Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases

Artificial intelligence (AI) is the umbrella discipline, machine learning (ML) is a technique within it, and generative AI is a specialized branch of machine learning — each term is a narrower circle inside the previous one. AI describes any system that performs tasks associated with human intelligence, such as understanding language, recognizing images, or making decisions. Machine learning achieves this by training models on data so they improve from experience instead of relying on explicitly programmed rules. Generative AI goes one step further: rather than only classifying inputs or predicting a number, it produces new content that did not exist before.

Natural language processing (NLP) is the field of AI concerned with understanding and generating human language, covering tasks such as translation, chat, and question answering. Much of NLP is not generative at all: sentiment analysis, text classification, and entity extraction read language and return a label or a list of fields. NLP is also the foundation on which large language models are built. So a model that reads job applications and pulls out each candidate's skills and years of experience as structured fields is NLP and machine learning, but it is not generative AI, because it never produces new content. Working with language and generating content are separate questions.

The reliable test is the output. If the system produces original content, it is generative AI; if it returns a label, a score, or a forecast, it is traditional (predictive) machine learning. One organization usually needs both. An energy utility estimating tomorrow's electricity load for each region from years of meter readings needs predictive ML, because the output is a number derived from historical patterns; the same utility drafting plain-language outage notices tailored to each neighbourhood needs generative AI, because the output is new text. Judge each proposal by what it must produce, not by which technology is newer.

A foundation model is a large AI model pretrained on a vast, broad dataset that can then be adapted to many different downstream tasks — one model serving as the foundation for summarization, drafting, classification, translation, and more. This reuse is the business breakthrough, and it is an example of transfer learning: knowledge learned during pretraining carries over to new tasks. Instead of building a separate model from scratch for every task, which needs large labeled datasets per task, an organization starts from a pretrained foundation model and adapts it with prompting or light tuning, which dramatically reduces the time, data, and cost needed to put AI to work.

A large language model (LLM) is a foundation model specialized in language. It is built in stages. Pretraining is self-supervised: the model reads an enormous body of unlabeled text and learns to predict the next words (tokens) in a sequence, so the text itself supplies the answers and no one has to label anything. A token is the unit of text a model reads and writes — a word or a piece of a word — and an LLM generates its output the same way, predicting one token at a time and appending each to the sequence. That is how it learns grammar, facts, and reasoning patterns, and why a trained LLM can answer questions, draft documents, translate, summarize, and write code. Later refinement steps use smaller, curated data: instruction tuning trains on labeled prompt-and-response examples so the model follows instructions, and reinforcement learning from human feedback (RLHF) turns human ratings of its answers into a reward so it becomes more helpful and safer. Google's Gemini family is the flagship example, and Gemma is Google's family of lightweight open models that organizations can run and adapt themselves.

A multimodal foundation model can take in and reason across more than one type of data — text, images, audio, and video — in a single model. Gemini is natively multimodal: you can show it a photo of a shelf display and ask for a written compliance report, or give it a short product video and ask for a step-by-step written guide. Do not confuse multimodal with multilingual. Multilingual means handling several human languages; a help-desk assistant that answers typed questions in French, Hindi, and Portuguese is still working with one modality, text.

Multimodality matters when the answer depends on how the inputs relate to each other. Consider an equipment-maintenance review that must catch when a technician's written report says a part was replaced but the inspection photo still shows the worn one. If one model first turns the photo into a short caption or label and a second model reads that caption alongside the report, the second model never sees the image itself, so any detail the caption leaves out is lost, and the team now maintains two models instead of one. A single multimodal model that receives the photo and the report together can check one against the other directly. For business leaders, multimodality matters because most enterprise information is not plain text; a model that reads documents, screenshots, and recordings unlocks far more of an organization's knowledge.

A diffusion model is a type of generative model that creates images by learning to reverse noise: during training it watches images being progressively degraded into random static, and it learns to run that process backwards, refining pure noise step by step into a coherent picture that matches a text description. This is the technique behind modern image and video generation. It works differently from a language model, which produces text one token at a time by predicting what comes next; a diffusion model refines a whole image over many denoising steps.

Diffusion is not the only generative approach for images. Generative adversarial networks (GANs) train two networks against each other — a generator that makes images and a discriminator that tries to tell them from real ones — and produce an image in a single pass. Variational autoencoders (VAEs) compress images into a compact representation and learn to generate new images by decoding from it. The feature that identifies a diffusion model is the iterative denoising: starting from noise and removing it step by step.

In the Google Cloud portfolio, Imagen is the foundation model for generating and editing high-quality images from text prompts, and Veo is the foundation model for generating video. A marketing team can use Imagen to produce campaign visuals and product imagery without a photo shoot, and use Veo to storyboard or generate short video content — compressing creative cycles from weeks to hours.

You do not need the mathematics. What matters is the mapping: language tasks point to large language models such as Gemini, image generation points to diffusion-based models such as Imagen, and video generation points to Veo. When a scenario describes creating visual assets from a written brief, diffusion models and Imagen are the concepts involved.

Prompt engineering and prompt tuning are both ways to get better output from a foundation model without retraining it from scratch, but they work at different levels. Prompt engineering is the human craft of writing effective instructions: giving the model a clear task, audience, context, examples, and a desired format. It requires no training or infrastructure — anyone can do it in the Gemini app — and it is usually the first, cheapest lever an organization pulls to improve results.

Prompt tuning is a parameter-efficient tuning technique: instead of a person rewording instructions, the system learns a soft prompt — a small set of numeric vectors, not readable words — from labeled examples and adds it to every request. The base model's weights stay frozen; only the soft prompt is learned. That makes it far cheaper than changing the model, while giving more consistent task-specific behavior, such as a house style, than hand-written prompts alone. Full fine-tuning, by contrast, updates all of the model's weights on your data, which costs the most and produces a new version of the model. Prompt tuning is a concept you need to recognize by name; on Google Cloud, the main managed tuning option for Gemini models is supervised fine-tuning that uses a parameter-efficient method, so only a small number of added parameters is trained (preference tuning is also available for some Gemini models).

Grounding is a different lever again. Retrieval-augmented generation (RAG) finds relevant documents at request time — from a company knowledge base or search — and adds them to the prompt so the model answers from current, specific facts. It changes what the model knows for that request; it does not teach the model a behavior and does not change its weights.

For a leader, the decision logic runs in order of cost and effort: start with prompt engineering (no training, immediate), add grounding when the problem is missing or outdated knowledge, move to prompt tuning when you need repeatable task-specific behavior at scale, and reserve fine-tuning for cases where the lighter options cannot reach the required quality. The right choice is usually the lightest technique that meets the need and the organization's constraints.

Machine learning models are trained in three main ways, and the difference comes down to what kind of feedback the model receives. Supervised learning trains on labeled data — examples paired with the correct answer — so the model learns to map inputs to known outputs. It powers spam filtering (emails labeled spam or not), fraud detection (transactions labeled fraudulent or legitimate), and sales forecasting.

Unsupervised learning trains on unlabeled data and discovers hidden structure on its own. There is no correct answer provided; the model finds groupings and patterns. The classic business example is customer segmentation — clustering shoppers by behavior without anyone defining the segments in advance — along with anomaly detection and market-basket analysis.

The two approaches complement each other. A supervised model can only recognize patterns that appear in its labels: an intrusion-detection model trained on labeled examples of known network attacks has no way to recognize a technique nobody has labeled yet. Unsupervised anomaly detection needs no labels, so it can flag network activity that simply departs from normal behavior, surfacing an unfamiliar pattern early; security analysts then investigate what was flagged, and the cases they confirm become labels that can later improve the supervised model.

Reinforcement learning trains an agent through trial and error: the agent takes actions in an environment and receives rewards or penalties, gradually learning a strategy that maximizes reward. It suits sequential decision problems such as recommendation systems, robotics, and resource optimization. It also refines large language models: in reinforcement learning from human feedback (RLHF), human reviewers judge the model's responses for qualities such as helpfulness and safety, a reward model learns from those judgments, and the LLM is then optimized to earn a higher reward.

Two related terms sit outside these three approaches. Self-supervised learning creates its own labels from unlabeled data — predicting the next word in a sentence is the example — and is how foundation models are pretrained. Transfer learning reuses what a model learned on one task for another, which is what adapting a foundation model does. The cue is the data and the feedback: labeled examples with known answers signal supervised learning, unlabeled data explored for patterns signals unsupervised learning, and learning by reward and penalty over repeated attempts signals reinforcement learning.

The machine learning lifecycle is the end-to-end journey a model travels: data ingestion, data preparation, model training, model deployment, and model management. Understanding it matters to leaders because most gen AI project cost and risk lives in the data and management stages, not in the model itself.

Data ingestion collects raw data from source systems — databases, applications, documents, event streams. Data preparation cleans, transforms, deduplicates, and organizes that data so it is usable for training and so evaluation is honest. Poor preparation is the most common reason AI projects underdeliver. Model training teaches the model from the prepared data, or adapts a pretrained foundation model to your task. Model deployment puts the trained model into production behind an endpoint, or runs it in batch, so applications and agents can use it. Model management is the ongoing work: monitoring quality, tracking versions, watching for drift as real-world data changes, and retraining when performance degrades.

Gemini Enterprise Agent Platform (Agent Platform for short, the evolution of what was previously branded Vertex AI) is Google Cloud's unified AI platform, and it supports training, deployment, and management from one environment. Two near-misses are worth separating. Looker is a business intelligence and reporting tool for exploring data and building dashboards; it does not run the pipelines that prepare training data. Cloud Monitoring alerts on an endpoint's operational health, such as latency, error rates, and uptime; a model can be fast and error-free while its predictions get steadily worse, so those alerts do not detect drift. Model monitoring watches the data the model receives.

Placing a problem in the right stage depends on timing. Drift is a gradual change after launch: the real world moves away from the data the model was trained on, so quality declines over weeks or months, and the fix belongs to model management — monitor, alert, and retrain. A model that underperforms from its first day in production and never matched its test results points earlier in the lifecycle. For example, if a source system that retries failed uploads wrote some machine sensor readings twice and nobody removed the repeats, identical records can land on both sides of the train/test split; the model is then partly tested on examples it has already seen, so the evaluation overstates quality. That is a data-preparation failure, and deduplicating before splitting the data is what keeps evaluation honest.

Remember the order and the ownership: the lifecycle is a loop, not a line, and management feeds back into new data and retraining for as long as the model serves the business.

Choosing a foundation model is a structured trade-off across eight factors, not a search for a single best model: match each requirement in the scenario to the factor it tests, and often more than one factor applies at once.

The practical pattern: a lightweight, lower-cost model often wins for simple high-volume tasks, while a frontier multimodal model with a large context window wins for complex reasoning over rich inputs. Set the quality bar first. When several models clear that bar, the remaining differences in quality matter less than price at high volume, so cost becomes the deciding factor for simple, repetitive work. Requirements also combine: answering questions about a long engineering manual supplied as scanned pages full of diagrams depends on both modality (images of pages, not plain text) and context window (the whole manual at once).

How the model is served matters too. Batch prediction sends a large set of requests to be processed together and returns the results later; it costs less than online requests and suits overnight jobs with a deadline, such as scoring a quarter's worth of product reviews for sentiment ahead of a planning meeting. Because it is asynchronous, it cannot serve a user waiting for an answer — real-time needs call for online prediction. Matching requirements to factors, rather than defaulting to the most capable model, is the skill being built here.

Google Cloud frames gen AI business value in four verbs — create, summarize, discover, and automate — and exam scenarios are usually a thin disguise over one of them. Create covers generating new content: text generation for marketing copy and proposals, image generation for creative assets, code generation to accelerate developers, and video generation for training or advertising material.

Summarize condenses long content into short, decision-ready form: distilling a fifty-page report into an executive brief, recapping meetings, or compressing a support ticket history before an agent picks it up. Discover surfaces knowledge that already exists but is hard to find: asking questions across an organization's documents, analyzing data in plain language, and spotting patterns a person would miss. Automate removes manual steps from a workflow: drafting replies, extracting fields from invoices, classifying feedback, and powering personalized user experiences that adapt content to each customer.

Consider a national retailer. It uses gen AI to create product descriptions and seasonal campaign imagery, summarize thousands of customer reviews into weekly insight reports, discover answers from internal merchandising policies through conversational search, and automate personalized email offers for each shopper segment. One foundation model strategy serves all four patterns — which is exactly why foundation models changed the economics of enterprise AI.