Exam-style
A homeware retailer approves four projects. (1) Marketing needs product shots for a seasonal catalogue that was never photographed, a fifteen-second clip for social, and a spoken voiceover over the clip. (2) Support wants every recorded call's transcript reduced to a paragraph a supervisor can skim before a callback. (3) Every inbound support email has to land in one of eleven fixed queues, and any email from an account on the priority contract has to skip the queue and page the duty manager. (4) The storefront wants an assistant a shopper can talk to, so that a question like what do I need to start baking sourdough comes back as a conversation that names products from the catalogue. Which of the four is not a job for a foundation model?
Reveal the answer
B. The third, because sorting email into eleven fixed categories is a text classification problem for a purpose-built AI service, and the priority-contract rule is a lookup against an account field that needs no model at all
The use cases named for generative AI are image, video, and audio generation; summarization; chatbots and other AI assistants; translation; code generation; customer service agents; search; and recommendation engines. Hold that list next to the four projects. Marketing’s catalogue shots, the fifteen-second clip and the voiceover are image, video, and audio generation, three modalities of one job, and a multi-modal model on Amazon Bedrock covers all three. The call paragraph is summarization, the most reliable thing a foundation model does, because the source text is in front of it and the work is compression rather than recall. The storefront assistant sits where several of those use cases meet: it is one of the chatbots on the list, it behaves as a customer service agent when the conversation turns to a delayed delivery, it does search across the catalogue in language a shopper actually uses, and it ends in recommendation engines territory when it names three products. That leaves the third project. Reading an email and deciding which of eleven queues it belongs in produces no new content; it produces one label out of a fixed set, which is classification, and Amazon Comprehend does custom classification from your own labelled examples at a fraction of the cost and with an accuracy figure you can measure. The priority-contract rule is smaller still: read the contract tier off the account record and branch. That is an if-statement, and a model asked to make the same decision will occasionally get it wrong, which is a strange thing to buy. The fourth option misreads the vocabulary, since diffusion models are one family of foundation model and generating images is squarely on the list. The third option treats recommendation engines as off-limits, when the list names them and the conversational form is exactly where a model beats a ranked grid.
Q. Catalogue images plus a clip and a voiceover, call transcripts reduced to a paragraph, inbound email sorted into eleven fixed queues, and a shopping assistant that talks. Which one is not foundation-model work?
A. The email routing. One label out of eleven is classification, not generation, and the priority-contract rule is a field lookup.
Why? Learn the list: image, video, and audio generation; summarization; chatbots; translation; code generation; customer service agents; search; and recommendation engines. Three of these projects land on it. The catalogue work is image, video, and audio generation. The call paragraph is summarization, which is the safest use of a model because the answer is in the source text rather than in the model’s memory. The storefront assistant touches four entries at once, being a chatbot, a customer service agent when a delivery goes missing, search over the catalogue in a shopper’s own words, and a recommendation engine when it names products. Routing touches none of them. It produces a label, and Amazon Comprehend does custom classification more cheaply and more consistently, with a measurable accuracy you can report. The contract-tier branch is plain code. The pull towards the model is that all four projects involve English text, so they feel like one kind of work; sort by what comes out the far end instead. New content is a foundation model. One category out of a known set is a classifier. One row out of a table is a query.