Generative AI, Explained: The 18% Loud Category That Created the Wave

By Diego Navia · BizBlocz · April 2026

Part of the AI Explained series. Start with the overview →

Generative AI is the most visible category of enterprise AI. It is also the one most people mean when they say "AI," and the one whose share of total enterprise AI value is smaller than the headlines suggest.

You type a prompt. Write a cover letter for a product manager role. Summarize this 40-page contract in three paragraphs. Draw a logo for a yoga studio. A response comes back. Fluent. Coherent. Usable, often on the first try. The leap in capability over the past three years feels dramatic because it is. ChatGPT, Claude, Gemini, Microsoft Copilot, and the "AI features" that appeared inside email, CRM, and word processors after 2023 are all the same category doing the same job at scale.

That visibility is what built the wave. It is also what makes the calibration question harder.

What generative AI actually does

Generative AI creates. Text, code, images, audio, video, slides. The deliverable at the end of the task is language or pixels.

At the core of generative AI is a class of models called transformers, introduced in a 2017 paper from Google Research titled Attention Is All You Need. A transformer is trained on a very large dataset of human-produced material (text from the web, code from open repositories, images paired with captions). During training, the model learns the statistical patterns in that material: which word tends to follow which, which shapes tend to appear together, how a sentence is usually structured. After training, when given a prompt, the model generates the most statistically plausible continuation, one token at a time.

The mechanism that made this work is called attention. When generating each next word, the model weighs which parts of the input matter most. Attention is the reason a modern model can hold context across thousands of words and produce output that reads as coherent rather than disjointed.

Three common tags describe variations on this core pattern. LLM (large language model). GPT (generative pre-trained transformer). Diffusion (the architecture behind most image generation). All of them are doing the same thing at a mechanical level: pattern completion at scale.

What it looks like from the outside is something much closer to creativity.

What it does not do is worth naming, because the market keeps mixing it up.

It does not retrieve facts. It produces plausible-sounding statements that may or may not be correct. A model asked a question it has no grounded answer to will usually produce a confident, well-structured, incorrect answer. This is commonly called hallucination.

It does not reason numerically. A generative model can write text that looks like arithmetic or financial reasoning, but it is not calculating. It is generating tokens that resemble calculation. Reliable numerical work requires bolting on a calculator, a code-execution environment, or a structured tool.

It does not take actions. It can draft an email but does not send it. It can describe how to update a record but does not update the record. Taking actions falls into the territory of agentic AI.

It does not know any specific company. Its knowledge is limited to its training data, which is public material up to a cutoff date. Policies, contracts, product details, client history are not in there. Giving a model access to a company's own documents requires a separate architecture called retrieval-augmented generation (RAG).

Each of these gaps can be closed with additional layers. None of them are solved by the model alone.

Commercial products

The market has consolidated around a relatively small number of foundation model providers, with a much larger set of enterprise platforms embedding those models into existing workflows. Three layers carry the category.

Foundation models. OpenAI (GPT-4, o1, o3) is the most widely deployed family, with broad capabilities across text, code, and reasoning. Anthropic Claude is often chosen for long-document analysis, careful reasoning, and regulated-industry use cases. Google Gemini brings strong multimodal capabilities and deep Google Workspace integration. Meta Llama is the open-weight option commonly used where on-premise deployment or data sovereignty matters. Mistral is the European provider often selected for EU data-residency requirements.

Enterprise platforms. Microsoft 365 Copilot is the most widely adopted enterprise generative AI deployment to date, embedded across Word, Excel, Teams, and Outlook. Salesforce Einstein and Agentforce sit inside the CRM for customer communication and opportunity summarization. ServiceNow Now Assist runs in IT and HR service management. SAP Joule sits inside ERP for finance, HR, and procurement. Workday AI handles HCM workflows like job-description drafting. ChatGPT Enterprise and Claude for Work serve the BYOLLM pattern where the company licenses the foundation model directly.

Enabling technologies. Retrieval-augmented generation (RAG) grounds model output in a company's own documents and is the dominant architecture for enterprise knowledge Q&A. Fine-tuning further trains a base model on company-specific data to improve performance on terminology and output formats. Reasoning models (OpenAI o1 and o3, Claude extended thinking) use additional computation at inference time to work through multi-step problems.

The pattern under all three layers: the dominant techniques (transformers, attention, RAG, fine-tuning) have been stable since 2022. The churn is in which model wins which workload, who hosts it, and how it is integrated into the workflow. Not in the math.

Examples in action

A marketing manager asks a model to produce twenty variants of an ad headline for a product launch. Thirty seconds later, twenty options appear, each in a slightly different tone. The manager picks three to test.

A procurement team uploads a 180-page RFP and asks for a draft response. Using the team's past proposals as a reference library, the model produces a first draft in minutes. The team spends its time editing, negotiating specifics, and reviewing numbers, the parts that require judgment.

A developer types a comment describing what a function should do. The IDE completes the function. The developer reviews, adjusts, moves on.

A customer support agent receives an incoming email. The system drafts a reply grounded in the company's help center articles. The agent reviews, edits two sentences, sends.

A credit officer pastes a 40-page loan agreement into a document assistant and asks for a summary of covenants, termination clauses, and change-of-control provisions. The summary is a starting point for review, not the final answer.

The common thread: the deliverable at the end of each task is language. Language is what generative AI does.

Where it fits well

Generative AI is the leading technology for a specific and identifiable set of enterprise tasks. The common pattern is consistent: the deliverable is language or a language-shaped artifact.

First-draft document production. Contracts, proposals, RFP responses, board memos, job descriptions, management commentary. Generative AI compresses the time from brief to first draft. The human role shifts toward editing, judgment, and approval.

Summarization. Earnings call transcripts to three paragraphs. Long due-diligence documents to a red-flag memo. Thousand-message support threads to an escalation brief. Summarization is the task attention mechanisms were originally designed for.

Knowledge Q&A over a company's own corpus. Policy questions, contract clause searches, past-project lookups. With RAG, a generative model can answer from the organization's own documents rather than the open internet. Answer quality depends heavily on the quality of the retrieval layer and the hygiene of the underlying documents.

Customer communication drafting. Service response drafts, personalized outreach, multilingual communications, complaint acknowledgments. The output is language addressed to a person, and quality is measured in tone and clarity.

Code generation. Boilerplate, test scaffolding, documentation, explanation of legacy code. GitHub's 2023 productivity research found a 55% velocity improvement on well-defined coding tasks for developers using Copilot.

High-volume marketing production. Product descriptions, ad variants, email campaign material, social content drafts. Low-stakes, high-volume language where speed and variation matter.

Where another category leads

The same test (what is the deliverable?) points elsewhere for a surprising number of enterprise processes.

Invoice processing produces a structured data record extracted from a document. Document AI leads. Generative AI helps draft exception communications.

Demand forecasting and planning produces a numerical forecast. Machine learning leads with decades of proven performance.

Payment execution, reconciliation, and close produce actions taken across systems. Agentic AI leads, supported by ML and rules.

Defect detection on a production line produces a visual classification. Computer vision leads. Generative AI has no primary role.

Fraud detection, risk scoring, credit decisioning produce a scored decision with a required audit trail. Machine learning leads, with regulatory explainability requirements that limit direct use of generative output.

The pattern is consistent. When the deliverable is a number, a decision, a structured record, or a physical action, a different category of AI usually does the primary work, and generative AI plays a supporting role at best.

Why generative AI is 18% of enterprise AI value

Across 127 enterprise subprocesses we mapped, generative AI is the leading technology in eleven of them and contributes roughly 18% of aggregate enterprise AI value. Larger than NLP (14%), document AI (9%), and computer vision (5%). Smaller than agentic (22%) and meaningfully smaller than machine learning (32%).

The 18% captures where generative AI is the leading technology, the eleven subprocesses where the deliverable is language and GenAI does the primary work. The number understates the category's full role. As the language layer of the AI stack, generative AI also acts as the connective tissue across subprocesses where another category leads. The discipline of work redesign has another name for these connections: handoffs. For thirty years the handoffs that made or broke a process flowed between humans. Generative AI is now the dominant medium through which those handoffs happen, whether they connect two humans, a human and an agent, or two systems exchanging structured data wrapped in language. It explains the prediction an ML model produces. It summarizes the action an agent takes. It narrates the extraction a document-AI pipeline pulled out of an invoice. It is the interface most enterprise users will ever experience to the work the other five categories actually do.

That cross-cutting role is something no other AI Six category carries to the same degree. Machine learning predicts and stops. Computer vision classifies and stops. Document AI extracts and stops. Generative AI extends into the supporting layer of nearly every other category, which is why its visibility outruns its primary share. The eighteen percent is what GenAI leads. The headline-grabbing presence in every Copilot demo, every CRM screen, every IT ticket update, every executive briefing summary, is the connective-tissue layer working in the background of all the other categories.

That distribution is the most useful piece of calibration for an enterprise AI portfolio. Generative AI is a meaningful contributor as a leading technology and a much larger contributor as the language layer on top of everything else. The MIT NANDA 2025 study, which found that 95% of generative AI pilots delivered no measurable impact on the bottom line, points to a related conclusion from a different angle: the technology works, but it is often deployed as the leading technology in places where language is not what the process actually needs, and as the connective layer in places where the underlying category was never properly funded.

The related question is which enterprise platforms own the process IP for the specific language-shaped tasks where generative AI does win. That is the fight we covered in the SaaSpocalypse piece, and generative AI sits on both sides of it. Incumbents with deep workflow context hold one kind of moat. Foundation-model providers with frontier capability hold another. The procurement question is which one your specific process actually needs.

The practitioner angle

Alan Turing imagined, in a 1950 paper, a future where a person on one end of a text conversation could not tell whether the response on the other end came from a human or a machine. He proposed that as a workable definition of intelligent. Generative AI is the first technology in history that passes a casual version of that test at scale.

The discipline is remembering what is still missing. Fluent does not mean correct. Coherent does not mean grounded. Persuasive does not mean defensible. The failure mode of generative AI is a confident wrong answer, and the engineering work to catch that failure mode (retrieval grounding, evaluation harnesses, escalation rules, human review on consequential outputs) is the part of generative AI deployment that most pilots skip and most production systems require.

The 95%-no-measurable-impact figure from MIT NANDA is, in part, a measurement of how often that engineering work was skipped.

Which of your processes have a deliverable that is genuinely language, where the value comes from speed of first draft and quality of the editing loop? And which ones got dressed up as language work because generative AI sold better than workflow or prediction in the 2026 budget cycle?

Next in the series: Machine Learning, Explained — the oldest and largest category of enterprise AI, the leading technology behind 32% of aggregate AI value, and the one almost nobody calls AI anymore. → ML Explained

Also in the AI-Explained series: Agentic AI, NLP, Document AI, Computer Vision. Start with the overview →

Sources: Vaswani et al., "Attention Is All You Need," Google Research (2017). MIT NANDA, "State of AI in Business 2025." GitHub, "Research on the productivity impact of GitHub Copilot" (2023). Gartner Hype Cycle for Artificial Intelligence (2025). Subprocess-level estimates are BizBlocz aggregate research, an analysis of 127 enterprise subprocesses and 245+ data points across 30+ independent research publications. Directional, not decimal-precise.