company[6]. A cleanly measured effect on a writing task says little about a firm's quarterly numbers.
Third, time. Robust long-run data is missing. Whether today's gains on content-driven work persist, or are partly eaten back by competence loss, saturation, and altered team dynamics, is open[7]. We are watching the beginning of a curve and treating it like its end.
Fourth, causality. Even where correlations look robust, the direction is often unclear. The finding that human-AI combinations do worse on decisions on average is strong, but it does not say every combination fails, it says design decides[5]. And that the effect of team size is empirically unclear does not mean "it doesn't matter," it means "we don't know yet"[12].
The honest consequence: at several points in this debate the right answer is not "yes" or "no" but "the evidence is insufficient." Anyone who hides that is selling opinion as finding.
When the thesis partly holds, and when it does not
The productive question is not whether small teams are the future, but when small is the right shape. The evidence supports a usable checklist. Small wins when several of these conditions hold at once:
- Knowledge-intensive and capital-light. Value comes from skill and output, not from assets, inventory, or data volume. Content, design, advisory, and software-adjacent work fits here.
- Content-driven, not decision-heavy. The core of the work is creating, not owning coordinated decisions under liability.
- A niche where speed beats scale. The market rewards domain depth and fast iteration more than reach, capital, or data stock.
- Access to scarce AI talent. The team has people who genuinely master the tools, rather than merely licensing them.
- Lightly regulated. It is not health, finance, or critical infrastructure, where compliance fixed costs favor the large.
Miss one of these conditions and the picture inverts. Without scarce talent, the tool becomes an expensive license[11]. Without a niche, you meet competitors who have the same AI plus more capital and more data[8]. In heavily regulated fields, size protects structurally. And even when every condition holds, an honest caveat remains: the long-run evidence to turn "can win" into "wins durably" is missing[6]. The thesis is not a law, then, but a special case with clear preconditions and an open time horizon.
What decision-makers should do about it
The most honest conclusion is not a shrug. The future does not belong categorically to small teams but to a heterogeneous mix: large platforms, mid-size specialists, and small niche players, each where its structural shape fits the task. For concrete decisions, several things follow.
First, treat "small" as a bet on a context, not a strategy in itself. Test your venture against the checklist above before you turn headcount into a virtue. If the context does not fit, a lean team is not a strength but an understaffed flank.
Second, deploy AI first where the evidence supports it, in content-driven work, and treat complex, coordinated decisions with care. Where humans and AI mesh poorly, naive automation costs quality on average[5]. Put the approvals where responsibility sits, not where it is convenient.
Third, invest in capability before tools. The mid-market bottleneck is talent, not software[11]. A team that truly masters its tools beats a larger one that merely owns them, and that is the real kernel of the thesis. It is not team size that wins, but focus.
Fourth, take your data, your regulation, and your capital base seriously before betting on "small." Where data is the moat, compliance the barrier, and capital the precondition, AI amplifies the advantages of the large rather than leveling them[1]. The macro evidence says the same: the gain hinges on execution, not on the tool[4]. And SME productivity turns on structural factors AI does not resolve on its own[10].
Fifth, measure honestly and expect spread. The 15% boost is an average with wide errors, not a promise[6]. Run your own small measurements before you make a number from a study the basis of a headcount decision. And accept where the evidence is silent: much of this debate is not settled yet[12].
The right lesson, then, is not "go small" but "get focused and honest about your context." That is the decision that makes a difference tomorrow, regardless of how many people sit at your table.
