AWS AI Services Compared for Real-World Use Cases

View the full Getting to Know AWS AI: Services, Use Cases and the Reality of Adoption →

AWS AI services are not interchangeable: Amazon Bedrock is designed for generative applications, SageMaker AI for building and operating machine-learning models, Textract and related APIs for bounded analysis tasks, and Amazon Q for ready-made employee or developer assistance. Choosing the wrong layer can add cost and complexity without improving the result.

That matters because the platform’s AI range can look more interchangeable than it is. Amazon Bedrock gives developers managed access to generative foundation models: broad models trained to produce content such as text, images or code. Amazon SageMaker AI is for building, training, deploying and operating machine-learning models and pipelines. Textract, Transcribe, Comprehend and Rekognition are pre-built APIs (application programming interfaces) for particular document, speech, language and image tasks. Amazon Q is an assistant product for employees and developers, rather than a blank application platform.

The sensible buying question isn’t “which AWS AI service is best?” It is: what decision, piece of work or customer interaction are we trying to improve, and how much control do we genuinely need?

For the wider map of AWS terminology, model types and service families, read Part 1 of this series, AWS AI Explained: The Services, Terms and Choices Behind the Platform. This instalment compares where each option fits, where it doesn’t, and what remains stubbornly human work.

Choose the job first, then the AWS layer

A useful dividing line is whether the system is being asked to recognise, predict, generate or assist.

Recognition means extracting known kinds of signal: text from a scanned form, entities from a complaint, a spoken conversation into a transcript, or labels from an image. These are AWS’s purpose-built AI APIs. Prediction is a model estimating a future or unknown value from historical examples: stock demand next month, likely customer churn, or a manufacturing defect category. That is conventional machine learning, where SageMaker AI is usually the more natural home.

Generation covers language and increasingly images, code and structured outputs: drafting a reply, summarising a case file, rewriting a product description, or answering a question against policy documents. Bedrock is AWS’s principal managed layer for this work. Assistance is the user-facing package: an employee asks a company assistant for information, or a developer asks for help understanding and changing an AWS workload. That is where Amazon Q enters.

There will be overlap. A claims workflow, for example, may use Textract to pull fields from a document, Comprehend to spot named entities in correspondence, Bedrock to draft a case summary, and a conventional SageMaker model to flag cases that deserve early investigation. The workflow is the product. The individual AI call is only one component.

For generative applications, Bedrock is usually the sensible first stop

Amazon Bedrock is managed model access plus the building blocks around it. A developer can call supported foundation models through a common AWS service layer, then add guardrails, prompt management, knowledge retrieval, flows and model customisation. It is aimed at teams that want to build a generative-AI feature without hosting the underlying large language model themselves.

That makes Bedrock a reasonable choice for common applications including:

• Summarisation: condensing support interactions, long reports, call transcripts or technical notes into a useful brief.
• Content assistance: producing a first draft of a response, knowledge-base article, product description or internal update from supplied facts and style rules.
• Structured extraction with explanation: turning messy narrative text into a proposed JSON record, a structured data format, then showing the rationale or source passages to a reviewer.
• Question answering over controlled information: helping staff find the relevant policy, procedure or product information from approved internal sources.
• Agent-style workflows: allowing a model-directed system to select approved tools, retrieve information and propose or carry out constrained actions.

Don’t confuse access to a model with a finished application. A generic chat prompt may demonstrate the possibility in an afternoon. A useful production system needs a proper user interface, authentication, logging, error handling, tests, escalation paths and a clear answer to a basic question: what happens when the model is wrong?

RAG is useful, but it isn’t a truth machine

Retrieval-augmented generation, normally shortened to RAG, gives a model relevant material at the point of a question rather than relying solely on its general training. In a typical setup, an internal assistant searches approved documents, retrieves relevant chunks and uses them as context when forming an answer. Bedrock Knowledge Bases provides managed and customer-managed approaches to this process, including retrieval and source citations in responses.

This is materially better than pasting an entire policy library into a prompt. It can also reduce the temptation for a model to improvise. But RAG quality rests on mundane foundations: whether documents are current, whether content has been split sensibly, whether retrieval finds the right passages, and whether the user is permitted to see them.

A policy assistant that retrieves an obsolete staff handbook with great confidence is still a bad policy assistant. Give users links or citations to source material where an answer has consequences. Test questions with ambiguity, conflicting policy versions and awkward terminology. Keep a human approval point for advice that changes money, rights, safety or compliance.

“Agents” need constrained authority, not a grander name

Agent-style systems are often presented as the next step: software that plans, retrieves information and calls tools to complete multi-stage jobs. There are useful, limited versions of this. An assistant might gather account details, check a product catalogue, create a draft support ticket and hand it to an employee. Or it might retrieve relevant contract clauses and populate a review checklist.

The practical limit is authority. A system that can draft is different from one that can issue a refund, change a record, place an order or alter infrastructure. Start with read-only retrieval and draft outputs. For actions, expose narrow tools with tightly defined inputs, enforce permissions outside the model, and require confirmation for meaningful changes. Bedrock’s service landscape is changing here too: AWS documents its older Bedrock Agents capability as “Classic” and directs new customers towards AgentCore-related capabilities. Don’t make long-lived architecture decisions from a product demo alone.

SageMaker AI is for machine-learning work that earns its complexity

Bedrock can customise some foundation models, but it isn’t a substitute for every machine-learning programme. Amazon SageMaker AI is AWS’s managed environment for the broader lifecycle: preparing data, training models, deploying inference endpoints, which serve model predictions to applications, monitoring quality and running machine-learning operations, often called MLOps.

The strongest reason to use SageMaker AI isn’t a more elaborate AI story. It is that the task depends on patterns in your own data, and success can be measured. Examples include forecasting demand across a retail range, predicting maintenance needs from equipment telemetry, detecting unusual transactions, classifying specialist images, or estimating the likelihood that a customer will require follow-up.

These aren’t generally jobs for a large language model with a clever prompt. They require a representative dataset, a meaningful target outcome, careful handling of missing or biased records, and comparison against a non-ML baseline. If a simple reporting rule or spreadsheet forecast performs nearly as well, the simpler method is often the better operational choice.

SageMaker AI also makes sense where the team needs to own the model-development process: custom features, reproducible training runs, evaluation metrics, approval gates, monitoring and retraining. That ownership matters when a model affects revenue, cost, safety or a regulated process. It is needless overhead for an internal tool that merely needs to summarise a paragraph.

There is a middle ground. AWS offers managed and low-code paths within SageMaker AI, including facilities for custom tabular, time-series and image models. They can lower the barrier to an initial model, but they don’t make data quality, evaluation or accountability disappear. A churn classifier trained on poorly defined customer outcomes will simply produce polished-looking uncertainty at scale.

Use the pre-built APIs when the task is recognisable and bounded

AWS’s older AI services remain some of its most practical. They may not attract the attention generative AI does, but a specific API can be cheaper, faster to evaluate and easier to control than asking a general model to imitate the same task.

Documents: Textract is extraction, not a complete back-office process

Amazon Textract extracts text and analyses document structures such as forms and tables. It fits organisations receiving repeatable document types: invoices, purchase orders, applications, receipts or scanned records. It can remove a good deal of manual keying, especially when the output feeds a clear validation process.

It cannot decide whether an invoice is legitimate, whether a field belongs to the right supplier record, or whether an unusual document should be paid. Build deterministic checks around it: supplier matching, totals validation, duplicate detection and thresholds for review. Put exceptions in front of people who understand the process. Low-confidence output is not a minor technical nuisance.

Speech and language: Transcribe and Comprehend deal in signals

Amazon Transcribe converts speech to text, with custom vocabulary support for domain terms, acronyms and proper nouns. That is valuable for meetings, contact-centre recordings and media workflows, but a bespoke vocabulary isn’t merely a configuration tweak: it needs testing against accents, audio quality and the terminology people actually use.

Amazon Comprehend applies natural-language processing to text, identifying elements such as entities, key phrases, language and sentiment. It can also detect certain personally identifiable information in supported text contexts. It is useful for triage, search enrichment, redaction workflows and basic analysis of large document sets. Treat its confidence scores as routing signals, not declarations of fact. Decide in advance what a score below the threshold does: reject, flag, redact, ask for review or continue with an audit trail.

Images and video: Rekognition can reduce review volume, not eliminate judgement

Amazon Rekognition analyses images and video for labels, text, people, faces, scenes and moderation-related categories, among other functions. It can support asset cataloguing, content moderation queues and searches across sizeable media collections. AWS returns confidence scores for label detection, which is useful precisely because those outputs should not automatically become final decisions.

Be especially cautious with face matching, identity-related work and sensitive judgements. Technical capability does not answer whether a deployment is lawful, proportionate, fair or acceptable to the people affected. The surrounding policy, retention rules, human review and evidence requirements matter more than the API name.

Amazon Q is an assistant product, not Bedrock with a different label

Amazon Q Business is positioned as a generative-AI assistant that organisations can tailor to business information and tasks. Amazon Q Developer is targeted at developers working to understand, build, extend and operate AWS applications and workloads. They use generative AI, but the purchase decision differs from Bedrock.

With Bedrock, a team builds its own application: it chooses the interface, model behaviour, knowledge architecture and connections to other systems. With Amazon Q, AWS provides more of the assistant product. That can suit internal knowledge discovery, writing support or developer help where a standard interaction model is sufficient.

It can be the wrong answer when the requirement is a customer-facing feature, a distinctive workflow, a tightly designed approval journey or a tool that must behave in a highly specific way. In that case, Bedrock may be the better foundation, with the accompanying engineering and operational responsibility.

Amazon Q doesn’t remove the hard work around permissions. A business assistant connected to company data needs identity-aware access and well-managed sources; otherwise it risks becoming a convenient way to surface material to the wrong colleague. For developer assistance, teams still need code review, tests and change controls. A plausible code suggestion is not the same as a safe production change.

The integration bill is normally larger than the model bill

The most misleading AI pitch is that selecting a service is the deployment. In real systems, value depends on integration: the document arriving in the right storage location, the identity system passing the right permissions, the customer record being found reliably, the result appearing in the existing work queue, and an employee being able to correct it without starting again.

This is particularly relevant for smaller organisations in the Channel Islands and elsewhere with limited specialist capacity. A well-bounded workflow that improves one irritating process is usually a better first project than a broad assistant connected to everything. Digital Jersey’s AI Project Lab has moved from training to prototypes, which gets at the practical point: technical experimentation becomes useful only when it meets real data, processes and deployment constraints.

Part 3, Should You Adopt AWS AI? A Practical Test for Cost, Risk and Readiness, examines the questions that should be settled before production: model and infrastructure cost, data handling, region and residency requirements, security permissions, skills, vendor dependence, support arrangements and who owns the result after launch.

For now, the buying guidance is straightforward. Don’t pay for an open-ended generative system when Textract or Transcribe solves the job. Don’t build a custom machine-learning pipeline when a managed assistant is enough. And don’t assume an assistant is a shortcut around data quality, access control or accountable human decisions. AWS offers a broad and capable AI toolkit. It is most useful when treated as one.