AI-Powered Video Analysis and Highlight Generation Platform

accenture · Media & Entertainment · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Video highlight generation and media operations

Accenture wanted to automate long-form video analysis and short-form highlight creation without losing editorial quality, brand alignment, or human review control.

Constraints:

• traditional highlight workflows require manual review, clip selection, editing, transitions, narration, checks, and formatting
• video-processing workloads vary from sports highlights to retail offer generation
• some tasks require fast inference while others require deeper multimodal reasoning
• brand standards and editorial quality still need review
• the source is an AWS/Accenture promotional article and does not disclose independent production accuracy metrics

Architecture

Accenture Spotlight uses Amazon Nova foundation models and Amazon Bedrock Agents in a multi-layered agent workflow. A central orchestrator decomposes the work and coordinates three specialized super agents. The video processing agent extracts metadata using research, visual analysis, and audio analysis utility agents. The short video generation agent selects sections of interest, builds the clip sequence, and postprocesses output for publishing. A reviewer agent validates quality, accuracy, brand alignment, and user requirements, with optional human-in-the-loop review.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  central-orchestrator-agent["central-orchestrator-agent"] --> video-processing-agent["video-processing-agent"]
  central-orchestrator-agent["central-orchestrator-agent"] --> short-video-generation-agent["short-video-generation-agent"]
  central-orchestrator-agent["central-orchestrator-agent"] --> reviewer-agent["reviewer-agent"]
  central-orchestrator-agent["central-orchestrator-agent"] --> task-driven-model-selector["task-driven-model-selector"]
  video-processing-agent["video-processing-agent"] --> research-agent["research-agent"]
  video-processing-agent["video-processing-agent"] --> visual-analysis-agent["visual-analysis-agent"]
  video-processing-agent["video-processing-agent"] --> audio-analysis-agent["audio-analysis-agent"]
  video-processing-agent["video-processing-agent"] --> short-video-generation-agent["short-video-generation-agent"]
  research-agent["research-agent"] --> visual-analysis-agent["visual-analysis-agent"]
  visual-analysis-agent["visual-analysis-agent"] --> section-of-interest-agent["section-of-interest-agent"]
  short-video-generation-agent["short-video-generation-agent"] --> section-of-interest-agent["section-of-interest-agent"]
  short-video-generation-agent["short-video-generation-agent"] --> video-generation-agent["video-generation-agent"]
  short-video-generation-agent["short-video-generation-agent"] --> video-postprocessing-agent["video-postprocessing-agent"]
  short-video-generation-agent["short-video-generation-agent"] --> reviewer-agent["reviewer-agent"]
  video-generation-agent["video-generation-agent"] --> reviewer-agent["reviewer-agent"]
  video-postprocessing-agent["video-postprocessing-agent"] --> reviewer-agent["reviewer-agent"]
  reviewer-agent["reviewer-agent"] --> human-reviewer["human-reviewer"]

Components

Data flow

Long-form or live video enters Spotlight. The orchestrator selects traditional AI or Nova models based on task complexity and latency needs, then invokes video processing agents to extract recipes, visual segments, people tags, timestamps, diarization, and transcripts. The short-video generation agent identifies sections of interest, assembles a sequence, and postprocesses for mobile formats or branded overlays. Reviewer agents and optional human review check quality before output is distributed.

Why this architecture

Spotlight uses agent decomposition because highlight creation is not one task: it needs research into what makes clips perform, multimodal video/audio understanding, segment selection, clip construction, post-processing, review, and optional human oversight. Task-driven model selection lets fast traditional models handle simple or time-sensitive operations while Nova handles deeper multimodal reasoning.

Models used: Amazon Nova · Amazon Bedrock Agents · traditional AI models · person recognition and tracking · speech diarization · transcription

Scale

• Spotlight supports both live streams and pre-recorded video.
  “It supports both live stream feeds and pre-recorded video”article, stated
• Custom highlight generation is reported to happen in minutes instead of hours or days.
  “custom highlight generation happening in minutes, reducing from hours or days”article, stated
• The comparison table cites minutes for 2-3-hour sessions.
  “Video Processing Latency Minutes for 2–3-hour sessions Hours to days”article, stated
• Accenture reports over 10x better cost savings with Amazon Nova-powered agents.
  “Amazon Nova-powered agents deliver over 10 times better cost savings over traditional highlight creation methods”article, stated

Alternatives considered (roads not taken)

Manual highlight creation

Rejected because: Manual review, cutting, transitions, narration, checks, and formatting preserve control but create scaling bottlenecks.

This post showcases how Accenture Spotlight delivers a scalable, cost-effective video highlight generation solution using [Amazon Nova"

— article, stated

One generic model for every video task

Rejected because: Spotlight dynamically chooses task-specific models or Nova FMs depending on latency and reasoning depth.

Pure automation without editorial review

Rejected because: The platform preserves human-in-the-loop quality assurance and collaborative refinement for brand and accuracy needs.

Patterns

multi-agent-orchestration (agent architecture)

A central orchestrator coordinates three super agents and utility agents across analysis, generation, and review.

model-routing (model operations)

Spotlight routes between traditional AI and Nova foundation models based on complexity and latency.

human-in-the-loop-gate (quality)

Content owners can use human review for quality assurance, editorial oversight, and brand alignment.

structured-output-contract (workflow)

Agents pass metadata, section-of-interest recommendations, segment patterns, and review results through defined workflow roles.

Tradeoffs

Use agent orchestration instead of a monolithic pipeline.

Chose Central orchestrator plus specialized agents over Single large model or one fixed pipeline.

Rationale: Specialized agents let Spotlight handle research, visual analysis, audio context, generation, post-processing, and review separately.

Deploy serverless and on-demand.

Chose Serverless scalable LLM invocation over Resource-heavy statically provisioned infrastructure.

Rationale: Video workloads can be bursty and cross-industry, so on-demand deployment reduces idle infrastructure cost.

Gotchas (what bit them)

Manual video editing preserves quality but does not scale.

Root cause: Editors must inspect footage, identify moments, cut clips, add narration/transitions, and run manual checks.

Resolution: Decompose the workflow into specialized agents while preserving reviewer and human-in-loop checks.

Different video tasks need different latency/intelligence tradeoffs.

Root cause: Retail offers at the edge and complex video processing have different response-time and reasoning requirements.

Resolution: Use task-driven model selection between traditional AI and Nova models.

Evaluation

Spotlight evaluates clips through reviewer-agent checks for quality, accuracy, brand alignment, and user requirements, with an optional human-in-the-loop review path. The article also reports comparative operational metrics for latency and cost.

Metrics: quality · accuracy · brand alignment · user requirement fit · video processing latency · highlight review cost · overall generation cost

Outcomes

• Spotlight supports personalized short-form generation, sports highlights, content matching, and realtime retail offers.
• Custom highlight generation is reported in minutes instead of hours or days.
• The article reports over 10x cost savings compared with traditional highlight creation methods.
• The platform can process both live streams and pre-recorded video.

The case reports minutes for 2-3-hour sessions, 10x lower review and generation cost, and cross-industry use cases rather than customer deployment counts.