← Agent Frameworks 🕐 15 min read
Agent Frameworks

Bedrock Data Automation Cost Tracking — Document, Audio, and Video Processing Economics (2026)

1. What Bedrock Data Automation Is · March 2026

Amazon Bedrock Data Automation (BDA) reached general availability on March 3, 2025.

Amazon Bedrock Data Automation (BDA) reached general availability on March 3, 2025. It is Amazon’s managed pipeline for extracting structured data from unstructured multimodal content — documents, images, audio, and video — using AI without requiring the customer to stitch together Textract + Comprehend + Rekognition + a foundation model invocation themselves. BDA was previously marketed under the “Intelligent Document Processing” umbrella across several AWS services; BDA consolidates that surface into a single API.

The core abstraction is a Project (a container for extraction configuration) paired with one or more Blueprints (JSON schemas that define which fields to extract and how). A project accepts any combination of the four supported modalities. Invocation is async-only via InvokeDataAutomationAsync; results land in an S3 output bucket, and completion events are delivered via EventBridge.

Regional availability as of mid-2026: US East (N. Virginia), US West (Oregon), EU (Frankfurt, Ireland, London), AP (Mumbai, Sydney), and GovCloud (US-West). BDA is not yet available in all Bedrock regions.

Supported modalities and inputs:

  • Documents: PDF, DOCX, XLSX, HTML, CSV, TXT; multi-page supported
  • Images: JPEG, PNG, TIFF, BMP, WEBP
  • Audio: MP3, WAV, FLAC, OGG, AMR; 11 languages including English, Spanish, French, German, Portuguese, Italian, Japanese, Korean, Chinese, Cantonese, Taiwanese
  • Video: MP4, MOV, MKV, AVI, FLV, WEBM

2. Pricing Structure

BDA charges per unit of consumed content, not per inference token. This is the key structural difference from raw Bedrock model invocation: costs are deterministic and predictable at the document/minute level, not at the token level.

Document and Image Pricing

Mode Unit Price (us-east-1)
Standard Output — Document per page $0.010
Standard Output — Image per image included in Standard (Knowledge Base integration)
Custom Output — Document (≤30 fields) per page $0.040
Custom Output — Document (>30 fields) per page $0.040 + $0.0005 × (fields − 30)
Custom Output — Image (≤30 fields) per image $0.005
Custom Output — Image (>30 fields) per image $0.005 + $0.0005 × (fields − 30)

Standard Output activates BDA’s default extraction: full text, layout, tables, key-value pairs, entities, and a semantic summary. This is sufficient for RAG ingestion into Bedrock Knowledge Bases. Custom Output applies a Blueprint and returns only the fields defined in it — structured JSON keyed to your schema. Custom Output costs 4× Standard for documents but is far more token-efficient downstream because consumers receive clean JSON rather than unstructured text.

The >30 field surcharge is $0.0005 per additional field per processed unit. A 50-field blueprint on a document page costs $0.040 + (20 × $0.0005) = $0.050/page. A 100-field blueprint on the same page costs $0.040 + (70 × $0.0005) = $0.075/page. Blueprint complexity has a direct cost multiplier — field count governance matters at scale.

Audio Pricing

Mode Unit Price
Standard Output — Audio per minute $0.006
Custom Output — Audio per minute Blueprint-driven; same $0.0005/field surcharge applies above 30 fields

Standard audio output includes full transcript, speaker diarization (channel-separated), sentiment per segment, topic classification, and entity extraction. At $0.006/min, a one-hour recording costs $0.36. A call center running 10,000 hours of calls/month spends ~$3,600/month on BDA audio before any downstream processing.

Video Pricing

Mode Unit Price
Standard Output — Video per minute $0.050
Custom Output — Video per minute Blueprint-driven; field surcharge applies above 30 fields

Video is the most expensive modality at $0.050/minute ($3.00/hour). Standard output includes frame-level scene segmentation, visual entity detection (objects, logos — BDA detects 35,000+ company logos), transcript from embedded audio, and a semantic summary. A one-hour meeting recording costs $3.00 in BDA video; 1,000 hours of archived video content costs $3,000.

Important: BDA does not publish separate Custom Output pricing for audio and video on the main pricing page; the field surcharge mechanism applies across all modalities, but base Custom Output rates for audio/video require confirmation against the current AWS Price List API for your region.

No Charge for Blueprint Creation or Storage

Blueprint creation, modification, and versioning carry no direct charge. Blueprint storage in the BDA control plane is free. The pricing effect of blueprints is indirect: using a Custom Output Blueprint triggers the $0.040/page (vs. $0.010/page) rate and field-count surcharges. Blueprint proliferation is a governance concern, not a direct cost driver.

3. Document Processing Cost Model

Per-Page Economics

Scenario Cost/page 10K pages/day Monthly (30d)
BDA Standard Output $0.010 $100 $3,000
BDA Custom Output (30 fields) $0.040 $400 $12,000
BDA Custom Output (50 fields) $0.050 $500 $15,000
Textract OCR only $0.0015 $15 $450
Textract AnalyzeExpense $0.010 $100 $3,000
Textract Forms + Tables $0.065 $650 $19,500
DIY: Textract OCR + Claude Sonnet 4.6 ~$0.012–0.025 ~$120–$250 ~$3,600–$7,500

The “DIY” estimate for Textract + Claude extraction assumes ~800–1,200 input tokens and ~200–400 output tokens per page at Claude Sonnet 4.6 pricing ($3/MTok input, $15/MTok output). The actual DIY cost depends heavily on prompt length and how many extraction passes are required.

BDA Custom Output vs. DIY crossover: For variable-format documents (contracts, medical records, regulatory filings), BDA Custom Output at $0.040/page is competitive with DIY pipelines once you account for orchestration overhead, retry logic, and error handling. For standardized high-volume documents (W-2s, invoices with fixed layouts), Textract’s purpose-built APIs remain 50–75% cheaper than BDA.

Hybrid Routing Model — Published Results

A hybrid approach routing standardized documents to Textract and variable documents to BDA has been documented at 54% cost reduction vs. all-BDA for a 100,000-document/month workload:

  • All-BDA: $4,000/month
  • Hybrid: $1,825/month

The classification call that decides the route (a Bedrock model invocation for document type detection) costs fractions of a cent — well under $50/month even at 100K documents. The hybrid approach becomes worthwhile above roughly 10,000 pages/month where the routing logic amortizes.

When BDA Wins Over DIY

  1. Mixed modality pipelines — a single BDA project handles PDF, image, audio, and video under one API; DIY requires four separate service integrations.
  2. Variable document schemas — BDA’s AI-powered extraction generalizes across document layouts without retraining; Textract Forms requires consistent field positions.
  3. Human-in-the-loop integration — BDA provides confidence scores per extracted field; low-confidence items route to Amazon A2I without custom orchestration.
  4. Speed to production — a BDA pipeline from S3 → structured JSON can be operational in hours vs. weeks for a custom extraction pipeline.
  5. Reduced token costs downstream — Custom Output returns clean JSON; RAG ingestion or downstream LLM calls consume far fewer tokens than processing raw extracted text.

When DIY Wins

  1. High-volume standardized documents (50K+/month) with known layouts — Textract AnalyzeExpense at $0.010/page matches BDA Standard Output but with higher accuracy on structured forms.
  2. OCR-only workloads where entity extraction is not needed — Textract OCR at $0.0015/page is 6.7× cheaper than BDA Standard.
  3. Custom model fine-tuning — organizations with labeled training data can fine-tune extraction models that outperform BDA’s generalist approach at lower per-unit cost.
  4. Existing Textract investments — teams with built pipelines, error handling, and volume discounts on Textract may not recoup BDA migration costs.

4. Audio and Video Processing Cost Model

Audio: BDA vs. Transcribe + Comprehend Pipeline

Component Service Rate 1,000 hours/month
BDA Standard Audio BDA $0.006/min $360
Transcribe (standard) Amazon Transcribe $0.024/min $1,440
Transcribe + Comprehend sentiment Transcribe + Comprehend $0.024 + $0.0001/unit ~$1,500+
Transcribe + Bedrock analysis Transcribe + Bedrock $0.024/min + ~$0.005–0.02/min ~$1,740–$2,640

BDA Standard Audio at $0.006/min is 4× cheaper than Amazon Transcribe alone for equivalent transcript + NLP output. The price difference is substantial: for 1,000 hours of call center recordings, BDA costs $360 vs. $1,440 for Transcribe alone (before any NLP analysis). This makes BDA Audio the economically dominant choice for call center analytics unless custom acoustic models are required.

BDA Audio Standard Output includes: full transcript with timestamps, speaker labels (channel-separated diarization), sentiment per speaker turn, topic detection, named entity extraction, and an AI-generated summary. Getting equivalent output from Transcribe + Comprehend + a summarization call would require three separate service invocations.

Audio Custom Output use cases:

  • Call center: extract resolution_status, product_mentioned, escalation_flag, agent_id, customer_sentiment_shift
  • Medical dictation: extract diagnosis_codes, prescribed_medications, follow_up_date
  • Meeting transcription: extract action_items, decisions_made, owners, deadlines

Video: BDA vs. Transcribe + Rekognition Pipeline

Component Service Rate 100 hours/month
BDA Standard Video BDA $0.050/min $300
Rekognition Video (label detection) Rekognition $0.10/min $600
Rekognition + Transcribe Rekognition + Transcribe $0.10 + $0.024/min $744
Full DIY (Rekognition + Transcribe + Bedrock) Multiple ~$0.15–0.25/min $900–$1,500

BDA Video at $0.050/min is 2–5× cheaper than equivalent DIY pipelines for content moderation or semantic analysis use cases. The economics improve further when blueprint-driven extraction eliminates downstream LLM calls.

Video Standard Output includes: per-scene transcript, visual entity labels, logo detection (35,000+ company logos), content moderation signals, and an AI-generated scene-by-scene summary.

Key video use cases where BDA pays off:

  • Earnings call archives: extract speaker statements, financial metrics mentioned, sentiment shifts by speaker
  • Training video libraries: extract key concepts, timestamps of concept introduction, quiz-generation structured output
  • Content moderation pipelines: BDA’s integrated moderation signals replace a separate Rekognition Moderation invocation

5. Blueprints and Custom Schemas

Blueprint Architecture

A Blueprint is a JSON schema defining extraction fields. Each field specifies:

  • name — output key in the JSON result
  • type — string, number, boolean, or array
  • inferenceType — extracted (directly present in source) or inferred (derived from context)
  • description — natural language instruction to the extraction model

Blueprints can be created via console (Blueprint Prompt assistant accepts natural language), JSON editor, or API (CreateBlueprint).

Project limits:

  • Up to 40 document blueprints per project
  • 1 image blueprint per project
  • 1 audio blueprint per project
  • 1 video blueprint per project

Versioning: CreateBlueprintVersion creates an immutable snapshot. Published versions are named {BlueprintName}_1, {BlueprintName}_2, etc. Invocation requests reference a specific version ARN, enabling production pinning. Drafts remain editable; published versions are read-only. Versions can be duplicated as new blueprint bases.

Blueprint and Token Efficiency

Custom Output blueprints are more token-efficient for downstream LLM consumption than Standard Output. Standard Output returns full extracted text, tables, and metadata — often 2,000–10,000 tokens for a typical business document. A Custom Output blueprint returning 20 structured fields yields a compact JSON object of 200–500 tokens.

For a pipeline where extracted content feeds a downstream LLM (classification, routing, summarization), blueprint-structured output can reduce input token costs by 80–95%. At Claude Sonnet 4.6 pricing ($3/MTok input), eliminating 2,000 tokens per document at 10K documents/day saves approximately $600/month in downstream inference alone — partially or fully offsetting the BDA Custom Output premium over Standard.

Blueprint Instruction Optimization (Dec 2025)

AWS added instruction optimization for document blueprints in December 2025. The feature accepts a small set of labeled example documents and automatically refines field extraction instructions to improve accuracy. This reduces the manual prompt-engineering cycle for blueprint development without additional runtime cost.

Blueprint Governance Recommendations

  1. Pin versions in production. Always reference a specific version ARN in InvokeDataAutomationAsync; using the draft blueprint ARN means production behavior changes when developers modify the schema.
  2. Count fields before deploying. The 30-field threshold is the primary cost inflection point. Audit blueprints for unused or redundant fields before deployment.
  3. Separate concerns across projects. Use distinct projects for document types with materially different extraction schemas; avoid stuffing 40 document blueprints into a single project when sub-5 blueprints with tight schemas would suffice.
  4. Test field surcharge math. At 10K pages/day, adding 10 fields beyond 30 adds $50/day ($1,500/month). The cost of additional fields compounds with volume.

6. Enterprise Document Pipeline Integration Patterns

Standard Architecture: S3 → BDA → Downstream

S3 Input Bucket
    │
    ├── S3 Event Notification → Lambda (trigger)
    │                               │
    │                               └── InvokeDataAutomationAsync
    │                                       │ (projectArn, inputS3Uri, outputS3Uri)
    │                                       │ (notificationConfiguration: EventBridge)
    │
    EventBridge Rule (BDA completion event)
        │
        └── Lambda (result processor)
                │
                ├── Read structured JSON from S3 output bucket
                ├── Write to DynamoDB / RDS / OpenSearch
                └── Trigger downstream workflow

The async invocation accepts:

  • inputConfiguration.s3Uri — individual file or prefix
  • outputConfiguration.s3Uri — output bucket prefix
  • blueprintConfiguration — list of blueprint ARNs to apply (BDA auto-classifies and routes)
  • notificationConfiguration.eventBridgeConfiguration.eventBridgeEnabled: true

Status polling alternative: GetDataAutomationStatus returns IN_PROGRESS, SUCCESS, or FAILURE. EventBridge notification is preferred over polling at scale.

Step Functions Orchestration Pattern

For pipelines requiring human review or multi-stage processing, AWS-published reference architectures use Step Functions:

  1. Map state fans out individual document invocations in parallel
  2. Wait-for-callback task token pauses for low-confidence extractions
  3. A2I human review loop resumes the callback token on reviewer completion
  4. Results merge into a final structured record

This pattern adds Step Functions costs (~$0.025 per 1,000 state transitions) — negligible vs. BDA costs at document scale.

Error Handling and Retry Costs

BDA does not retry failed invocations automatically. Failed jobs (FAILURE status) do not incur a BDA charge — billing is on successful completion only. Retry logic in Lambda or Step Functions re-invokes InvokeDataAutomationAsync, which will incur a full BDA charge on success. Design retry backoff carefully for large documents where transient failures are more likely.

Partial extraction: BDA does not bill partial pages. A 10-page PDF that fails on page 7 is not billed for pages 1–6. The invocation fails and no charge applies; a retry processes the full document again.

BDA vs. Bedrock Flows for Document Pipelines

Bedrock Flows is a visual pipeline builder for LLM workflows. BDA is a separate service focused on multimodal extraction. The two are complementary: BDA handles extraction, Bedrock Flows can orchestrate downstream processing (classification, routing, enrichment) using the structured JSON BDA produces. Bedrock Flows charges for model invocations within the flow at standard inference rates — BDA costs are additive, not substituted.

For pure document extraction without downstream LLM steps, BDA alone is sufficient. For enrichment pipelines (extract → classify → enrich → store), BDA + Bedrock Flows is a natural pairing.

7. Cost Allocation for BDA in CUR

How BDA Appears in CUR 2.0

BDA charges do not appear as token-based line items. They appear with usage type patterns reflecting the modality and output type:

Usage Type Pattern Modality Mode
{region}-BDA-Document-StandardOutput-Pages Document Standard
{region}-BDA-Document-CustomOutput-Pages Document Custom
{region}-BDA-Image-CustomOutput-Units Image Custom
{region}-BDA-Audio-StandardOutput-Minutes Audio Standard
{region}-BDA-Video-StandardOutput-Minutes Video Standard

The exact usage type strings should be confirmed against the AWS Price List API (GetProducts for AmazonBedrock service). AWS CUR documentation notes that BDA line items are distinct from bedrock-runtime model inference line items and will not aggregate with token-based spend.

Critical reconciliation note: BDA costs will not appear in standard Bedrock cost summaries that filter on token types (input/output/cache). Filter on product_servicecode = 'AmazonBedrock' AND line_item_usage_type LIKE '%BDA%' to isolate BDA spend from model inference spend in the same account.

Tagging Strategy for BDA Cost Attribution

BDA jobs inherit resource tags from the IAM role executing InvokeDataAutomationAsync. As of April 2026, Bedrock supports IAM principal-based cost allocation in CUR 2.0 — tags on the executing role propagate to CUR as iamPrincipal/{key} columns.

Recommended tagging schema for enterprise BDA pipelines:

IAM Role Tags:
  team: finance-ops
  env: production
  workload: invoice-extraction
  cost-center: CC-4412
  document-type: contract

These tags appear in CUR 2.0 after activation in the AWS Billing console (24-hour propagation lag).

Athena Query: Isolate BDA Spend by Document Type

SELECT
  resource_tags_team AS team,
  resource_tags_workload AS workload,
  line_item_usage_type,
  SUM(line_item_unblended_cost) AS total_cost,
  SUM(line_item_usage_amount) AS total_units
FROM
  your_cur_database.your_cur_table
WHERE
  product_service_code = 'AmazonBedrock'
  AND line_item_usage_type LIKE '%BDA%'
  AND line_item_usage_start_date >= DATE '2026-01-01'
GROUP BY
  resource_tags_team,
  resource_tags_workload,
  line_item_usage_type
ORDER BY
  total_cost DESC;

Bedrock Projects for BDA Cost Grouping

Bedrock Projects (the cost-management construct, separate from BDA Projects) allow tagging at the application level for inference API costs. At time of writing, Bedrock Projects primarily tag inference calls via the Responses API and Chat Completions API endpoints. BDA invocations via bedrock-data-automation-runtime may not flow through Bedrock Projects tagging — confirm against current CUR output for your account before relying on this mechanism for BDA attribution.

Cost Anomaly Detection Setup

For BDA pipelines processing variable document volumes, configure AWS Cost Anomaly Detection with a monitor on AmazonBedrock filtered to BDA usage types. Set alert thresholds at 20–30% above expected daily spend. Common anomaly causes:

  • Runaway S3 event triggers processing the same files repeatedly
  • Retry loops on persistently failing documents
  • Blueprint field count increase after a schema update (crossing the 30-field threshold)

8. BDA vs. Alternatives — Comparison

Document Processing Cost Comparison (10K pages/day scale)

Service Use Case Price/page 10K pages/day Strengths Weaknesses
BDA Standard Mixed unstructured docs $0.010 $100/day Single API, multimodal, managed No volume discount, 4× premium for custom
BDA Custom Output Structured extraction $0.040 $400/day Clean JSON output, blueprint governance Expensive vs. purpose-built APIs
AWS Textract OCR Text extraction only $0.0015 $15/day Cheapest, fast, volume discounts No semantic extraction
AWS Textract AnalyzeExpense Invoices, receipts $0.010 $100/day Purpose-built accuracy Fixed document types only
AWS Textract Forms+Tables Structured forms $0.065 $650/day High accuracy on known layouts Expensive, layout-dependent
Azure Document Intelligence (custom) Custom extraction ~$0.010–0.030 $100–$300/day Strong form accuracy, commitment discounts Azure-only, less flexible on variable layouts
Google Document AI (Form Parser) Forms, invoices ~$0.065 $650/day High accuracy Expensive for general extraction
Google Document AI (custom extractor) Custom schemas ~$0.100 $1,000/day Flexible Most expensive at scale
Reducto Complex doc parsing ~$0.015–0.030/pg $150–$300/day Excellent accuracy on complex layouts, agentic Not AWS-native, data residency questions
Unstructured.io RAG preprocessing ~$0.030/pg $300/day Best-in-class RAG chunking Higher cost, primarily preprocessing
DIY (Textract + Claude Sonnet 4.6) Flexible extraction ~$0.012–0.025 $120–$250/day Controllable, tunable Engineering overhead, token cost variability
Mistral OCR / open-source OCR only ~$0.001–0.002 $10–$20/day Cheapest No extraction, infrastructure required

Audio Processing Comparison (1,000 hours/month)

Service Cost Output
BDA Standard Audio $360/mo ($0.006/min) Transcript + diarization + sentiment + entities + summary
Amazon Transcribe $1,440/mo ($0.024/min) Transcript + diarization only
Transcribe + Comprehend ~$1,500/mo Transcript + NLP (separate API calls)
Transcribe + Bedrock (Claude) $1,740–$2,640/mo Transcript + LLM analysis
Azure Speech Services ~$720/mo ($0.012/min) Transcript + basic sentiment
Google Speech-to-Text ~$360/mo ($0.006/min) Transcript only (no NLP)

BDA matches Google Speech-to-Text on price while delivering substantially more structured output. It is the cost-dominant choice for call analytics workloads that need NLP output, not just transcription.

Video Processing Comparison (100 hours/month)

Service Cost Output
BDA Standard Video $300/mo ($0.050/min) Transcript + visual entities + logo detection + moderation + summary
AWS Rekognition Video $600/mo ($0.10/min) Visual labels, faces, moderation
Rekognition + Transcribe $744/mo Visual + transcript
DIY (Rekognition + Transcribe + Bedrock) $900–$1,500/mo Full analysis, custom
Azure Video Indexer ~$180–$360/mo Transcript + faces + scenes (limited free tier)
Google Video Intelligence ~$120–$360/mo Visual labels + transcription (feature-dependent)

Azure Video Indexer and Google Video Intelligence are cheaper for video, but neither is natively integrated into an AWS document pipeline. For AWS-native architectures, BDA Video is the clear choice over DIY Rekognition pipelines.

Service Selection Decision Tree

Document type is standardized (invoice, W-2, tax form)?
  └── Volume > 50K/month? → Textract purpose-built API (AnalyzeExpense, AnalyzeID)
  └── Volume < 50K/month? → Either; hybrid routing pays above ~10K/month

Document layout varies significantly (contracts, medical records, filings)?
  └── Need clean JSON output for downstream LLM? → BDA Custom Output
  └── Need raw text for RAG ingestion? → BDA Standard Output

Mixed modalities in same pipeline (docs + audio + video)?
  └── BDA (single API, single billing surface)

Audio call center analytics?
  └── BDA Standard Audio ($0.006/min) — dominant vs. Transcribe + NLP

Video content moderation/analytics in AWS?
  └── BDA Standard Video — dominant vs. DIY Rekognition pipeline

Non-AWS or specialized accuracy requirement?
  └── Reducto (complex layouts), Unstructured.io (RAG preprocessing)

9. Key Numbers for Financial Modeling

Metric Value
BDA Standard Document $0.010/page
BDA Custom Document (30 fields) $0.040/page
BDA Custom Document (each field above 30) +$0.0005/page
BDA Standard Audio $0.006/minute
BDA Standard Video $0.050/minute
10K document pages/day, Standard ~$3,000/month
10K document pages/day, Custom (30 fields) ~$12,000/month
BDA Audio vs. Transcribe savings ~75%
BDA Video vs. DIY Rekognition savings ~50–80%
Hybrid routing savings (BDA + Textract) ~54%
Blueprint creation/versioning No charge
Failed invocation charge None
Volume discounts None (flat pricing)
S3 event notification cost ~$0.10 per 1M events (negligible)

Sources