Choosing an AI solution

Buy what is common, and build what sets you apart

You bringReady inSets you apartPre-trained APIa finished skillonly the requesthoursno: everyone has itFoundation modelGemini, with a prompta good promptdaysa littleTuned or AutoMLtrained on your datayour labelled examplesweeksyes: it knows your businessCustom modelbuilt from the startdata, code, an ML teammonthsmost: it is yours alonefaster and easiermore your ownThe less you build, the sooner it works. The more you build, the more it is yours alone.A common task, such as translation, should be bought ready made. Something that sets the business apartis worth training on its own data. Most organisations use several of the four at once.

From a pre-trained API in hours to a custom model in months: speed falls as differentiation rises.

Choosing an AI solution · Going deeper

Five things to weigh

Five things to weigh when choosing an AI solutionAskWhat the answer points toImplementation speedHow soon must it be working?soon: a pre-trained API or GeminiDevelopment effortHow much building can we afford?little: a managed productDifferentiationMust it be better than what rivals can buy?no: ready-made is enoughTechnical expertiseDo we have ML engineers?no: AutoML or an APIChoice and flexibilityDo we need many models, or to change later?yes: a platform with many modelsSpeed, effort and expertise push toward ready-made. Differentiation pulls toward your own model.Choice and flexibility is the case for a platform that offers many models, Google's and others, in one place.

Speed, effort, differentiation, expertise, and choice and flexibility.

Pre-trained APIs

A pre-trained API is a finished skill that you call

a photo of a receiptVision APIthe shop, date and total, as texta recorded phone callSpeech-to-Text APIa written transcripta help page in JapaneseCloud Translation APIthe same page in Englisha question, with a video attachedGeminian answer written for youA pre-trained model is a finished skill. You send it your content and get the result, with no training.The first three do one thing each: see, hear, translate. Gemini is a foundation model: one model that readstext, images, audio and video and can be asked for almost anything. All four are called through an API.

The Vision API sees, Speech-to-Text hears, Cloud Translation translates, and Gemini does many things.

Pre-trained APIs · Going deeper

APIs and foundation models

Pre-trained APIs and foundation modelsVision APIFinds objects, text, faces and unsafe content in imagesSpeech-to-Text APITurns spoken audio into text, in many languagesCloud Translation APITranslates text between languagesGeminiGoogle's family of foundation models: multimodal, for reasoning and generationAgent Platform APIThe one way to call Gemini and other models from your own applicationsFoundation modelA large model trained on broad data, adaptable to many tasks without retrainingA single-purpose API does one job well. A foundation model does many jobs from a prompt."Multimodal" means it takes more than one kind of input: text, images, audio, video and code.

Single-purpose APIs do one job. A foundation model does many from a prompt.

Check yourself

recordings ofsupport calls?written transcriptsthat can be searched

A contact centre wants what the drawing shows. Which product belongs at the "?"

The Agent Platform

One platform, from choosing a model to running an agent

GEMINI ENTERPRISE AGENT PLATFORMChoose a modelGemini, and manyothers, in one placeAgent Studiotest prompts, and tunewith your own dataBuild an agentconnect tools andcompany dataRun and governdeploy, scale, watchand secure itAutoML: train a custom model from labelled data, no codeAgent Platform API: call models from your applicationsOne platform takes an idea from choosing a model to running an agent in production.It was called Vertex AI until April 2026, and older material uses that name. Its value to a business is thatmodels, data connections, agents, security and monitoring are in one governed place, so a prototype doesnot have to be rebuilt to go live.

Gemini Enterprise Agent Platform: choose a model, use Agent Studio, build an agent, then run and govern it.

The Agent Platform · Going deeper

What gets built on it

What organisations build on the Agent PlatformUseFor exampleCustomer service agentsAnswer from company policy and complete the requestSearch over company knowledgeOne question across documents, tickets and wikis, with sourcesContent generationProduct descriptions, marketing drafts, summaries of long reportsAssistance for developersExplaining, writing and reviewing codePredictions from your own dataDemand forecasts and risk scores from custom or AutoML modelsGrounding an agent in company data is what turns a general model into a useful one.A model alone knows the world up to its training. Connected to your data, it knows your business today.

Service agents, search over company knowledge, content, code help and predictions.

Custom models

Three ways to teach a model your own data

BIGQUERY MLfor people who know SQLCREATE MODEL shop.churnOPTIONS (model_type = "logistic_reg")AS SELECT * FROM shop.customers;A model made with one SQL statement.The data never leaves the warehouse.AUTOMLno codeyour labelled examplesa trained modelAGENT STUDIOstart from Geminia foundation modeland your examplesa model tunedto your businessThree ways to put your own data into a model, none of which needs a team of ML engineers.BigQuery ML suits analysts whose data is already in BigQuery. AutoML suits a team with labelled data andno ML code. Agent Studio suits a team that wants a foundation model to speak in its own terms.

BigQuery ML uses SQL, AutoML uses no code, and Agent Studio tunes a foundation model.

Custom models · Going deeper

The three compared

Custom models comparedBigQuery MLAutoMLAgent StudioWho uses itdata analyststeams without ML engineersdevelopers and prompt writersYou supplytables in BigQuery, and SQLlabelled tables, images or textprompts and example answersIt producesa predictive model in BigQuerya custom predictive modela tuned generative modelTypical useforecast sales, score customersspot defects in photosanswer in the house style"Using standard SQL" in a question means BigQuery ML. "Without writing code" means AutoML.Both phrases come straight from the way Google describes the two products.

"Standard SQL" means BigQuery ML. "Without code" means AutoML.

Check yourself

The datafive years of customer records, already in BigQueryThe teamanalysts who write SQL every dayThe aimpredict which customers are about to leave

A retailer describes its situation as shown. Which product fits best?

AI Hypercomputer

Chips, software and ways to pay, designed as one

HardwareGPUs, and TPUs, the chips Google designs for AISoftwareopen frameworks and open standards, tuned to the chipsConsumptionpay on demand, reserve ahead, or use spare capacityDesigned togetherso models train andanswer faster, at acost you can controlThe AI Hypercomputer is the chips, the software and the ways to pay, designed as one system.It is what runs Gemini and what customers rent to train and serve their own models. The third layer isthe business point: AI computing is expensive, and how you buy it changes the bill as much as the chip.

The AI Hypercomputer: GPUs and TPUs, open software, and flexible consumption to control cost.

AI Hypercomputer · Going deeper · 1 of 2

GPUs, TPUs and ways to pay

Inside the AI HypercomputerGPUA graphics processor: the general-purpose chip for AI, supported by every frameworkTPUTensor Processing Unit: Google's own chip, built only for machine learningOpen softwarePopular open source frameworks run on it, so work is not tied to one supplierOn demandPay by the second for what you use, with no commitmentCommitted usePromise a level of use for a year or more, for a lower priceSpot and schedulingSpare capacity at a discount, or a queued start when chips are freeTPU is the term to remember: it is Google's chip, and no other cloud has it.Dynamic Workload Scheduler, the queued start, returns in lesson 10 as a way to control cost.

A TPU is Google's own AI chip. Committed use and spot capacity lower the bill.

AI Hypercomputer · Going deeper · 2 of 2

Clue words for AI offerings

When the question saysThe answer isBecause it is"extract text or objects from images"Vision APIpre-trained for images"transcribe calls"Speech-to-Text APIaudio to text"content in many languages"Cloud Translation APItext to text"multimodal", "generate", "reason"Geminia foundation model"build, deploy and govern agents"Gemini Enterprise Agent Platformone platform, end to end"prototype prompts", "tune a model"Agent Studioworkspace on the platform"custom model, no ML expertise"AutoMLtrains from labelled data"build a model using SQL"BigQuery MLML inside the warehouse"GPUs and TPUs", "train at scale"AI Hypercomputerinfrastructure for AI"fastest to implement"A pre-trained APInothing to train"unique to our business"A custom or tuned modeltrained on your dataDecide first how much is to be built. Then the product follows from the phrase.

Decide how much to build, then match the phrase to the product.

Check yourself

The needput 4,000 help pages into six languagesThe teamno machine learning specialistsThe datelive next month

A company describes its need as shown. What should it do?

Lesson complete

You finished “Google Cloud AI offerings”

Answer the questions along the way to see your score here.

Open in a new tab