Where AI Value Lands

Weber et al. — AI Startup Business Models (2021)

8 min read

Weber et al. — AI Startup Business Models: Key Characteristics and Directions for Entrepreneurship Research

Citation

Weber, M., Beutter, M., Weking, J., Böhm, M., & Krcmar, H. (2021). AI Startup Business Models: Key Characteristics and Directions for Entrepreneurship Research. Business & Information Systems Engineering, 64(1), 91–109. https://doi.org/10.1007/s12599-021-00732-w. Received 10 March 2021; accepted 18 October 2021; published online 13 December 2021.

Central claim

AI startups inherit most value-delivery and value-capture machinery from common IT-related business models (SaaS, subscription, transaction-based revenue), but they depart from those models in three specific aspects: (1) new value propositions that shift IT into the knowledge and service work domain, (2) a different role for data — used to train embedded models rather than to generate insights — and (3) a more technology-centered overall business logic where mastery of AI capability dominates other business-model dimensions.

The paper’s contribution is operational: a business-model taxonomy of AI startups (11 dimensions, 39 characteristics) distilled empirically from 100 Crunchbase startups, and a four-pattern clustering of those startups into archetypal business models.

Method (one paragraph)

100 AI startups randomly drawn from Crunchbase (October 2020), filtered by stable operating status, ≥$1M funding, founded post-2010, AI-as-core-product. Taxonomy developed via Nickerson et al. (2013) — 5 iterations alternating conceptual-to-empirical and empirical-to-conceptual. Hierarchical agglomerative Ward clustering on dichotomized dimensions yielded four clusters; cluster assignment manually validated at 84% agreement.

The taxonomy — 11 dimensions × 39 characteristics

Organized under the four Massa et al. (2017) / Teece (2010) business-model meta-categories:

CategoryDimensionCharacteristics
Value propositionCore AI valueCognitive insights / Monitoring & anomaly detection / Process & task support / Autonomous robots & bots
Continuous learningCentral learning & updates / Learning at customer side / Not provided
Value creationPrimary AI technologyMachine learning / Natural language processing / Computer vision / Robotics
Data typeNumeric / sensor data / Textual / document data / Natural language data / Visual data / Mixed data
Data sourceSelf-generated / Acquired / Publicly available / Customer provided on demand / Customer transmitted continuously
Hardware provisionYes / No
Value deliveryDelivery modeSoftware application / Programmable interface / Base technology / AI-produced output
Level of customizationStandardized product / service / Tailoring / individualization / Full customization
CustomerB2B / B2C / Both
Industry scopeIndustry focused / Industry agnostic
Value captureCustomer chargeFree of charge / Subscription-based / Transaction-based / One-time payment

See ai-startup-business-archetypes-weber for the operational classifier.

The four archetypal patterns

Briefly (each has its own page):

  1. ai-charged-product-service-provider (26/100) — products/services with readily trained AI models embedded. Standardized. B2B-dominant. Industry-focused. Top of H2_u-curve-of-value. Examples: Overjet, Alegion, Synapse.
  2. ai-development-facilitator (25/100) — tools that help other firms build their AI (APIs, SDKs, no-code chatbot builders). Industry-agnostic. NLP-dominant. Middle of the U. Examples: BotXO, Mindsay.
  3. ai-data-analytics-provider (30/100, largest) — integrate and analyze customer data for decision support; classical analytics with ML inside. Subscription. Shoulder of the U. Examples: Kubit, Falkonry, Zebrium.
  4. ai-deep-tech-researcher (19/100) — frontier R&D as the product, in robotics / autonomous driving / drug discovery / neurotech. Often pre-revenue, externally funded. Bottom of the U when defensible. Examples: Cerenion, Syrius Robotics.

Key data points

  • 27,900 Crunchbase startups tagged “Artificial Intelligence” as of September 2021. Compare ~97,000 in 2026 → see ai-startup-count-crunchbase-2021-2026.
  • 100-startup case base, sampled from 8,076 post-2010 AI-tagged Crunchbase startups passing the operating-status + funding + AI-core filters.
  • 3.8 data sources per startup on average (websites, industry portals, whitepapers, investment interviews).
  • 84% cluster-assignment agreement between manual and Ward-method clustering.
  • Pattern shares: 26 / 25 / 30 / 19 — Data Analytics Provider is the largest cluster, Deep Tech Researcher the smallest.

Three distinctive aspects (the §5 argument)

The paper’s “what is genuinely new about AI startup BMs” answer has three legs:

  1. New value propositions through AI capabilities. AI startups shift IT into knowledge and service work — call centers, fraud detection, disease diagnosis, customer support. Previous data-driven business models targeted backend operations; AI startups target human-labor-dominant domains. This connects to displacement and ai-task-exposure-decomposition in the project.
  2. Different roles of data. Not all AI startups are equally data-dependent. Specifically: in the ai-charged-product-service-provider pattern, data is used to train models that are then embedded — not to generate live insights. The other patterns extend the wiener-traditional-it-archetypes triad.
  3. Technology-centered business logic. Mastering the underlying AI dominates other dimensions of the business model. The competitive question becomes: can another startup catch up with a bigger dataset and better algorithms, or compensate with better usability / branding? The paper raises this as an open question; the project’s answer (so far) is in H2_u-curve-of-value and distribution-moat.

Hypothesis touches

  • H1_L0-L7-ladder — Weber’s pattern is the seller-side parallel: a startup at a given ladder position can sell at a different ladder position. Modestly reinforces H1 by giving its seller-side mirror.
  • H2_u-curve-of-value — the four Weber patterns map onto the U:
    • AI-charged P/S Provider → top of U (validated in 2026 by Cursor, Windsurf, Perplexity — see where-value-lands-2026).
    • AI Development Facilitator → the squeezed middle.
    • Data Analytics Provider → shoulder. Existence of 30 viable startups here is mild evidence against the strong form of “middle dies.”
    • Deep Tech Researcher → bottom of U when defensible.
    • Net effect: reinforces H2 qualitatively, complicates it empirically — the middle is not entirely dead at the firm-count level (though it might be at the value-capture level).
  • H3_orthogonal-axes-under-priced — direct strong reinforcement. The 11 Weber dimensions are the kind of cross-cutting axes H3 names. Two startups at the same L1–L2 ladder position can sit in opposite Weber patterns.
  • H4_rl-specialization-value-pocket — modest reinforcement. Deep Tech Researcher pattern is the structural slot H4 inhabits in 2026, with the substitution that the startup no longer trains its own base model.

Contradictions surfaced

  • The 30-startup Data Analytics Provider cluster (the largest) is a mild counter-evidence to the strong “middle layer is dead” version of H2_u-curve-of-value / middle-layer-defensibility. The middle layer is densely populated and reasonably profitable — but the value-capture share at exit may still be small relative to top-of-U archetypes. The hypothesis needs to distinguish firm count from value capture.
  • Weber et al. expected to find ethics as a salient dimension; they did not. The taxonomy contains no ethics dimension. This is worth flagging for any future-paper section on AI-startup ethics.

Open questions surfaced (verbatim from Table 4)

  • When is data not essential to value creation for AI startups?
  • How can AI startups create competitive advantage (e.g., via AI model leadership)?
  • What type of AI technology is easier to replicate than others?
  • How can AI startups challenge competitors that have an AI training advantage?
  • What are the implications of continuous learning and data network effects for entrepreneurship?
  • How and when do AI startups challenge existing service and knowledge work dominant industries?

These belong on 05_open-questions — at least the first three intersect with active project questions on value-capture and distribution-moat.

2021 → 2026 vintage caveats

The paper is pre-ChatGPT (received March 2021, sample taken October 2020). Three structural shifts since:

  1. Foundation-model commoditization has eaten most thin-wrapper ai-development-facilitator startups; the pattern survives in narrower niches (agent frameworks, fine-tuning APIs — see rl-apis).
  2. AI-startup population growth. Crunchbase tag has grown ~3.5×; see ai-startup-count-crunchbase-2021-2026.
  3. A fifth pattern is plausibly emerging: agentic-workflow providers — multi-step foundation-model orchestrators for business workflows. The 2021 taxonomy folds them into AI-charged P/S Provider, but the value-creation logic (action orchestration over context-control + autonomy-slider) is different enough that future-paper sections may want to treat it separately. The vertical-domain instance is vertical-ai-orchestration from deloitte-ai-dossier-eri.

The full inventory of dimensions this 2021 taxonomy is missing for a 2026 sample — model sourcing, autonomy level, orchestration depth, distribution architecture, data flywheel, ethics-as-product — is consolidated in weber-taxonomy-2026-gaps, together with the source-hunting plan for closing each gap.

Independent adoption — Gültekin et al. (2025)

The first independent peer-reviewed empirical study in the wiki to adopt this taxonomy as classifier is gultekin-pinarbasi-commercialisation-ai-2025 — interviews with 10 Turkish AI startups (Istanbul / Antalya technoparks, Nov 2022). They use Weber’s four-pattern classification without modification. Distribution in their sample: 1 ai-development-facilitator, the rest ai-charged-product-service-provider or ai-data-analytics-provider; zero ai-deep-tech-researcher (consistent with the country’s startup mix). The study independently re-confirms Weber et al.’s own surprised non-finding on ethics — the dimension is present as concern, absent as product surface — across a different geography. See weber-taxonomy-2026-gaps for the dimension-by-dimension reading of what this independent sample evidences and what it leaves untouched.

See also

Graph View

Backlinks