Innovating with AI

Definitions

Relationship: AI > ML > deep learning > generative AI, each nested inside the one before. Analytics and BI describe what happened; ML predicts and generates.

Kinds of ML

• Supervised: learns from labelled examples (emails marked spam or not). Classification and regression.
• Unsupervised: finds structure in unlabelled data (customer segments).
• Reinforcement: learns by trial, reward and penalty.

Predictive AI forecasts or classifies. Generative AI produces something new. Both can appear in one solution.

How agentic AI changes work

Agents go beyond chat. They can break down a goal, use tools and systems, and carry out multi-step tasks, with people supervising.

Key benefits of Google Cloud AI

• Best infrastructure for AI: AI Hypercomputer with TPUs and GPUs.
• AI-ready data cloud: BigQuery and other data services connected to AI so data need not be copied.
• Sophisticated first-party models: Gemini and other Google models, plus a wide Model Garden of third-party and open models for choice.
• All-in-one AI developer platform: Gemini Enterprise Agent Platform covers model choice, tuning, building agents, deployment and governance in one place.
• Prebuilt AI agents and applications: ready-to-use products such as Gemini Enterprise for employees and customer engagement agents, so teams need no model building.

Business problems ML solves

ML fits when rules are too many or too complex to write, when patterns are hidden in large data, or when decisions must scale.

Other use cases: demand forecasting, churn prediction, predictive maintenance, document processing, image inspection, translation, content generation. If a problem has a simple fixed rule, ML is not needed. ML needs data; with no usable data, it is the wrong tool.

Data quality

AI is “garbage in, garbage out”. Models trained on poor data give poor, biased or unreliable results.

Quality problems cause wrong predictions, bias and lost trust. Data governance and cleaning pipelines come before model building.

Explainable and responsible AI

Explainable AI means people can understand why a model produced a result. Responsible AI means building and using AI fairly, safely and accountably.

Business implications:

• Trust and adoption: customers and staff use systems they understand.
• Regulation and compliance: some industries require explaining decisions (lending, hiring, health).
• Bias and fairness: biased training data causes unfair outcomes and reputation damage.
• Privacy and security: protect personal data and guard against misuse and attacks.
• Accountability: humans stay responsible; keep human oversight for high-impact decisions.
• Reliability of generative AI: models can hallucinate (produce confident but false content). Grounding answers in trusted company data reduces this.

Google publishes AI Principles and offers tools such as Explainable AI features in the platform and Model Armor to screen prompts and responses.

Choosing Google Cloud AI solutions

Think of a spectrum from ready-made to custom.

Considerations the guide names: implementation speed, development effort, potential for differentiation, technical expertise required, and choice and flexibility. Rule of thumb: use the most ready-made option that meets the need, and move toward custom only when the use case is a source of competitive advantage or no ready option fits.

Gemini Enterprise Agent Platform

The unified platform (formerly Vertex AI) for building, deploying and managing AI models and agents.

• Model Garden: catalogue of Google, partner and open models to choose from (choice and flexibility).
• Agent Studio (formerly Vertex AI Studio): a place to prompt, test, tune and prototype with models, including low-code agent building.
• Agent building and orchestration tools: build agents that use tools and company data, with grounding, search and connectors to enterprise systems.
• AutoML: train custom models on your own data with no coding.
• Custom training and MLOps: for data scientists who want full control; pipelines, model registry, monitoring.
• Governance and security: access control, evaluation and monitoring for models and agents.

Business use cases: customer-service agents, document summarisation and extraction, search over internal knowledge, code assistance, marketing content generation, personalised recommendations.

Pre-trained APIs and foundation models

Pre-trained APIs need no training data and no ML expertise; you call them. Gemini is multimodal: one model handles text, images, audio, video and code.

Building custom models on your own data

When general models are not enough, an organisation can create value by using its own data:

• AutoML on Agent Platform: upload labelled data and the system builds and tunes a model for you. Good for teams with data but few ML specialists.
• Agent Studio: tune a foundation model with company examples so it matches your tone, terms and tasks, and ground it in your data.
• Custom training: full control for expert teams.

Custom models turn proprietary data into differentiation competitors cannot copy.

AI Hypercomputer

An integrated system for training and serving AI at scale. Components:

Benefits: better performance and efficiency, lower cost per training run, and control over spend. GPUs are general-purpose accelerators; TPUs are Google-designed for ML.

BigQuery ML

BigQuery ML lets analysts create and run ML models using standard SQL, directly where the data lives in BigQuery.

• No data movement and no separate ML platform to learn.
• Opens ML to people who know SQL, not only data scientists.
• Supports forecasting, classification, regression, clustering, recommendations and calls to generative models.
• Fast experimentation: CREATE MODEL then ML.PREDICT in a query.

Trap: if the scenario says “analysts who know SQL want to build a model on warehouse data”, the answer is BigQuery ML, not custom training.