AI and machine learning for business

By November, every department at Birchline wanted to talk about AI. Jules wanted a shopping assistant on the website that could answer questions such as "which sleeping bag is warm enough for a late-September trip in the mountains?" The customer service team wanted help sorting thousands of emails. The buying team wanted better demand forecasts so that stores stopped running out of popular tents in June and drowning in unsold snowshoes in March. Dana wanted to know what any of it would cost and how Birchline would know it was working. Farah asked Owen to sit in on the meetings and help turn the excitement into a short list of sensible projects. This lesson follows that work and covers what the Cloud Digital Leader exam expects about AI: the core ideas, the business problems AI can solve, the importance of data, responsible use, and how to choose between Google Cloud's AI offerings.

AI, machine learning, generative AI and agents

Artificial intelligence (AI) is the broad field of building computer systems that perform tasks normally associated with human intelligence, such as understanding language, recognizing images, making predictions and planning. Machine learning (ML) is the main approach used in modern AI: instead of a programmer writing every rule, a system learns patterns from examples. Shown thousands of past sales together with the weather, holidays and promotions at the time, an ML model can learn to predict next week's sales.

Generative AI is a kind of AI that creates new content, such as text, images, code or audio, in response to a prompt. It is powered by very large models trained on huge amounts of data, often called foundation models; Google's Gemini models are an example. Agentic AI goes a step further. An AI agent is a system that can pursue a goal by planning steps, using tools such as search, databases or other applications, and acting on the results, with some degree of autonomy. A chatbot that answers a question is generative AI; a system that reads a customer's return request, checks the order, applies the returns policy, issues a shipping label and updates the order record is agentic.

It is equally important to know what AI is not. Data analytics examines data to describe what happened and why, such as last month's sales by region. Business intelligence (BI) packages analytics into dashboards and reports for decision makers, as Looker does. Both describe and explain the past using rules and queries that people write. ML predicts or generates using patterns it has learned. A useful test is to ask whether the system follows rules someone wrote (analytics and BI) or learned its behaviour from examples (ML).

Business problems ML can solve

ML is worth using when a problem involves patterns too complex or too numerous for people to write rules for, when there is enough good data to learn from, and when the answer can be checked. Common categories include forecasting (demand, staffing, cash flow), classification (is this email a complaint, a question or an order change?), recommendation (customers who bought this also bought), anomaly detection (this payment looks unlike the customer's normal behaviour), image and document understanding (read the supplier's invoice, check a photo of damaged goods) and personalization (show each shopper the products most relevant to them). Generative AI adds content creation, summarization, conversation and search over an organization's own documents.

The business value usually falls into a few buckets: saving time on repetitive work, making better predictions than people or simple rules can, scaling a service without scaling headcount at the same rate, and creating new experiences for customers. Farah's shortlist for Birchline had three items: demand forecasting for stores (better predictions), sorting and drafting replies to customer emails (saving time), and the shopping assistant (a new customer experience). For each, Dana asked for one measurable outcome: fewer stockouts, shorter email response times, and a higher share of website visitors who buy.

Data quality decides model quality

An ML model can only be as good as the data it learns from. Lesson 3's six quality dimensions matter even more here. Incomplete data leaves the model blind to some situations; duplicated records make some patterns look more common than they are; late data means predictions based on yesterday's world; invalid or inaccurate values teach the model wrong lessons; and inconsistent definitions across systems mean the model learns a muddle. A demand forecast trained on sales data that silently drops returns will consistently over-order products that customers send back.

Training data must also represent the situations the model will face. Birchline's sales history is dominated by its large urban stores, so a model trained on it may forecast poorly for small rural stores unless the team checks performance store by store. Data used for AI also raises privacy questions: customer emails used to train an email classifier contain personal information, and Canadian privacy law and Birchline's own privacy policy govern whether and how they may be used.

# Checking a demand model: invented forecast and actual weekly sales for four stores.
stores = {"Downtown": (420, 400), "Lakeside": (260, 275), "Highway mall": (310, 300), "Small town": (60, 95)}
for name, (forecast, actual) in stores.items():
    error = abs(forecast - actual) / actual
    print(f"{name:13} forecast {forecast:4}  actual {actual:4}  off by {round(error * 100):3} in 100")
overall = sum(abs(f - a) for f, a in stores.values()) / sum(a for f, a in stores.values())
print("Overall the forecast is off by about", round(overall * 100), "units in every 100 sold")

Downtown      forecast  420  actual  400  off by   5 in 100
Lakeside      forecast  260  actual  275  off by   5 in 100
Highway mall  forecast  310  actual  300  off by   3 in 100
Small town    forecast   60  actual   95  off by  37 in 100
Overall the forecast is off by about 7 units in every 100 sold

The overall figure looks respectable, but the small-town store is badly under-forecast. Averages hide exactly the kind of uneven performance that responsible AI practice asks teams to look for.

Explainable and responsible AI

Explainable AI refers to methods that help people understand why a model produced a particular output, for example which factors most influenced a forecast or a loan decision. Explanations build trust, help teams find errors and are sometimes required by regulators or customers. Responsible AI is the broader practice of building and using AI so that it is fair, safe, private, secure, transparent and accountable. Google publishes its own AI principles, and the exam expects you to recognize the general themes: avoid creating or reinforcing unfair bias, test for safety, protect privacy, keep people accountable for decisions and be open about where AI is used.

For Birchline, responsible AI is practical rather than abstract. The shopping assistant must not invent product specifications, so it will answer only from Birchline's own catalogue and say when it does not know. Email drafts will be reviewed by an agent before they are sent. The forecast will be checked for every store, not just on average. And customers will be told when they are talking to an AI assistant. Generative models can produce fluent but wrong answers, often called hallucinations, and grounding them in an organization's own trusted data is one of the main ways to reduce that risk.

Choosing an approach: pre-trained, low-code or custom

Google Cloud offers AI at several levels, and the exam guide lists the factors that should drive the choice: speed of implementation, development effort, differentiation (does a custom model give a real competitive edge?), the expertise available in the team, and the choice and flexibility needed. At one end, ready-made models and APIs can be used in days with little ML skill but offer no special advantage, because competitors can use the same ones. At the other end, custom models take longer and need specialists but can be tuned to an organization's own data and problems.

Pre-trained APIs solve common problems out of the box: the Vision API labels images and reads text in them, the Cloud Translation API translates between languages, the Speech-to-Text API transcribes audio, and Gemini models generate and understand text, images and more. BigQuery ML lets analysts build and run ML models using SQL inside BigQuery, which is ideal when the data is already there and the team knows SQL but not Python. AutoML trains a model on an organization's own labelled data without the team having to design the model itself.

Google brings these together in an all-in-one platform for building with models and agents. The current exam guide calls it the Gemini Enterprise Agent Platform; earlier Google materials and many existing courses refer to the same family of tools as Vertex AI. It provides access to Google's first-party models and partner models, tools such as Agent Studio for building agents and custom models, AutoML, and the services to deploy, monitor and govern them. Underneath sits the AI Hypercomputer, Google's integrated AI infrastructure of GPUs, Google-designed Tensor Processing Units (TPUs), software and flexible ways to pay for capacity. Most digital leaders never touch that layer directly, but it explains why Google emphasizes AI infrastructure as a strength.

Birchline's choices followed the ladder. Email sorting and draft replies use Gemini models through the platform, grounded in Birchline's policies, because the need is common and speed matters. Demand forecasting starts with BigQuery ML, because the sales history is already in BigQuery and Ravi's analysts know SQL. The shopping assistant is the one place Birchline wants differentiation, so it is built as an agent on the platform, connected to the product catalogue and stock levels, with Jules's team owning the experience.

Measuring whether AI is working

Dana's question about how Birchline would know an AI project was working deserves its own answer, because many AI projects stall after an impressive demonstration. Farah's rule is to agree three things before any build starts. First, a business measure: the outcome the project exists to change, such as fewer stockouts in the top hundred products or a shorter average reply time to customer emails. Second, a baseline: the value of that measure today, measured honestly over enough weeks to smooth out normal ups and downs. Third, a decision point: a date and a threshold at which the team will either expand the project, change it or stop it.

Technical measures, such as how often the email classifier picks the right category, matter too, but they are means rather than ends. A classifier that is right nine times in ten creates no value if staff do not trust it and re-sort every email by hand. That is why Birchline's pilots include the people who will use the results, gather their feedback weekly and track adoption alongside accuracy. Projects that cannot name a business measure go to the bottom of the list, however exciting the technology looks.