AI Startup Business Archetypes (Weber et al. 2021)
A four-pattern taxonomy of AI-startup business models, derived empirically from a 100-startup Crunchbase sample by weber-ai-startup-business-models. Sits parallel to ecosystem-business-archetypes (Choudary’s horizontal taxonomy) and taker-shaper-maker (WEF buyer-side typology) — together these form the three off-the-shelf sub-vocabularies the project uses for H3_orthogonal-axes-under-priced.
The four archetypes (one-line each)
| Pattern | Definition | Sample share | Canonical example |
|---|---|---|---|
| ai-charged-product-service-provider | Products/services with readily trained AI models embedded — value delivered by the trained model itself, not by analyzing the customer’s data | 26 / 100 | Overjet (dental X-ray analysis); Alegion (data-labeling) |
| ai-development-facilitator | Tools that help other firms build their own AI — SDKs, APIs, no-code workbenches, customizable chatbot platforms | 25 / 100 | BotXO (custom conversational AI); Hugging Face (NLP APIs) |
| ai-data-analytics-provider | Integrates and analyzes vast data (internal + external) for decision support; classical analytics with ML inside | 30 / 100 (largest) | Kubit (retention/profitability anomaly detection); Falkonry (sensor → machine-state) |
| ai-deep-tech-researcher | Research-led; develops frontier AI tech for niche problems (robotics, autonomous driving, drug discovery, neurotech); not yet standardized; often pre-revenue | 19 / 100 | Cerenion (brain-activity interpretation); Syrius Robotics (warehouse robots) |
The full morphological box — 4 categories · 11 dimensions · 39 characteristics
Every AI startup is a point in this space. The four archetypes are clusters found by Ward-method hierarchical clustering over the dichotomized dimensions (Weber et al. §4). All characteristic names below are verbatim from the paper.
Category 1 — Value proposition: what the startup promises the customer
| Dimension | What it captures | Characteristics |
|---|---|---|
| Core AI value | The fundamental kind of value the AI delivers | Cognitive insights · Monitoring & anomaly detection · Process & task support · Autonomous robots & bots |
| Continuous learning | Whether the model keeps improving after deployment | Central learning & updates · Learning at customer side · Not provided |
Category 2 — Value creation: what the startup mobilises to make the promise real
| Dimension | What it captures | Characteristics |
|---|---|---|
| Primary AI technology | The dominant AI discipline behind the product | Machine learning · Natural language processing · Computer vision · Robotics |
| Data type | The modality of data the system ingests | Numeric / sensor data · Textual / document data · Natural language data · Visual data · Mixed data |
| Data source | Where the training and inference data come from | Self-generated · Acquired · Publicly available · Customer provided on demand · Customer transmitted continuously |
| Hardware provision | Does the startup ship hardware alongside the software? | Yes · No |
Category 3 — Value delivery: the form factor in which the value reaches the customer
| Dimension | What it captures | Characteristics |
|---|---|---|
| Delivery mode | How the customer touches the product | Software application · Programmable interface · Base technology · AI-produced output |
| Level of customization | How much the product is shaped to each customer | Standardized product / service · Tailoring / individualization · Full customization |
| Customer | Who pays | B2B · B2C · Both |
| Industry scope | How wide a target market | Industry focused · Industry agnostic |
Category 4 — Value capture: how the startup turns value into revenue
| Dimension | What it captures | Characteristics |
|---|---|---|
| Customer charge | The pricing logic | Free of charge · Subscription-based · Transaction-based · One-time payment |
The interactive version of this table (with archetype presets and a radial dendrogram) is at outputs/weber-ai-startup-taxonomy.html.
Unit of analysis — one BM, not one firm
Prerequisite (added 2026-05-25)
The classifier applies to a single business model, not to the whole startup. A firm running multiple BMs — a business-model-portfolio — is classified by applying the four-pattern taxonomy separately to each BM. Forcing a multi-BM startup into one pattern destroys signal. Source for the BMP frame: westerveld-business-model-portfolio-2023. Operational consequence in this wiki: tag source pages with
weber-pattern:per BM when the firm runs more than one.
How to classify a startup (the salient-characteristic shortcut)
Weber et al. built an 11-dimension / 39-characteristic taxonomy (value proposition, value creation, value delivery, value capture), then clustered. For practical classification, the salient characteristics that separate one pattern from the others are usually enough:
Is the AI model already trained when delivered?
├─ Yes, standardized product → AI-charged Product/Service Provider
├─ No, customer trains via tools you provide → AI Development Facilitator
└─ AI runs over customer's own data, no embedded model → Data Analytics Provider
Is the technology itself the deliverable (not productized yet)?
└─ Yes → Deep Tech Researcher
The full discriminating-dimension table (data type, data source, hardware provision, customization level, customer, industry scope, revenue model) is in each pattern page’s “Salient characteristics” section.
Mapping to project hypotheses
-
H2_u-curve-of-value — where each pattern sits on the U:
- AI-charged Product/Service Provider → top of U (workflow-embedded, distribution-dependent). Cursor, Windsurf, Perplexity are 2026 instances of this pattern.
- AI Development Facilitator → the squeezed middle. Generic SDKs and chatbot-builders are exactly what middle-layer-defensibility argues fails — except when distribution survives (Hugging Face).
- Data Analytics Provider → shoulder of U (mature, profitable, but not where the new AI margin sits — closer to traditional ai-productivity-firm-level BI).
- Deep Tech Researcher → bottom of U when the niche is genuinely defensible (frontier weights, robotics, biology). Closely related to H4_rl-specialization-value-pocket.
-
H1_L0-L7-ladder — seller-side parallel: the Weber taxonomy is a seller-side sibling of the ladder, which is mostly buyer/substrate-side. An AI-charged Product/Service Provider typically sells at L1–L2 of the buyer’s stack; a Deep Tech Researcher sells the L4–L6 substrate.
-
H3_orthogonal-axes-under-priced — direct evidence. The 11 dimensions Weber et al. extracted are exactly the kind of cross-cutting axes H3 names. Two startups at the same ladder position can sit in opposite Weber patterns.
What Weber et al. argue is genuinely new about AI startup BMs
From the paper’s discussion (§5):
- New value propositions — shifting IT into the domain of knowledge and service work (replacing humans rather than streamlining backends).
- Different roles of data — in the AI-charged P/S Provider pattern especially, data is used to train an embedded model, not to generate insights. This is the qualitative break from wiener-traditional-it-archetypes (data users / data suppliers / data facilitators).
- Technology-centered business logic — not all AI startups are equally data-dependent. The pattern decides.
2021 → 2026 vintage shift (what has happened since)
The paper is December 2021 — pre-ChatGPT. Three things have shifted:
- Foundation-model commoditization. Most of what was “AI Development Facilitator” in 2021 (custom NLP stacks, build-your-own-chatbot) has been eaten by the GPT-4 / Claude API surface. The pattern now mostly survives only where distribution + infra sticks (Hugging Face) — see middle-layer-defensibility.
- Crunchbase AI-startup count. Sept 2021: 27,900 → 2026: ~97,000 (≈3.5×). See ai-startup-count-crunchbase-2021-2026.
- A fifth pattern is plausibly emerging in 2026: agentic-workflow providers — startups that orchestrate multiple foundation-model calls into multi-step business workflows (Cursor’s task agents, LangGraph apps, the vertical-ai-orchestration pattern from ER&I). The 2021 taxonomy folds these into “AI-charged Product/Service Provider”, but the value-creation logic (action orchestration over context-control) is different enough that it may deserve its own pattern.
A consolidated diagnosis of the dimensions Weber’s 11-dimension box is missing for 2026 — model sourcing, autonomy level, orchestration depth, distribution architecture, data flywheel, ethics-as-product — plus the agentic-workflow fifth archetype, lives in weber-taxonomy-2026-gaps. That page is the working-draft input to the eventual GTM sibling taxonomy.
Why this matters for the project
The user’s stated goal at ingest was to enable future-startup classification using this taxonomy as new startups appear in the wiki. The four-pattern table above + the salient-characteristic decision tree are the operational classifier. Use it whenever a new AI-startup source is ingested, and tag the resulting source/concept page with the Weber pattern in its frontmatter (weber-pattern: ai-charged-product-service-provider etc.).
See also
- weber-ai-startup-business-models — the source.
- business-model-portfolio — the unit-of-analysis prerequisite: each BM is classified separately when a firm runs multiple BMs.
- westerveld-business-model-portfolio-2023 — the source for the BMP frame.
- weber-taxonomy-2026-gaps — diagnostic page on what this taxonomy fails to encode about 2026 startups; six missing dimensions + the fifth-archetype sketch; working-draft input to the GTM sibling taxonomy.
- gultekin-pinarbasi-commercialisation-ai-2025 — first independent empirical study in the wiki to use this classifier (Turkish AI startups, n=10, Nov 2022); 9/10 firms fall into Pattern 1 or 3.
- ecosystem-business-archetypes — Choudary’s parallel four-archetype taxonomy (horizontal-ecosystem lens, not AI-tech lens). The two read together cover Axis 3 of 07_analytical-vocabulary.
- taker-shaper-maker — WEF’s three-archetype buyer-side typology.
- wiener-traditional-it-archetypes — the pre-AI IT-business-model triad (data users / suppliers / facilitators) the Weber paper positions itself against.
- H1_L0-L7-ladder, H2_u-curve-of-value, H3_orthogonal-axes-under-priced, H4_rl-specialization-value-pocket.