OCI AI Services: Vision, Speech, Language, and Document

The pretrained AI services occupy a specific slot in the OCI stack: above the infrastructure where you train your own models, below the generative platforms where language models reason, and beside the workflows of ordinary applications. Their economic logic is sharp. A model that took millions to train is rented by the API call, and the buyer's entire engineering problem reduces to integration. When the task fits, nothing beats this tier on time to value, and when the task does not fit, no amount of integration effort will close the gap. The skill is knowing which is which before the pilot, not after.

This article is part of our complete guide to AI on OCI. The generative layer has its own article in the OCI Generative AI service; here we cover the perception and understanding services that predate it and quietly outperform it on cost for narrow tasks.

The catalogue, service by service

Vision

Vision analyzes images: object detection, image classification, text extraction from photographs, and face detection. The pretrained models handle general scenes well, and the service accepts custom training on your own labeled images when the subject matter is domain specific, defect detection on your production line rather than objects in general. Typical buyers are operations teams automating visual inspection and content teams tagging media libraries.

Document Understanding

Document Understanding is Vision's specialized sibling for paperwork: it extracts text, tables, and key value pairs from invoices, receipts, passports, and forms, returning structured data rather than raw strings. This is the service with the most reliable enterprise payback, because almost every organization pays humans to retype documents into systems, and the arithmetic of replacing that work is rarely subtle. It pairs naturally with generative models downstream: extraction first, reasoning second.

Speech

Speech converts audio to text, batch transcription of recorded calls and meetings, with timestamps and confidence scores, plus text to speech in the other direction. Contact centers are the archetypal consumer: transcribe everything, then run analytics or generative summarization over the text. Accuracy depends on audio quality and vocabulary, and a pilot on your actual recordings, not the sample files, is the only number that matters.

Language

Language handles text understanding: sentiment, named entity extraction, key phrase detection, language detection, classification, and translation. Custom classification and entity models can be trained on your labels. Its sweet spot is high volume, low ceremony text processing, routing tickets, screening feedback, normalizing multilingual content, where a per record price in fractions of a cent makes the build versus buy question rhetorical.

Anomaly Detection

Anomaly Detection learns normal behavior from your time series, sensor readings, transaction volumes, infrastructure metrics, and flags departures. It is the most underrated service in the set for operations teams, because it converts the impossible task of writing thresholds for everything into the tractable one of reviewing flagged windows.

What they share

The services have a common operational personality, and it is the reason they are easy to govern. They are stateless APIs under standard IAM policy, so access control is ordinary OCI policy work. They price per transaction, per image, per page, per audio hour, per text record, so cost scales with usage and is trivially attributable. They run regionally, so data residency follows the region choice. And your data is not used to train shared models. The integration pattern is always the same: events or batches flow from Object Storage or an application queue, the service annotates, and structured results land back in the data platform.

Pretrained, custom trained, or build your own

The escalation ladder has three rungs, and each costs roughly ten times the one below. Pretrained models, used as shipped, win whenever your task matches the general case: reading typical invoices, transcribing clear calls, classifying ordinary sentiment. Custom training within the services, your labels on their architecture, wins when the domain is specific but the task shape is standard: your defect types, your document keys, your ticket categories. Building your own model, on the OCI Data Science platform or raw GPU capacity, is justified only when the task shape itself is nonstandard or the model is the product. The expensive failure mode is starting on rung three for ego reasons; the cheap failure mode is staying on rung one after the evidence says your domain is too far from general. Both are avoidable with a measured pilot.

A selection and pilot framework

The integration architecture that ages well

The services are easy to call and easier to call badly. The pattern that ages well is asynchronous and event driven: documents, images, or audio land in Object Storage, an event triggers the processing flow, the service annotates, and results land in a queue or table that downstream systems consume at their own pace. This shape absorbs volume spikes without drama, retries failures without human attention, and keeps the AI call out of any user facing request path where its latency would be felt. The synchronous pattern, application calls service and waits, belongs only where a human is actively waiting on the result. Two further habits pay for themselves within a quarter. Wrap each service behind a thin internal interface owned by your team, so a pricing change, a regional gap, or a better model is an implementation swap rather than an application rewrite. And store the raw service responses, confidence scores included, not just your interpretation of them, because the day you want to reprocess history with better thresholds, the stored responses are the difference between a query and a reprocessing bill.

Watching quality after launch

Pretrained services drift in a way infrastructure does not: your inputs change while the model stands still. The documents gain a new layout, the call audio moves to a new telephony stack, the product names in tickets evolve, and accuracy erodes with no error message anywhere. The defense is the same evaluation discipline that chose the service in the first place, kept running. Hold out a labeled sample, score the service against it monthly, and chart the trend beside the exception queue rate, because a rising review queue is usually the first symptom the business notices. Confidence score distributions are the cheap early warning: when the average confidence on a document type sags quarter over quarter, the inputs have drifted even if accuracy has not yet visibly broken. These reviews take an afternoon a month and belong on the same operational calendar as patching and capacity reviews, which is where we put them for estates under Managed Monthly retainers.

Where these services fit in a generative world

The generative wave did not retire this tier; it gave it a new job. The emerging enterprise pattern is a pipeline: perception services convert the physical world, documents, audio, images, into clean structured text, and generative models reason over the result. Document Understanding extracts the contract fields, the language model summarizes the risk. Speech transcribes the call, the language model drafts the follow up. Using a flagship generative model to do raw OCR or transcription is paying reasoning prices for perception work, and the bill notices even when the architecture diagram does not.

The independent view

The AI services tier is where we send clients who want AI value this quarter without standing up an AI program: the wins are real, the governance is ordinary OCI policy work, and the costs are legible. The decisions that benefit from independent eyes are the pilot design, the escalation calls, and the exception economics, which is assessment and architecture work, not procurement. Our OCI consulting practice runs that engagement on a fixed Project fee, and for estates already running these pipelines we fold the metering and quality reviews into the Managed Monthly retainer.