Small cohorts and the economics of serious education

Small cohorts are valuable when education depends on diagnosis, feedback, oral defense, and independent project supervision.
Cohort size should be chosen from the educational production function, not treated as a symbol of exclusivity.
A sustainable specialized school must price and organize scarce faculty attention without diluting completion standards.

Why Cohort Size Matters

Some educational content scales almost without limit. A recorded derivation can reach ten students or ten thousand at low marginal cost.

Other parts do not scale so easily:

• diagnosing a student's model;
• commenting on an open response;
• conducting an oral examination;
• reviewing a data audit;
• supervising a dissertation;
• resolving an ethical or empirical ambiguity; and
• designing a new transfer problem after correction.

Specialized AI education combines scalable content with scarce judgment. Cohort design must respect both.

A Capacity Model

Let faculty time available in a term be $T$. For student $i$, the expected time requirement is

Feasible enrollment satisfies

The difficulty is that $t_i$ is uncertain at admission. Students with larger foundation gaps, weaker writing, or ambitious research projects may require more feedback. A school that plans using only average delivery time will underestimate the tail.

A capacity reserve is therefore necessary:

where $\rho$ protects time for unexpected correction and supervision.

Congestion in Feedback

As enrollment rises, feedback can be delayed, shortened, standardized, or delegated.

Let the average feedback quality be

where $\bar n$ is the cohort size at which congestion becomes material.

The expression is illustrative, not an estimated law. It makes a design point: once scarce attention is saturated, additional students can reduce the educational input received by everyone.

Large programs can offset congestion with more faculty, teaching assistants, automated checks, and standardized course design. A small institution cannot assume those inputs expand instantly.

Small Does Not Automatically Mean Good

A small class can still be poorly designed. It may contain unclear objectives, weak instruction, no useful peer interaction, or arbitrary assessment.

Small size creates capacity for:

• individualized diagnosis;
• detailed written feedback;
• participation by every student;
• faculty knowledge of progress;
• flexible case discussion; and
• close project supervision.

The institution must convert that capacity into structured learning.

Table 1. How cohort size affects educational quality

Small-cohort opportunity	Required design
Faculty knows each student	Capability records rather than informal impressions
Every student can speak	Structured prompts and evidence-based discussion
Feedback can be detailed	Rubrics and revision requirements
Cases can be adapted	Stable outcomes and comparable standards
Projects can be ambitious	Scope controls and milestone review
Peer network can be strong	Norms of preparation and technical criticism

Exclusivity without educational design is only scarcity.

Peer Effects

Students learn from one another's questions, errors, domains, and models.

Let student $i$'s learning be

where $F_i$ is faculty input and $w_{ij}$ measures meaningful peer interaction.

A larger cohort offers more possible connections, but not all connections become educational. If discussion fragments into passive observation, effective $w_{ij}$ approaches zero.

Small cohorts can increase interaction density. They can also become intellectually narrow if every student has the same background. Cohort design should seek complementary diversity in disciplines and applications while preserving a common readiness threshold.

Heterogeneity and Instructional Range

The relevant issue is not demographic or disciplinary diversity alone. It is the range of preparation relative to one task.

Let readiness for concept $k$ have variance

If the range is very large, one lecture may be remedial for some and inaccessible for others. The solution need not be a homogeneous cohort. It can be:

• foundation modules;
• differentiated problem sets;
• optional derivation sessions;
• structured peer explanation;
• separate program tracks; and
• common integrative cases.

A small school can personalize routes more readily, but only if it has mapped the differences.

The Economics of High-Feedback Education

Suppose tuition revenue is $pn$, fixed cost $F$, and variable educational cost per student is $c(n)$. Program surplus is

If feedback congestion requires additional faculty or supervision, $c(n)$ may rise after a threshold. If the school suppresses those costs, measured surplus rises while quality falls.

The economically relevant objective is not short-term surplus alone:

where $Q(n)$ is educational quality and $R(n)$ the institution's future capacity and reputation.

The weights should not be rhetorical. A school that values quality must budget the activities that produce it.

Completion Rates and Cohort Economics

Completion is often treated as a simple measure of success. In a demanding program, it must be decomposed.

A non-completion can arise from:

• poor admissions placement;
• insufficient foundation support;
• unrealistic workload;
• life and financial constraints;
• ineffective teaching;
• failure to meet a transparent standard; or
• a rational decision that the program is not the right fit.

The school should estimate transition probabilities:

where states may include foundation, active study, pause, remediation, dissertation, completion, and withdrawal.

The objective is not to force every entrant to complete. It is to remove avoidable failure while preserving the meaning of completion.

Scalable and Non-Scalable Layers

A sustainable model separates what can scale from what should remain scarce.

Read as a diagnostic rather than a second scorecard, the framework becomes clear: Recorded lectures — Scaling potential: High; Quality control: Content review and updates. Automated prerequisite practice — Scaling potential: High; Quality control: Item validation and explanations. Standard technical checks — Scaling potential: Moderate to high; Quality control: Human review of edge cases. Case discussion — Scaling potential: Moderate; Quality control: Facilitator preparation and participation design. Open-response feedback — Scaling potential: Low to moderate; Quality control: Rubrics and reviewer calibration. Oral assessment — Scaling potential: Low; Quality control: Standardized prompts and multiple evidence. Dissertation supervision — Scaling potential: Low; Quality control: Scope limits and workload allocation. Independent examination — Scaling potential: Low; Quality control: Qualified readers and conflict controls.

Technology should scale repetition and routine checks. Faculty time should concentrate where judgment changes the student's model.

Small Cohorts and Online Education

Online delivery is sometimes assumed to mean mass delivery. It can instead separate access from cohort size.

Students can access lectures asynchronously across locations and time zones while entering limited feedback cohorts for cases, assessment, and dissertation work.

Let content access have near-zero marginal cost $c_c$, while cohort services have cost $c_h$:

Honest program design should not price or describe the entire education as if it were only content access.

The Dissertation Bottleneck

Dissertations create a particularly strong capacity constraint.

If each dissertation requires expected supervision $h_s$, examination $h_e$, and administration $h_a$, total dissertation capacity is

Projects with weak questions or excessive scope consume more time. Early milestones and scope review are therefore not bureaucratic additions; they protect both student progress and institutional capacity.

A program that scales coursework without planning dissertation capacity produces a queue at the point where independent judgment should be demonstrated.

The Optimal Cohort Is Not the Maximum Cohort

Suppose educational value per student is $q(n)$, peer value is $p(n)$, and net financial contribution is $\pi(n)$. An institution may choose cohort size by

subject to minimum standards:

The maximum feasible enrollment is the largest $n$ that avoids operational failure. The optimum may be smaller because marginal students create congestion before hard capacity is reached.

The optimum can also be larger than an arbitrarily exclusive cohort. More students may improve peer diversity, stabilize course offerings, and finance better faculty support. “Small” should not become an untested ideology.

The decision should be revisited as instructional technology, faculty capacity, and program mix change.

When to Add Students

Enrollment should increase when the institution can show that:

1. prerequisite preparation is functioning;
2. feedback latency remains acceptable;
3. assessors apply standards consistently;
4. case and supervision capacity exists;
5. qualified faculty are available; and
6. later student work preserves the intended outcome.

The threshold is not a fixed number. It depends on the delivery model, faculty structure, student preparation, and program level.

The SIAI GSB Cohort Principle

For SIAI GSB, small cohorts should be understood as a consequence of the educational model, not a prestige claim.

Recorded and self-paced coursework can widen access. Technical assessment, case criticism, and the final dissertation require limited and accountable faculty attention. Cohort size should follow that capacity.

The school should expand only when the scarce layers can expand without turning independent work into a procedural formality.

The marginal cost of an additional student is therefore not the cost of another login or lecture stream. It is the additional claim on diagnosis, feedback, oral examination, project supervision, and peer attention. These costs arrive unevenly: a cohort may appear scalable during content delivery and become congested near case presentations or dissertations. Capacity planning should follow the bottleneck rather than the average week. Recorded explanations and automated practice can scale the common foundation, but they do not eliminate the faculty time needed to examine an unusual model choice or help a student recover from a conceptual error. Small cohorts matter most where variation in student reasoning is itself the object of teaching. The relevant operating metric is not students per lecturer in the abstract, but unresolved analytical work per qualified reviewer at the busiest stage of the program. That measure reveals congestion before it appears in final completion statistics. It also distinguishes genuine instructional capacity from nominal staffing that cannot supervise the program’s most demanding work. Growth should wait until that review capacity expands without weakening the standard of individual scrutiny.

The economics of scale changes when assessment must reveal process. SIAI’s analysis of AI-detection failure recommends drafts, revision evidence, and explanation rather than automated suspicion. The Economy’s work on teacher AI literacy places more responsibility on faculty interpretation, and its analysis of the training gap emphasizes capabilities that require guided adaptation. Each response is educationally stronger than mass checking, but consumes scarce attention. Cohort design must therefore price feedback honestly rather than assuming that content delivery is the whole product.

Conclusion

The economics of serious education begins with a distinction between content and judgment.

Content can scale. Diagnosis, feedback, defense, and supervision remain comparatively scarce. Small cohorts are valuable when they protect those inputs and create dense peer learning.

The purpose is not to remain small. It is to avoid growth that changes the educational product while preserving the same name.

A specialized AI school should scale what technology can carry and preserve human attention where judgment matters.

References

Benjamin S. Bloom, “The 2 Sigma Problem”, Educational Researcher 13, no. 6 (1984): 4-16.
George D. Kuh, High-Impact Educational Practices, Association of American Colleges and Universities, 2008.
Standards and Guidelines for Quality Assurance in the European Higher Education Area, ESG 2015, 2015.
Swiss Institute of Artificial Intelligence (2026) ‘Why AI Detection Is Failing Higher Education’, SIAI AI Memo, 30 August.
The Economy Editorial Board (2026) ‘The Training Gap Behind the Rise of SuperHuman Labor’, The Economy Review, 28 July.
The Economy Editorial Board (2026) ‘Teacher AI Literacy Is the Real Test of AI in Education’, The Economy Review, 22 June.