Prebuilt AWS AI Services: Vision, Language, Speech and Conversation

Everything you built so far: data pipelines, algorithms, training jobs, endpoints: was yours: your data, your model, your bill for instance-hours. This lesson is about the other half of the AIF-C01 surface: the services where AWS already owns the model and you rent the answer by the unit. That single change in ownership changes every decision you make. You no longer tune hyperparameters; you choose the right API, the right feature set inside that API, and the right meter, and the exam tests exactly that, because "which service?" and "what does it cost?" are the two question shapes that recur across all five domains.

By the end of this lesson you will be able to:

• map modality to service and defend the choice against three plausible distractors;
• state the billing unit of every service in scope and compute a monthly bill from AWS's own worked examples;
• recall the free tiers that differ in duration: 12 months for most, 3 months for Textract, 60 days for Amazon Q Business;
• apply the Rekognition Group 1 / Group 2 metering split and explain why two APIs on one image cost twice;
• pick the correct Textract API and feature combination for OCR, forms, tables, queries, IDs, expense and lending;
• defend the Personalize minimum of 1 provisioned TPS and the Amazon Forecast lifecycle answer;
• separate prebuilt services, SageMaker custom builds and Amazon Bedrock with the comparative verdict;
• answer ten exam-style questions written in the official voice, including four arithmetic items.

---

1. The prebuilt landscape: what each service is for

1.1 Ten services, four modalities

The AIF-C01 exam does not ask you to remember API method names from memory. It asks you to read a workload description and name the service. That is a classification task, and it starts from one input: what is the customer handing you?

1.2 What "prebuilt" means on this exam

A prebuilt service is an API with a fixed output contract. You send a JPEG and you get back labels with confidence scores and bounding boxes; you send 550 characters and you get back entities with offsets. You cannot change the label vocabulary in Rekognition DetectLabels (that is what Rekognition Custom Labels is for), you cannot make Comprehend return a paragraph of reasoning (that is a foundation model task), and you cannot ask Textract what the document means (that is Textract Queries for structure, or Comprehend/Bedrock for semantics).

Three consequences drive the whole exam:

1. Pricing is per unit, not per hour: images, pages, characters, minutes, requests, users. No instance to size, no idle endpoint to forget about (with one loud exception: Personalize campaigns, section 8.1).
2. Feature selection inside one service changes the bill: Textract OCR is $0.0015 per page and Textract Forms is $0.05 per page, a 33× multiplier on the same document.
3. Free tiers are not uniform: 12 months for most services, 3 months for Textract, a 60-day trial for Amazon Q Business, and nothing for Personalize.

• 📚 Did you know? The AIF-C01 exam guide's in-scope AI/ML list contains 14 services: Amazon Augmented AI (A2I), Amazon Bedrock, Amazon Comprehend, Amazon Fraud Detector, Amazon Kendra, Amazon Lex, Amazon Personalize, Amazon Polly, Amazon Q, Amazon Rekognition, Amazon SageMaker, Amazon Textract, Amazon Transcribe and Amazon Translate. Eight of the ten services in this lesson come straight off that list: Amazon Forecast does not (it is a lifecycle question, not a construction question), and Amazon Q was added in a later guide revision along with Aurora, Bedrock AgentCore, Kiro, Strands Agents, SageMaker JumpStart and AWS Transform. Re-check the current in-scope list before exam day; the list has changed before and will change again.

---

2. Amazon Rekognition: images, video, faces and moderation

2.1 Two API groups, one image

Rekognition is the service with the most misunderstood meter in the whole prebuilt set, and the reason is historical. In the 9 November 2021 price reduction AWS split the image APIs into two groups that are metered separately, and that split still governs billing in 2026.

The rule the exam is really testing: image billing is per image, per API call. Run DetectLabels and then DetectModerationLabels on the same 2.5 million images and you have processed 5 million billed images, not 2.5 million. Video behaves the same way: a stored video minute analyzed by both the label API and the content moderation API bills two minutes, one per API.

2.2 The meters

2.3 Hard limits and free tier

2.4 Worked AWS examples

Example 1: 2.5 million moderation images. AWS's own pricing example runs DetectModerationLabels over 2,500,000 images in a month:

$$1{,}000{,}000 \times \$0.0010 + 1{,}500{,}000 \times \$0.0008 = \$1{,}000 + \$1{,}200 = \$2{,}200 \text{ per month}$$

Add a second API on the same images (say DetectText for watermarks) and the bill doubles to $4,400/month, because each API bills its own image count.

Example 2: video API versus frame sampling. For a 60-second video:

The gap widens with length: in an AWS blog analysis of 4,271 minutes of video, the video API cost $427.14 while 2 fps image sampling cost $512.57 for the images plus $164.23 for the Lambda that sampled them: $676.80 total. Below roughly a minute the two are close; above it, the video API wins on price and returns timestamps, shot boundaries and person tracking that frame sampling cannot.

• 📚 Did you know? The face-metadata meter has two official spellings of the same number: the pricing page says $0.00001 per face per month and the FAQ says $0.01 per 1,000 face vectors per month. They are identical ($0.01 ÷ 1,000 = $0.00001). Exam options love to present one as a discount over the other; they are the same price, and both are storage: a collection of 500,000 faces costs $5.00/month whether or not you search it.

---

3. Amazon Comprehend: what the text is saying

3.1 The unit and the minimum

Comprehend meters in 100-character units with a 3-unit (300-character) minimum per request. The standard rate in AWS's worked examples is $0.0001 per unit, which is $1.00 per 1,000,000 characters.

The minimum is where candidates lose easy marks. A 120-character tweet is one unit of real content but bills as 3 units = $0.0003. A 600-character support ticket bills as 6 units = $0.0006. Batch requests are the same: you pay for what you send, rounded up per request.

Worked example: support tickets. 10,000 tickets of 550 characters each: 550 ÷ 100 = 5.5 → 6 units per request → 10,000 × 6 = 60,000 units × $0.0001 = $6.00 per month. Note what that number teaches: at this volume Comprehend is effectively free, which is exactly why the exam pairs it with a feature selection question rather than a cost question.

3.2 What the API surface includes

3.3 Custom Comprehend: where the meter changes

3.4 Limits worth memorizing

• 📚 Did you know? Comprehend's prompt safety classification is the bridge between this lesson and Bedrock: AWS added it so you can screen a user prompt for harmful content before it reaches a foundation model. On the exam, a question about "screening an input before generation" has two valid-looking answers and only one is right: Comprehend classifies and scores the text; Bedrock Guardrails applies policies at generation time. Read carefully which side of the call the question puts you on.

---

4. Amazon Textract: structure from documents

4.1 Five APIs, one feature switch

Textract exposes five APIs: Detect Document Text, Analyze Document, Analyze Expense, Analyze ID and Analyze Lending. Only Analyze Document carries feature types, Forms, Tables, Queries, Custom Queries, Signatures and Layout, and they are combinable in a single call, priced per page per enabled feature.

4.2 The price ladder (us-west-2, first 1M pages per month)

Two facts fall out of that table. First, enabling Forms multiplies the OCR bill by 33×. Second, Forms + Tables + Queries at $0.070 is about 47× OCR-only; which is why the exam loves the "which features does the workload actually need?" question.

4.3 Queries: priced per page, not per answer

Queries let you ask natural-language questions of a document, "What is the account number?", and get back answers with page coordinates. The caps are 15 queries per page synchronously and 30 per page asynchronously, and the critical pricing rule is that you are billed per page regardless of how many answers come back. Asking one question or fifteen questions on the same page costs the same. Asking the same question of 100,000 pages costs 100,000 × $0.015.

4.4 Free tier: three months, not twelve

Page counting is simple and deliberately exam-friendly: one image (PNG, TIFF or JPEG) is one page, and each page of a PDF is one page.

Worked example, feature selection. OCR of 100,000 pages costs $150. Forms + Tables on 5,000 pages costs 5,000 × ($0.05 + $0.015) = $325. The second job touches 20× fewer pages and costs 2.2× more, $0.065 per page versus $0.0015, a 43× unit difference. The lesson is not "avoid Forms"; it is "enable only the features you will read".

---

5. Amazon Translate: high-volume machine translation

5.1 Rates

Characters include spaces, so you cannot game the meter by removing whitespace.

5.2 Free tier

• 2,000,000 characters per month for 12 months for standard and batch translation;
• Active Custom Translation: 500,000 characters per month for 2 months;
• No free tier at all for real-time document translation.

5.3 What you are not charged for

AWS's FAQ draws an unusually precise line: you are not charged when the source language equals the target language, and you are not charged when the call errors out and returns no translation. You are charged for non-translatable content that still produces output. In practice this means a misconfigured loop that "translates" English to English is free, and also that the exam can ask about billing exceptions without any arithmetic.

5.4 Custom terminology

Custom terminology is how you pin brand names, product codes and legal terms. Its limits are all quotable exam material:

Worked example: localizing a manual. 100 pages × 20,000 characters = 2,000,000 characters:

---

6. Amazon Polly and Amazon Transcribe: the speech pair

Polly and Transcribe are the two halves of the same problem, and the exam likes to put them in the same option list to see whether you know which direction the audio flows.

6.1 Amazon Polly: text to speech

The double-billing rule: Speech and Speech Marks are separate billed requests. Requesting both for the same 1,000 characters bills 2,000 characters. Speech marks are what a web player uses to highlight words in sync with audio, so the pattern is common in eLearning, and it doubles the Polly line on the invoice.

Worked example: one million characters:

Polly is billed per character, never per request: an option that says "per synthesis request" is a distractor.

6.2 Amazon Transcribe: speech to text

Transcribe bills in 1-second increments with no minimum, which is unusual and worth memorizing on its own: a 3-second clip costs three seconds, not a rounded minute.

What is included versus what costs extra:

Two structural rules: stereo audio bills the total duration once, not per channel, and Call Analytics returns insight fields, talk time, interruptions, non-talk, participant sentiment, that plain transcription does not.

Worked example: a 200,000-call contact center. 200,000 calls × 10 minutes = 2,000,000 minutes of post-call analytics:

$$250{,}000 \times \$0.0300 + 750{,}000 \times \$0.0186 + 1{,}000{,}000 \times \$0.0138 = \$7{,}500 + \$13{,}950 + \$13{,}800 = \$35{,}250 \text{ per month}$$

Limits and free tier:

• 📚 Did you know? Polly's Standard 5M-character free tier is the one figure on the AWS pricing page that omits the "first 12 months" qualifier used for Neural, Long-Form and Generative, while third-party pages confidently call it "always free". Do not build an answer on that: memorize the page as it reads, and treat any claim of a permanently free Polly tier as unverified. The same discipline applies to Lex's streaming rates, which circulate in third-party calculators but are not on AWS's own pricing excerpt, the AWS-published Lex numbers are $0.004 speech and $0.00075 text.

---

7. Amazon Lex: intents, slots and the per-request meter

7.1 The bill

AWS's worked example: 8,000 speech requests + 2,000 text requests = 8,000 × $0.004 + 2,000 ×$0.00075 = $32.00 + $1.50 = $33.50 per month.

7.2 What counts as one request

• Every user input is one billed request. Five typed messages in a conversation are five requests.
• In a streaming conversation, all turns ride on a single API call: Lex V2 streams over HTTP/2, and the whole exchange is one billable call rather than five.
• Speech input is capped at 15 seconds per request.
• Responses carry NLU confidence scores, sentiment and an interpretationSource field of either Bedrock or Lex: the exam's hint that Lex can be backed by a foundation model for generative meaning while still returning a structured intent.

7.3 Free tier

10,000 text + 5,000 speech requests per month for the first year.

7.4 Where Lex stops

Lex answers "what did the user mean?": intent, slots, session state, fulfillment state. It does not answer open-ended questions over enterprise content, and it does not generate prose. That boundary is the single most common distractor set in this lesson: Q Business for permission-aware answers over company data, Bedrock for open-ended generation, Lex for structured intent routing into a Lambda.

---

8. Amazon Personalize and Amazon Forecast

8.1 Personalize: the meter with a floor

The trap: minimum provisioned throughput. By default every active campaign must keep at least 1 provisioned transactions per second (TPS), and AWS states plainly that the minimum applies "even if you make no requests"; you are billed on the greater of the minimum and the actual TPS.

The arithmetic is brutal and AWS's own Knowledge Center publishes it: 1 TPS held for one day is

$$1 \times 24 \times 60 \times 60 = 86{,}400 \text{ billed recommendations per day}$$

which at $0.0556 per 1,000 works out to roughly $144 per month for a campaign nobody is calling (about 2.59 million recommendations a month). The operational answer is short: delete idle campaigns.

8.2 Amazon Forecast: a lifecycle question, not a construction question

---

9. Amazon Q: the per-user answer

Amazon Q is explicitly in scope for AIF-C01, and it is the only service in this lesson priced per user per month rather than per unit of work. That difference is the whole question.

Three readings the exam tests:

1. Q Business is permission-aware enterprise RAG. It answers questions over your connectors (S3, Confluence, Salesforce and dozens more) with each user's own permissions respected, and it cites sources. That is a different product from "an anonymous public chatbot".
2. The trial is 60 days, not 12 months: the shortest free period in this lesson, and the one candidates most often confuse with the 12-month AWS free tier.
3. Q Developer's Pro overage is per line of code, at $0.003/LOC over the pooled 4,000 LOC/month: a meter nobody expects, and therefore a favorite for a numerical option.

---

10. Which AI service? The decision tree

Everything above collapses into one diagram. Read the input first, then the intent, then the meter; that order is exactly how the exam writes its options.

Practice the mapping until it is automatic:

• 📚 Did you know? Amazon Lex V2 responses carry an interpretationSource field that can read Bedrock instead of Lex: the bot's NLU can be powered by a foundation model while the response stays a structured intent with slots and confidence. That is why "use Bedrock instead of Lex" is a partially wrong distractor: Bedrock gives you generation and reasoning, Lex gives you the conversation state machine, session, fulfillment hooks and the per-request meter the contact center is budgeted on.

---

11. Free tiers, limits and the numbers you must memorize

11.1 Free tier: memorize the outliers

The three durations are the exam's gift: 12 months for almost everything, 3 months for Textract, 60 days for Q Business.

11.2 Hard limits worth memorizing

11.3 AWS worked examples, all in one table

11.4 Input in, output out

• 📚 Did you know? Two services in this lesson do not charge you when nothing happens: Translate does not bill a call whose source language equals its target language or that errors without returning a translation; while Comprehend's 300-character minimum means a 40-character request costs the same as a 300-character one. The exam loves pairing a "no charge" rule with a "minimum charge" rule in the same option list to see whether you are recalling a specific service or just the general vibe of per-unit pricing.

---

12. Comparative verdict

---

13. Exam traps for this lesson

---

Service lifecycle watch

AWS retires and renames AI services on a published schedule, and part of this exam is telling a maintenance announcement apart from a shutdown. The vocabulary comes from the AWS General Reference (Service Lifecycle, last updated 24 Sep 2026) and is the same vocabulary section 8.2 already used for Amazon Forecast:

• Maintenance: closed to new customers and to new features, but still supported for existing customers;
• Sunset: a planned end of operations, typically announced about 12 months ahead;
• Full Shutdown: the service is removed from the AWS portfolio entirely.

2025 to 2026 Updates

Every row below is verified against an AWS availability-change page, an AWS What's New post, an AWS blog or the AIF-C01 exam guide. Nothing here comes from a third-party tracker.

The exam itself moved at the same time. AWS published exam guide v1.0 on 26 March 2026 and v1.1 on 30 April 2026, and states that guide changes appear on the live exam about one month after publication:

Two of this lesson's answers therefore gained a 2026 sibling: Forecast → SageMaker Canvas (closed 29 Jul 2024) and now Kendra → Bedrock Managed Knowledge Base (closed 30 Jul 2026). Both are lifecycle questions, not construction questions: the exam wants the pair (new customers → replacement, existing customers → keep running, no new features).

• 📚 Did you know? The exam logistics did not move while the content did: AIF-C01 is still 65 questions (50 scored + 15 unscored) in 90 minutes for USD 100, pass mark 700/1000, valid for 3 years, but the Italian and German language versions were retired after 15 October 2026. AWS also says guide updates reach the live exam roughly a month after publication, so v1.1's new objectives: agentic AI with MCP, context engineering, token-based pricing, prompt versioning and hallucination grounding: were examinable from around late May 2026.

---

Real-World Case Studies

The pricing tables tell you what a service costs; AWS's published customer stories tell you when to reach for it. Every figure below is quoted from the AWS case study or AWS Machine Learning Blog post named in the Source column. All of them are customer- or AWS-claimed and unaudited, and where a source says "up to", treat it as a ceiling rather than an expectation.

Case 1, Anthem: prebuilt Textract first, humans for the tail

Anthem's claim-form extraction reportedly cost millions of dollars a year and took 20 minutes per claim when handled manually. The AWS-published architecture routes portal documents through Amazon Textract OCR plus an ML index/classify step. The reported outcome: 80% of the workflow automated, with a 90%+ target, across thousands of claims per day. The exam-relevant part is what happens to the rest: a regulated, structured extraction task with a matching prebuilt API gets the prebuilt service first, and the remaining 10 to 20% of hard cases goes to a human exception path; which is exactly what Amazon A2I exists to manage. Source: aws.amazon.com/solutions/case-studies/anthem (re:Invent 2020 session).

Case 2, Chronomics: four months of DIY vision → 3 to 4 weeks of AutoML

Chronomics spent 4 months building an in-house computer-vision model to read COVID-19 test results and never reached its target. Switching to Amazon Rekognition Custom Labels shipped in 3 to 4 weeks at 96.5% accuracy / 97.9% F1, scored with DetectCustomLabels. The threshold arithmetic is the exam-relevant detail: at a threshold of 0.99, accepted predictions are 99.6% correct but 5% of predictions are discarded; at 0.999, precision rises to 99.87% while 27% are discarded. Somebody must decide what happens to the discarded tail, the human-review/A2I answer, before the threshold is set. Source: AWS Machine Learning Blog, 13 Dec 2022, aws.amazon.com/blogs/machine-learning/chronomics-detects-covid-19-test-results-with-amazon-rekognition-custom-labels.

Case 3, Sun Finance: the LLM-only OCR trap, and the fix

Sun Finance (fintech lending across 9 countries) had 60% of microloan applications needing manual review, each taking anywhere from 10 minutes to 20 hours. Attempt 1: sending ID photos straight to Claude Sonnet 4 for JSON extraction: scored 61.8% overall and only 43% on ID numbers and was rejected, because privacy protections block direct PII extraction. Attempt 2, built with the AWS Generative AI Innovation Center, separates the stages: Textract for OCR → Rekognition as fallback and face checks → the foundation model only for structuring → validation rules, with Titan Multimodal Embeddings in S3 Vectors for fraud similarity, evaluated on 585 images. Results: overall accuracy 79.73% → 90.80% (ID number 74.32% → 89.40%, document type 78.43% → 96.40%), −91% cost per document, 20 hours → under 5 seconds, fraud detection at 81%. Exam angle: OCR is not reasoning: the deterministic prebuilt API reads the characters, and the model only words the answer. Source: AWS Machine Learning Blog, 30 Apr 2026, aws.amazon.com/blogs/machine-learning/sun-finance-automates-id-extraction-and-fraud-detection-with-generative-ai-on-aws.

Case 4, Alnylam: Amazon Q Business with citations under GxP

Alnylam Pharmaceuticals' complaint triage took 3 to 4 days, and finding one internal answer took 15+ minutes. Using Amazon Bedrock, Amazon S3 and Amazon Q Business, the team built an intake-and-triage prototype in 3 months under GxP constraints, plus AskALNY, a Slack assistant serving 2,000 employees and 1,000 contractors that returns answers with source links. Reported outcomes: triage fell from 3 to 4 days to hours, information search from 15 minutes to 30 seconds, and 250+ use cases followed the first two. Exam angle: "permission-aware answers over company data, with citations" is Amazon Q Business at $3 Lite / $20 Pro per user per month: not Lex (intent routing into a Lambda) and not a bare Bedrock call (no connector or permission layer). Source: aws.amazon.com/solutions/case-studies/alnylam-case-study.

• 📚 Did you know? The only production-readiness rate AWS publishes is 65%: of more than 1,000 Generative AI Innovation Center implementations, 65% reached production in 2025, some in as little as 45 days, using AWS's Five V's framework (Value → Visualize → Validate → Verify → Venture). Never memorize "every pilot succeeds": the other 35% is the number that teaches you to baseline the metric before you build, and to plan a human path for what the model gets wrong.

---

Practice Questions

---