There Is an App for That — Until There Isn't
Abstract
Digital health applications and psychotherapy follow different capacity, evidence, and reimbursement paths. Baumol’s cost disease explains one pressure on labour-intensive care; it does not explain a particular German fee decision or establish that an app substitutes for a therapist.
Contents
I have built small tools that compress a repetitive task into a few minutes of machine-assisted execution. The result is satisfying. A piece of software can often serve another user at low marginal cost, even though compute, support, security, and maintenance never become literally free.
Psychotherapy has a different production structure. Much of it requires trained attention delivered over time. German digital health applications (Digitale Gesundheitsanwendungen, or DiGA) can be prescribed and reimbursed through their own regulatory path, while access to psychotherapy depends on workforce, regional planning, indication, and available appointments. The two routes are easy to place in one rhetorical comparison and difficult to compare like for like.
William Baumol’s cost-disease model clarifies one pressure in that comparison. It does not decide the policy.
The Seduction of Leverage
The phrase vibe coding describes one version of software development with a language model: describe desired behaviour, let the model draft code, run it, and iterate in natural language. It can lower the barrier to a prototype. It does not remove requirements, domain knowledge, testing, security, or the need to understand a system whose failure matters.
The attraction is leverage. Some software outputs can scale much faster than the labour that produced their first version. That changes which projects feel economically and culturally exciting. It does not follow that every activity should be reorganised to exhibit the same curve.
What Therapy Requires—and What Can Change
A psychotherapy session commonly involves sustained attention between a clinician and a patient. A meta-analysis by Flückiger and colleagues found a robust association between therapeutic alliance and adult psychotherapy outcomes, a correlation of about $r = 0.28$ across 295 studies [[1]]. Association does not make alliance the sole mechanism of every therapy, establish that techniques are interchangeable, or prove that all care must be one-to-one.
Individual, group, digital, and blended therapies are different interventions with different evidence, risks, and capacity profiles. Technology can change administration, monitoring, exercises, communication, and sometimes delivery. It does not make clinician time and an app the same product. How much a service design changes capacity and outcomes is an empirical question, not a fixed biological ceiling.
What Baumol’s Model Actually Says
Baumol and Bowen’s 1966 example was the performing arts [[2]]. A string quartet still needs four performers for the duration of that quartet. If productivity and wages grow faster elsewhere, a labour-intensive performance experiences relative cost pressure even when its own quality has not declined.
Baumol’s 1967 paper turned the example into a two-sector model [[3]]. Output per hour of labour, $a_i(t) = a_i(0)\,e^{g_i t}$, grows at rate $g_A$ in a progressive sector A and at a smaller rate $g_B$ in a stagnant sector B; Baumol, who numbers the stagnant sector 1, sets $g_B = 0$. Ignore costs other than labour, as he does, and let both sectors pay the same hourly wage $w(t)$. One unit of output then costs $c_i = w/a_i$, and
$$ \frac{c_B(t)}{c_A(t)} = \frac{a_A(t)}{a_B(t)} = \frac{a_A(0)}{a_B(0)}\,e^{(g_A - g_B)\,t}. \tag{1} $$The wage cancels. Whatever happens to pay, as long as it happens in both sectors, the stagnant output becomes dearer relative to the progressive one at the rate of the productivity gap; at a gap of 2 per cent a year the ratio doubles in $\ln 2 / 0.02 \approx 35$ years (Fig. 1). In hours of work, a unit of B costs $1/a_B(t)$ hours, constant when $g_B = 0$. The quartet does not become dearer in working time. Everything else becomes cheaper.
Eq. (1) needs three assumptions: labour is the only cost, wages move together across sectors, and the gap persists. If society nonetheless fixes the ratio of the two outputs, “more and more of the total labor force must be transferred to the nonprogressive sector”, and overall growth slows [[3]].
I previously wrote that the model does not say costs rise without bound. Within the model they do, relative to A: Baumol’s first proposition is that the stagnant unit cost “will rise without limit” [[3]]. The model does not say that a real service follows Eq. (1) indefinitely (Baumol called the split “a matter of degree rather than an absolute dichotomy”), that every component of healthcare is technologically fixed, or that one software advance causes one therapy fee decision. Baumol extended it to education and hospitals in 1967 [[3]] and, at book length, to healthcare in 2012 [[4]].
Psychotherapy contains labour-intensive components, so the analogy is useful. It is not exact. A one-to-one session of fixed length is close to Baumol’s case $g_B = 0$; a one-off reorganisation, such as group treatment, changes the level $a_B$ but not the rate $g_B$. Treatment modalities, administration, training, regulation, regional supply, and clinical need all affect capacity. “Cost disease” is not a diagnosis of laziness and not a complete model of a health system.
There Is an App for That
Germany’s Digital Healthcare Act (Digitale-Versorgung-Gesetz), in force since 19 December 2019, created the fast-track procedure through which eligible DiGA enter a federal directory and can be prescribed [[5]]. The Federal Institute for Drugs and Medical Devices (Bundesinstitut für Arzneimittel und Medizinprodukte, BfArM) decides within three months of a complete application. A product whose comparative study shows a “positive healthcare effect” (positiver Versorgungseffekt) is listed permanently; one still lacking that study can be listed provisionally for a trial of up to twelve months, extendable once, in individual cases, to at most 24, and is reimbursable from listing either way [[5]]. The live BfArM directory is the authoritative place to check a product’s current indication, listing status, and evidence documentation [[6]].
The latest dated count I could open is the annual report of the GKV-Spitzenverband, the statutory insurers’ national association: by 31 December 2025, 74 DiGA had been admitted, 60 first listed on trial, and 16 removed again, leaving 58 (48 permanent, 10 on trial), of which 32 addressed mental disorders; completed trials averaged 20 months against the twelve intended [[7]]. Data to July 2024 showed most trials overrunning [[8]]. Delays and removals are not evidence that every DiGA is ineffective. Nor does availability establish that a digital product is clinically substitutable for psychotherapy. Suitability depends on the indication, evidence, patient, risks, and intended role in care.
The tempting arithmetic is to divide an app budget by a session fee and report how many therapy sessions the money “could have bought.” I have removed that calculation. A budget does not create trained clinicians, regional seats, appropriate treatment, or appointment time. Comparing financing priorities is legitimate; translating euros mechanically into clinical capacity is not.
Policy Is Not an Equation Output
German fee schedules, needs planning, professional training, crisis services, and DiGA reimbursement have separate institutions and legal mechanisms. A statement by a professional body, an insurer report, and a live product directory may each be accurate within its scope while answering different questions.
Baumol’s model cannot identify the cause of a particular reimbursement or workforce decision. It also cannot show that policy intends to replace therapists with apps. Establishing either claim would require a dated policy record and causal evaluation, not two headline numbers moving in opposite directions.
The narrower inference survives. Care capacity depends partly on trained people whose time remains costly, while digital products can follow a more scalable delivery path. Public policy should evaluate both access and evidence without pretending they are interchangeable units.
The Uncomfortable Part
I enjoy leverage. I also think labour-intensive care should be funded deliberately rather than treated as inefficient merely because its measured productivity does not grow like software. That is my political judgment, not a deduction from Baumol.
The comparison with a string quartet is useful but limited. A performance, a therapy session, and a crisis call all contain time that cannot simply be copied. Their outcomes, labour markets, technologies, and funding rules are different. Some parts can change; some encounters still need someone’s attention.
Evidence-based digital tools can be useful. They should not become an accounting fiction that makes trained human capacity look obsolete. Policy decides whether that capacity is available.
Literature and calculations checked through 2026-10-24.
References
[1] Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340. https://doi.org/10.1037/pst0000172
[2] Baumol, W. J., & Bowen, W. G. (1966). Performing Arts, The Economic Dilemma: A Study of Problems Common to Theater, Opera, Music and Dance. Twentieth Century Fund.
[3] Baumol, W. J. (1967). Macroeconomics of unbalanced growth: The anatomy of urban crisis. The American Economic Review, 57(3), 415–426. https://www.jstor.org/stable/1812111
[4] Baumol, W. J. (2012). The Cost Disease: Why Computers Get Cheaper and Health Care Doesn’t. Yale University Press.
[5] Bundesinstitut für Arzneimittel und Medizinprodukte (2025). Das Fast-Track-Verfahren für digitale Gesundheitsanwendungen (DiGA) nach § 139e SGB V: Ein Leitfaden für Hersteller, Leistungserbringer und Anwender [The fast-track procedure for digital health applications under Section 139e of Social Code Book V: A guide for manufacturers, providers and users], version 3.6, 10 December 2025. https://www.bfarm.de/SharedDocs/Downloads/DE/Medizinprodukte/diga_leitfaden.pdf?__blob=publicationFile
[6] Bundesinstitut für Arzneimittel und Medizinprodukte. DiGA-Verzeichnis [DiGA directory]. https://diga.bfarm.de/de
[7] GKV-Spitzenverband (2026). DiGA-Bericht des GKV-Spitzenverbandes 2025: Bericht über die Inanspruchnahme und Entwicklung der Versorgung mit Digitalen Gesundheitsanwendungen [DiGA report 2025: Use and development of care with digital health applications], reporting period 1 September 2020 to 31 December 2025, published 1 April 2026. https://www.gkv-spitzenverband.de/media/dokumente/krankenversicherung_1/telematik/digitales/2025_DiGA_Bericht_GKV_SV.pdf
[8] Goeldner, M., & Gehder, S. (2024). Digital Health Applications (DiGAs) on a Fast Track: Insights From a Data-Driven Analysis of Prescribable Digital Therapeutics in Germany From 2020 to Mid-2024. Journal of Medical Internet Research, 26, e59013. https://doi.org/10.2196/59013
Changelog
- 2026-07-11: Removed unsupported current fee, budget, product-count, and budget-to-therapy arithmetic; distinguished alliance evidence from a universal therapy mechanism; and bounded Baumol’s relative-cost model away from causal claims about German policy, workforce collapse, or AI.
- 2026-10-03: Corrected the venue of reference 5 (Goeldner & Gehder, 2024) from JMIR mHealth and uHealth, volume 12, to Journal of Medical Internet Research, volume 26, e59013, as given by the DOI record (Crossref).
- 2026-10-24: Stated Baumol’s two-sector model, which the article invoked without stating, as Eq. (1) with its assumptions (labour-only costs, a common wage, a persistent productivity gap), sourced to Baumol (1967), now reference 3; added Fig. 1, that with flat productivity the stagnant output costs the same in working time, and that a one-off reorganisation changes the level, not the rate. Corrected “It does not say that costs literally rise without bound”: in the model the stagnant unit cost, relative to the progressive one, “will rise without limit” (Proposition 1). Noted that the 1967 paper already covers education and hospitals. Spelled out BfArM and dated the DiGA facts: the Act in force since 19 December 2019, the fast-track rules and reimbursement from listing (BfArM guide, version 3.6), and the GKV-Spitzenverband’s count for 31 December 2025 (58 listed, 32 for mental disorders, 16 removed; completed trials averaging 20 months). Goeldner & Gehder’s delay finding replaces “listings can later change”. Added the alliance–outcome correlation (r ≈ 0.28). References are renumbered in order of citation; reference 5 of the 2026-10-03 entry is now reference 8.