Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

From Playbook to Practice: Exploring Lean Methodology Alignment in Swedish AI Startups

Mc Crea, William LU and Thuresson, Oliver Oskar Nelsen (2026) INTM01 20261
Innovation Engineering
Abstract (Swedish)
The Lean Startup Methodology (LSM) has become a near-canonical heuristic for
early-stage venture building, yet its empirical foundation remains thin and its
applicability to artificial intelligence (AI) ventures is under-theorised. AI-native
startups operate under conditions that strain the methodology’s core assumptions:
outputs are probabilistic rather than deterministic, performance is contingent on
data availability and quality, and the distance between a working prototype and a
production-grade system is itself a first-order source of uncertainty. This thesis
asks how Swedish AI startups adopt, adapt, or reject LSM constructs in practice,
and which dimensions of the methodology survive, mutate, or break under
AI-specific... (More)
The Lean Startup Methodology (LSM) has become a near-canonical heuristic for
early-stage venture building, yet its empirical foundation remains thin and its
applicability to artificial intelligence (AI) ventures is under-theorised. AI-native
startups operate under conditions that strain the methodology’s core assumptions:
outputs are probabilistic rather than deterministic, performance is contingent on
data availability and quality, and the distance between a working prototype and a
production-grade system is itself a first-order source of uncertainty. This thesis
asks how Swedish AI startups adopt, adapt, or reject LSM constructs in practice,
and which dimensions of the methodology survive, mutate, or break under
AI-specific conditions.
The study employs a qualitative, exploratory multiple-case design across eight
Swedish AI startups spanning pre-seed to Series B, drawing on semi-structured
interviews with founders and operators and a structured cross-case analysis. The
findings indicate partial and asymmetric adherence. The epistemic functions of
LSM such as hypothesis-driven strategy, validated learning, and pivot logic,
persist in adapted forms optimized for learning while canonical mechanisms are
reworked. Several operational artefacts, notably the prototype minimum viable
product (MVP), the linear Build-Measure-Learn (BML) loop, and innovation
accounting, require substantive adaptation or are bypassed altogether. The thesis
reframes LSM as a set of situated epistemic functions rather than a prescriptive
playbook and identifies the prototype-to-production gap as a defining feature of AI
venture building. (Less)
Popular Abstract (Swedish)
Sedan över ett decennium har Lean Startup-metoden präglat hur
entreprenörer bygger nya bolag: börja smått, släpp produkten till
kunderna, lär, och byt riktning om gissningen var fel. Vår studie av åtta
svenska AI-bolag, från kodverktyg till juridiska AI-agenter, visar att
metoden fortfarande innehåller goda grundprinciper att förhålla sig till,
men att ett centralt antagande inte stämmer. Avståndet mellan en AI-demo
som fungerar och en produkt som kunderna verkligen vågar lita på är
betydligt större än metoden förutsätter.
Lean Startup-metoden, som formulerades av Eric Ries 2011, lärs ut på
handelshögskolor och i acceleratorer världen över och har präglat tusentals
startupbolags arbetssätt. Men metoden tillkom innan den nuvarande... (More)
Sedan över ett decennium har Lean Startup-metoden präglat hur
entreprenörer bygger nya bolag: börja smått, släpp produkten till
kunderna, lär, och byt riktning om gissningen var fel. Vår studie av åtta
svenska AI-bolag, från kodverktyg till juridiska AI-agenter, visar att
metoden fortfarande innehåller goda grundprinciper att förhålla sig till,
men att ett centralt antagande inte stämmer. Avståndet mellan en AI-demo
som fungerar och en produkt som kunderna verkligen vågar lita på är
betydligt större än metoden förutsätter.
Lean Startup-metoden, som formulerades av Eric Ries 2011, lärs ut på
handelshögskolor och i acceleratorer världen över och har präglat tusentals
startupbolags arbetssätt. Men metoden tillkom innan den nuvarande AI-vågen, och
en ny generation grundare bygger nu sina produkter under förutsättningar som det
ursprungliga ramverket inte tog höjd för.
För att förstå vad som händer i praktiken intervjuade vi grundare och
nyckelpersoner på åtta svenska AI-bolag, från pre-seed till Series B. Urvalet
sträcker sig över kodverktyg, dokumentautomation för försäkringsbranschen,
juridiska AI-agenter, klinisk dokumentation och tillämpad AI-konsultverksamhet.
Huvudfyndet är att Lean Startup håller delvis. De principer som verkligen betyder
något, att lära sig av kunderna, att iterera snabbt och att våga byta riktning, är
fortfarande användbara men följs inte rakt av. Men flera av metodens mer
receptartade redskap håller inte måttet i ett AI-sammanhang.
Den mest påtagliga friktionen rör det som metoden kallar Minimum Viable
Product, den minsta version av en produkt som man sätter i händerna på kunderna.
I traditionell mjukvara ligger en MVP ofta nära den färdiga produkten. I AI ligger
en fungerande demo nästan alltid långt ifrån ett driftsatt system, eftersom
komplexiteten är betydligt större: svaren kan variera från gång till gång,
prestandan beror på data som teamet inte styr över, och kunderna börjar inte lita på
systemet förrän det är inarbetat i deras egna arbetsflöden. Vi kallar detta gapet
mellan prototyp och produktion, och menar att det är ett definierande drag i
AI-baserat entreprenörskap.
Två fynd överraskade oss. Formell mätning av framsteg saknas i stort sett helt:
grundarna förlitar sig på kundsamtal och eget omdöme snarare än på dashboards.
Och det klassiska beslutet om att byta riktning, en pivot, utlöses ofta inte av att en
kundhypotes föll, utan av att de underliggande AI-modellerna plötsligt klarar
något nytt, eller av grundarnas egen intuition och magkänsla.
Slutsatsen är att Lean Startup bäst läses som en uppsättning principer snarare än ett
steg-för-steg-recept. Principerna gäller fortfarande. De specifika verktygen
behöver däremot anpassas till AI-eran. (Less)
Popular Abstract
For more than a decade, the Lean Startup method has shaped how
entrepreneurs build new companies: start small, get the product into
customers’ hands, learn, and change course if the guess was wrong. Our
study of eight Swedish AI startups, from coding tools to legal AI agents,
shows that the method still offers sound guiding principles, but that one of
its central assumptions no longer holds. The distance between an AI demo
that works and a product customers genuinely trust is far greater than the
method assumes.
The Lean Startup method, set out by Eric Ries in 2011, is taught in business
schools and accelerators around the world and has shaped the way thousands of
startups operate. But it predates the current AI wave, and a new... (More)
For more than a decade, the Lean Startup method has shaped how
entrepreneurs build new companies: start small, get the product into
customers’ hands, learn, and change course if the guess was wrong. Our
study of eight Swedish AI startups, from coding tools to legal AI agents,
shows that the method still offers sound guiding principles, but that one of
its central assumptions no longer holds. The distance between an AI demo
that works and a product customers genuinely trust is far greater than the
method assumes.
The Lean Startup method, set out by Eric Ries in 2011, is taught in business
schools and accelerators around the world and has shaped the way thousands of
startups operate. But it predates the current AI wave, and a new generation of
founders is now building products under conditions the original framework never
anticipated.
To understand what actually happens in practice, we interviewed founders and key
team members at eight Swedish AI companies, ranging from pre-seed to Series B.
The sample spans coding tools, document automation for the insurance industry,
legal AI agents, clinical documentation, and applied AI consulting.
The main finding is that Lean Startup holds up in part. The principles that matter
most, learning from customers, iterating quickly, and daring to change direction,
remain useful and at times necessary. On the other hand, several of the method’s
operational tools cannot be followed step by step in an AI setting and need to be
adapted.
The clearest friction concerns what the method calls the Minimum Viable Product,
the smallest version of a product you put in customers’ hands. In traditional
software, an MVP is often close to the finished product. In AI, a working demo is
almost always a long way from a deployed system, because the complexity is far
greater. Answers can vary, performance depends on data the team does not control,
and customers only begin to trust the system once it is embedded in their own
workflows. We call this the prototype-to-production gap, and we argue that it is a
defining feature in the sample.
Two findings surprised us. Formal measurement of progress is almost entirely
absent: founders rely on customer conversations and their own judgement rather
than dashboards. And the classic decision to change direction, the pivot, is often
triggered not by a failed customer hypothesis, but by the underlying AI models
suddenly becoming capable of something new, or by the founders’ own intuition
and gut feel.
The conclusion is that Lean Startup is best read as a set of principles rather than a
step-by-step recipe. The principles still apply. The specific tools, however, need
adapting to the AI era. (Less)
Please use this url to cite or link to this publication:
author
Mc Crea, William LU and Thuresson, Oliver Oskar Nelsen
supervisor
organization
alternative title
När teorin möter verkligheten: hur lean-metodik tillämpas i svenska AI-startups
course
INTM01 20261
year
type
H2 - Master's Degree (Two Years)
subject
keywords
The Lean Startup Methodology (LSM) has become a near-canonical heuristic for early-stage venture building, yet its empirical foundation remains thin and its applicability to artificial intelligence (AI) ventures is under-theorised. AI-native startups operate under conditions that strain the methodology’s core assumptions: outputs are probabilistic rather than deterministic, performance is contingent on data availability and quality, and the distance between a working prototype and a production-grade system is itself a first-order source of uncertainty. This thesis asks how Swedish AI startups adopt, adapt, or reject LSM constructs in practice, and which dimensions of the methodology survive, mutate, or break under AI-specific conditions. The study employs a qualitative, exploratory multiple-case design across eight Swedish AI startups spanning pre-seed to Series B, drawing on semi-structured interviews with founders and operators and a structured cross-case analysis. The findings indicate partial and asymmetric adherence. The epistemic functions of LSM such as hypothesis-driven strategy, validated learning, and pivot logic, persist in adapted forms optimized for learning while canonical mechanisms are reworked. Several operational artefacts, notably the prototype minimum viable product (MVP), the linear Build-Measure-Learn (BML) loop, and innovation accounting, require substantive adaptation or are bypassed altogether. The thesis reframes LSM as a set of situated epistemic functions rather than a prescriptive playbook and identifies the prototype-to-production gap as a defining feature of AI venture building.
language
English
id
9252248
date added to LUP
2026-10-06 16:22:19
date last changed
2026-10-06 16:22:19
@misc{9252248,
  abstract     = {{The Lean Startup Methodology (LSM) has become a near-canonical heuristic for
early-stage venture building, yet its empirical foundation remains thin and its
applicability to artificial intelligence (AI) ventures is under-theorised. AI-native
startups operate under conditions that strain the methodology’s core assumptions:
outputs are probabilistic rather than deterministic, performance is contingent on
data availability and quality, and the distance between a working prototype and a
production-grade system is itself a first-order source of uncertainty. This thesis
asks how Swedish AI startups adopt, adapt, or reject LSM constructs in practice,
and which dimensions of the methodology survive, mutate, or break under
AI-specific conditions.
The study employs a qualitative, exploratory multiple-case design across eight
Swedish AI startups spanning pre-seed to Series B, drawing on semi-structured
interviews with founders and operators and a structured cross-case analysis. The
findings indicate partial and asymmetric adherence. The epistemic functions of
LSM such as hypothesis-driven strategy, validated learning, and pivot logic,
persist in adapted forms optimized for learning while canonical mechanisms are
reworked. Several operational artefacts, notably the prototype minimum viable
product (MVP), the linear Build-Measure-Learn (BML) loop, and innovation
accounting, require substantive adaptation or are bypassed altogether. The thesis
reframes LSM as a set of situated epistemic functions rather than a prescriptive
playbook and identifies the prototype-to-production gap as a defining feature of AI
venture building.}},
  author       = {{Mc Crea, William and Thuresson, Oliver Oskar Nelsen}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{From Playbook to Practice: Exploring Lean Methodology Alignment in Swedish AI Startups}},
  year         = {{2026}},
}