AI startups captured $202.3 billion in venture funding in 2025, close to half of every venture dollar deployed worldwide and up from a 34% share the year before. The vocabulary that surrounds that capital is not decoration: how a company gets classified changes its valuation multiple, its risk profile, and the questions a partner asks in diligence.
This AI startup glossary defines the 65 terms that carry weight in an investment decision, grouped into five sections: company classifications, core technology, deal mechanics, startup metrics, and market intelligence.
On this page
- AI Company Classifications: The Five Types Every Investor Should Know
- Core AI Technology Terms: Deep Learning, Language Models, and RAG
- AI Investment Terms Every VC Should Know
- Startup Metrics That Matter for AI Companies
- AI Augmented vs AI Native: Why the Distinction Matters
- Market Intelligence Terms for AI Deal Sourcing
- Frequently Asked Questions
- How This Glossary Is Maintained
AI Company Classifications: The Five Types Every Investor Should Know
Bot Memo sorts every company in its repository into a five-tier classification. The tier a company lands in reflects how its value proposition depends on artificial intelligence, and that dependency drives valuation, risk, and exit path.
AI Native
Built with AI from the ground up. The product would not exist without AI capabilities.
Perplexity (answer engine), Harvey (legal and professional services), and Midjourney (image and video generation) are textbook examples.
These companies were founded AI-first, typically 2021 or later. Every core feature depends on machine learning models.
AI Augmented
Established products enhanced with AI features. AI improves the product but is not fundamental to its core value proposition.
Salesforce added Einstein AI to its CRM in 2016. Adobe built Sensei into Creative Cloud. The underlying product existed and had market fit before AI entered the picture.
AI Adjacent
Enables AI builders without running models itself. The picks-and-shovels play.
Scale AI provides data labeling. CoreWeave sells GPU cloud compute. Their customers build AI; these companies supply the infrastructure underneath.
AI Platforms
Trains foundation models from scratch or designs AI-specific silicon. Two questions settle it:
- Does the company train its own foundation models from scratch?
- Does it design custom chips for AI workloads?
A “yes” to either qualifies. OpenAI, Anthropic, Mistral, Cerebras, and NVIDIA all pass.
Companies whose business is hosting, fine-tuning, or orchestrating models built elsewhere do not. Hugging Face has released open models of its own, but its business is distribution and hosting rather than frontier training. LangChain orchestrates calls to models it does not build.
Non-AI
No material AI or machine learning use in core operations. These companies surface in AI-focused deal flow when they raise, then fail verification because there is no real AI underneath.
For the decision trees and the edge cases, see our full breakdown of AI Native, AI Augmented, and AI Adjacent classifications.
| Classification | AI Role | Founded | Example | Investor Signal |
|---|---|---|---|---|
| AI Native | Core product IS AI | 2021+ typically | Perplexity | High growth, high risk |
| AI Augmented | AI enhances existing product | Pre-2023 typically | Salesforce | Stable base plus AI upside |
| AI Adjacent | Enables others’ AI | Any | Scale AI | Picks-and-shovels |
| AI Platforms | Builds foundational AI tech | Any | OpenAI | Capital-intensive, winner-take-most |
| Non-AI | No material AI in the core product | Any | Traditional SaaS | Standard SaaS metrics apply |
Source: Bot Memo classification of 13,500+ AI companies (2023 to 2026)
Core AI Technology Terms: Deep Learning, Language Models, and RAG
Investors hear technical claims daily. These are the terms that separate signal from noise in a pitch deck.
Artificial Intelligence (AI): Software that performs tasks normally requiring human reasoning, such as recognizing images, writing text, or planning a sequence of actions. The umbrella term under which everything below sits.
Machine Learning (ML): The branch of AI where a system learns patterns from data instead of following hand-written rules. Nearly every commercial AI product on the market today is a machine learning system.
Deep Learning: Machine learning built on multi-layer neural networks. Deep learning is what made image recognition, speech, and modern language models work at commercial quality, and it is the reason compute costs became a line item investors care about.
Neural Network: A model made of layered nodes that pass weighted signals to each other, loosely inspired by biological neurons. The substrate under deep learning.
Supervised Learning: Training on labeled examples, where each input has a known correct answer. Accuracy depends on label quality, which is why data labeling became a venture-fundable category in its own right.
Unsupervised Learning: Training on unlabeled data, letting the system find structure on its own. Clustering and anomaly detection are the common commercial uses.
Foundation Model: A large model trained on broad data that can be adapted to many downstream tasks. Credible estimates for frontier training runs have climbed from single-digit millions to roughly $78 million for GPT-4 and $191 million for Gemini Ultra, and the industry expects billion-dollar runs at the frontier tier.
The term matters because it separates the companies that build platforms from the companies building on top of them.
Large Language Model (LLM): A foundation model trained primarily on text and code. GPT-4, Claude, Gemini, and Llama are language models, and they power most of the generative AI products an investor will see.
Artificial General Intelligence (AGI): A hypothetical system matching or exceeding human performance across any cognitive task. No such system exists. Treat AGI in a pitch deck as a statement of ambition, not a product roadmap.
Multimodal Model: A model that handles more than one input type, such as text plus images, audio, or video. Multimodal capability is increasingly the default at the frontier rather than a differentiator.
Fine-Tuning: Adapting a pre-trained model to a specific task using a smaller, targeted dataset. Costs a fraction of training from scratch. A startup claiming “proprietary AI” is often fine-tuning an open-source model.
Retrieval-Augmented Generation (RAG): A technique that feeds relevant external data to a language model at query time, grounding answers in retrieved documents rather than the model’s memory. Common in enterprise products and cheaper than fine-tuning for most use cases.
Context Window: How much a model can consider at once, measured in tokens. Larger windows let a product reason over whole contracts or codebases. The window itself is free; filling it is not, because you pay for every token you send.
Inference: Running a trained model to generate outputs. This is the ongoing cost of operating an AI product, as distinct from one-time training costs. Inference cost per query determines unit economics.
Hallucination: When a model generates plausible-sounding but factually incorrect output. A critical risk factor in regulated industries such as legal, medical, and financial services.
Transformer: The neural network architecture underlying virtually all modern language models, introduced by Google researchers in 2017. When a startup says “transformer-based,” it means the same architecture family as GPT and Claude, which tells you little on its own.
Embeddings: Numerical representations of text, images, or other data that capture semantic meaning. Used for search, recommendations, and similarity matching.
Vector Database: A store optimized for querying embeddings. Pinecone and Weaviate are the venture-backed names in this category.
Prompt Engineering: Designing input instructions to get desired outputs from a language model. A skill, not a product category. Be cautious of startups whose primary moat is prompt engineering.
Agentic AI: Systems that plan autonomously, execute multi-step tasks, and call tools. The dominant investment thesis of 2025 and 2026, and the source of a large share of current deal flow.
AI Accelerator: A chip designed specifically for AI workloads rather than general computing. GPUs, TPUs, and custom silicon from companies such as Cerebras fall here. Access to accelerators is a real constraint on how fast a model company can move.
AI Investment Terms Every VC Should Know
These cover the mechanics of how AI deals actually get done. The dollar ranges below are the bands commonly quoted in the market rather than hard thresholds, and AI rounds have been running at the upper end of each.
Pre-Seed: The earliest institutional funding stage, commonly quoted at $50K to $1M. Companies typically have a prototype or a founding team and little else.
Seed: First significant round, commonly $1M to $5M. Product exists, with early users or revenue.
Series A: First major institutional round, commonly $5M to $25M. Investors usually expect evidence of product-market fit, though in AI the round often arrives on team and traction alone.
Series B: Growth-stage funding, commonly $15M to $60M. Scaling a proven model.
Series C through G+: Late-stage growth rounds. Amounts vary widely, and in AI these frequently exceed $100M.
Growth Investment: Large capital deployment into proven companies, often $100M or more. A round that carries an explicit label keeps it, so a Series D stays a Series D. Growth is the classification for large capital that arrives without a stated series, where the size and the company’s maturity rule out an early-stage round.
Equity Financing: Raising capital by selling ownership rather than borrowing. The default structure for venture rounds, and the reason dilution matters.
Strategic Investment: Funding from a corporation rather than a VC fund, where the investor has business alignment with the startup. Microsoft’s investment in OpenAI is the defining example. Strategic money often carries commercial strings that a financial round would not.
Angel Investor: An individual investing personal capital, typically at pre-seed or seed, usually $25K to $500K.
Startup Accelerator: A fixed-term program that provides small-cheque funding, mentorship, and a demo day, usually in exchange for equity. Y Combinator and Techstars are the reference examples, and accelerator batches have become a visible source of early AI deal flow.
Convertible Note: Short-term debt that converts to equity at a future round. Common in pre-seed and seed.
SAFE (Simple Agreement for Future Equity): Created by Y Combinator. Similar to a convertible note but simpler, with no interest rate and no maturity date. The dominant early-stage instrument today.
Term Sheet: A largely non-binding document outlining the key terms of a proposed investment. Precedes the definitive legal agreements.
Lead Investor: The firm that sets terms, runs primary diligence, and usually takes a board seat. Other participants follow the lead’s terms.
Cap Table: The record of ownership across founders, investors, employees, and option pools. It gets complicated fast after multiple rounds.
Dilution: The reduction in existing shareholders’ ownership when new shares are issued. Founders commonly give up somewhere in the mid-teens to mid-twenties in percentage terms per round, with later rounds usually diluting less than seed.
Valuation (Pre-Money vs. Post-Money): Pre-money is company value before the new investment. Post-money is pre-money plus the new capital. A $10M raise at $40M pre-money means $50M post-money.
Due Diligence: The investigation before an investment closes. For AI companies this adds technical work on model architecture, data rights, and inference costs. Our AI startup due diligence guide walks through what to test.
Startup Metrics That Matter for AI Companies
Standard SaaS metrics apply to AI companies, with a few additions specific to model-driven businesses.
Revenue Metrics
ARR (Annual Recurring Revenue): Annualized value of recurring subscription revenue. The north-star metric for SaaS and AI-SaaS companies.
MRR (Monthly Recurring Revenue): ARR divided by 12. Useful for tracking month-over-month growth.
Revenue Per Employee: Revenue divided by headcount. AI-native companies are structurally leaner than traditional SaaS at the same stage, and the best of them post revenue-per-head figures in the millions, a level that used to belong to chip companies. Investors now treat this as a first-pass test of whether a company’s AI is doing real work.
Cost Metrics
Burn Rate: Net monthly cash out, meaning cash spent less cash collected. AI companies often burn faster because of GPU compute.
Runway: Months of cash remaining at current burn. Cash balance divided by monthly burn.
Unit Economics: Revenue and cost per unit of service delivered. For AI that means cost per query, per API call, or per generated output.
AI-Specific Metrics
Inference Cost: The compute cost of a single prediction or generation. This decides whether margins improve or collapse as usage scales.
Token Cost: Price per token, roughly four characters of English text, for model input and output. Output tokens typically run three to ten times the price of input tokens.
GPU Utilization: The share of GPU capacity actually in use. Low utilization is wasted capital, and with worldwide AI spending running into the trillions, much of it on compute infrastructure, the gap compounds quickly.
Market Sizing
TAM (Total Addressable Market): Total revenue opportunity at 100% market share. For AI the number keeps moving; the global AI market is projected to reach $4.8 trillion by 2033, up from $189 billion in 2023.
SAM (Serviceable Addressable Market): The slice of TAM a startup can realistically target with its current product and go-to-market.
SOM (Serviceable Obtainable Market): The portion of SAM a startup can capture near term. Investors expect founders to work backward from SOM, not forward from TAM.
LTV (Lifetime Value): The gross profit expected from a customer over the whole relationship. Use contribution rather than raw revenue, or the LTV/CAC ratio flatters a low-margin business. AI products with high switching costs tend to hold strong LTV.
CAC (Customer Acquisition Cost): Sales and marketing spend divided by new customers acquired. An LTV/CAC ratio above 3:1 is the standard venture benchmark.
AI Augmented vs AI Native: Why the Distinction Matters
This is the single most consequential classification call in the repository. It sets the valuation framework, the moat argument, and the risk profile.
The “Remove AI” Test
One question decides it: if all AI capabilities were removed, would the product still exist?
- Yes, AI Augmented. Salesforce without Einstein AI is still a CRM. Adobe without Sensei is still Photoshop.
- No, AI Native. Perplexity without language models is nothing. Harvey without AI is a blank screen.
Why Investors Should Care
AI Native companies tend to attract higher valuation premiums because AI is the moat, not a feature.
They also carry concentration risk. If a foundation model provider changes pricing or capability, the product can be disrupted overnight.
AI Augmented companies hold more defensible ground through existing customer relationships and product lock-in. The AI layer accelerates growth but is not existential. If AI costs spike, they can dial features back without killing the product.
Edge Cases That Trip Up Investors
Notion added AI features in 2023. AI Native or AI Augmented? AI Augmented. The core workspace product existed and had product-market fit well before AI.
Cursor forked VS Code and rebuilt it around AI code generation. AI Native or AI Augmented? AI Native. Despite the borrowed editor shell, the entire value proposition is AI-driven code generation. Remove the AI and no differentiation remains.
Why the Call Is Hard
Almost every established software company now ships an AI feature, which means self-description is worthless as a signal. The classification has to be made against what the product does without AI, not against what the marketing site claims.
Market Intelligence Terms for AI Deal Sourcing
These describe how professional investors source, evaluate, and track AI deals, and how the repository is structured to support that work.
Industry Vertical: The primary market a startup serves. Bot Memo tracks 18 consolidated verticals including FinTech, Health and Biotech, Cybersecurity, Developer Tools, and Enterprise SaaS.
Sub-Vertical: A narrower category inside a vertical. “Fraud detection” within FinTech, or “drug discovery” within Health and Biotech.
Technology Theme: The technical approach or market niche a company works in, independent of its vertical. Computer vision, natural language processing, and autonomous systems are themes; the repository tracks more than 105 of them.
Market Map: A visual layout of companies in a sector, organized by category. Used to find white space and to see where competition has already stacked up. Our AI agent market map is an example.
Deal Flow: The rate and quality of investment opportunities reaching an investor. Systematic sourcing is what separates a fund that sees a category early from one that reads about it.
AI Washing: Overstating AI capability to attract a higher valuation. It is common enough that verification is now a standard diligence step, and companies do get reclassified as Non-AI once checked. Our guide on how to spot AI washing covers the tells.
Technical Moat: A defensible advantage rooted in technology. For AI companies that means proprietary training data, specialized architectures, domain-specific fine-tuning, and feedback loops that improve with usage. Note what is absent from that list: access to a general-purpose model everyone else can also call.
Frequently Asked Questions
What is an AI native startup?
An AI native startup is a company built with artificial intelligence as its foundational technology from day one. The product would not exist without AI.
Examples include Perplexity (answer engine), Harvey (legal and professional services), and Midjourney (image and video generation). These differ from AI augmented businesses, which added AI features to products that already worked.
What is the difference between AI native and AI augmented?
AI native companies were built around AI from inception. Remove the AI and the product ceases to exist. AI augmented companies had established products before adding AI features.
The test is one question: could the core product function without AI? Salesforce, an AI augmented company, would still be a CRM. Perplexity, AI native, would be nothing. The distinction drives how investors set valuation, judge risk, and assess the moat.
What does AI adjacent mean in venture capital?
AI adjacent describes companies that enable AI development without running AI models themselves. These are the picks-and-shovels businesses: data labeling platforms such as Scale AI, GPU cloud providers such as CoreWeave, and MLOps tooling companies.
They profit from AI growth without carrying model-development risk, which usually means lower risk and more predictable revenue than an AI native bet.
What is a foundation model in AI?
A foundation model is a large AI model trained on broad datasets that can be adapted to many downstream tasks. GPT-4, Claude, Gemini, and Llama are foundation models.
Training one is capital-intensive. Credible estimates for frontier runs reach roughly $191 million, and the industry expects billion-dollar runs at the frontier. Only companies classified as AI Platforms train foundation models from scratch.
What is RAG in artificial intelligence?
RAG, or retrieval-augmented generation, feeds relevant external data to a large language model at query time, grounding its answers in retrieved documents rather than in what the model memorized during training.
For investors, RAG matters because it costs far less than fine-tuning a custom model and has become the default architecture for enterprise AI products handling proprietary data. Most AI native startups in legal, financial, and healthcare use some form of it.
What is the difference between machine learning and deep learning?
Machine learning is the broad field where systems learn patterns from data instead of following hand-written rules. Deep learning is a subset that uses multi-layer neural networks.
The practical difference for an investor is cost and capability. Classical machine learning runs cheaply on ordinary hardware. Deep learning needs specialized accelerators, which is why compute spend shows up as a structural line item in an AI company’s burn.
How do investors classify AI companies?
Professional investors use classification frameworks to sort AI companies by how central AI is to the business. Bot Memo’s five tiers are AI Native, AI Augmented, AI Adjacent, AI Platforms, and Non-AI, applied across 13,500+ companies.
Classification drives the investment thesis, the valuation benchmarks, and portfolio construction, because each tier carries a distinct risk and reward profile and needs a different diligence approach.
What are the different types of AI startups?
AI startups fall into five categories based on how central AI is to the business. AI Native startups such as Perplexity or Harvey are built entirely on AI. AI Augmented companies such as Salesforce or Adobe added AI to existing products.
AI Adjacent businesses such as Scale AI or CoreWeave provide the infrastructure others build on. AI Platforms such as OpenAI or Anthropic train foundation models from scratch or design AI chips. Non-AI companies have no material AI integration. See the full classification framework for the decision trees.
What does pre-seed vs seed mean for AI startups?
Pre-seed and seed are the two earliest institutional funding stages. Pre-seed rounds are commonly quoted at $50K to $1M and fund initial research, prototyping, or team formation. Seed rounds run roughly $1M to $5M and fund product development with early users or revenue.
For AI startups, pre-seed often covers model experimentation and dataset acquisition, while seed money absorbs training compute and the first production deployment. The distinction matters because AI companies at pre-seed carry higher technical risk: the core model may not work yet.
How This Glossary Is Maintained
Coverage: 13,500+ AI companies across all funding stages and geographies, spanning 2023 to today.
Classification: Every record is independently verified. Operating companies land in one of the five AI tiers above; the remaining four values (Venture Firm, Acquisition, News Article, Insufficient Info) mark records that are not operating companies at all. The definitions are version-controlled and applied consistently across the full repository.
Funding data: All amounts standardized to USD. Companies operating in multiple cities or verticals receive full attribution to each, so category sums intentionally exceed repository totals.
Definitions: Terms are reviewed against how investors and operators actually use them, not against textbook definitions that have drifted from practice.
About the author
Editorial Staff
The Editorial Staff at Bot Memo is a team of writers, analysts, and AI agents dedicated to mapping the global AI startup ecosystem. Led by Chintan Zalani, the team tracks thousands of funding rounds, classifies companies across verticals, and distills it all into actionable intelligence for investors and founders.
