SUPPLY CHAIN HANDBOOK

AI services: models, data and application integration

Prepared by IndustryAtlas Editorial principles & corrections

Compare compute, data, models, applications and implementation across language, visual generation and machine vision. Model capabilities, software features and deployed outcomes need separate verification.

Most recent source review: 2026-09-29

Reading guide revised: 2026-10-30

What must a business add after acquiring model capability?

Models process or generate information; applications must connect them to relevant data, permissions and real workflows. Knowledge tools can retrieve documents before asking a model to answer. Creative products integrate generation with editing and delivery. Machine vision starts with dependable image capture and connects recognition to equipment actions. These routes share some infrastructure, but differ in data preparation, acceptance criteria and error handling. Deployment value therefore extends beyond the model itself.

Enterprise documents and retrieval → Knowledge applications
Preparation and indexing make relevant material retrievable. The model receives retrieved context, while the application manages permissions, references and interfaces. Retrieval-augmented generation is distinct from retraining.
Content and generative models → Creative workflows
Models generate or modify images and video. Tools add controls, iterative editing and asset management; source-content rights and output usage terms must be checked for the relevant product.
Captured images and recognition models → Validated inspection systems
Cameras, lenses and lighting supply usable inputs. Integrators connect recognition to equipment and validate results against actual objects, production timing and agreed acceptance criteria.

What changes the choice?

Select the work to improve before comparing platforms. Knowledge retrieval, controlled creation and production inspection need different acceptance criteria; a general model score cannot substitute for them.

Sources for this overview

AI software & services

AI Software and Services

Turning compute resources into usable applications

AI extends from chips, servers and data centers to cloud compute, then models, data, development platforms and integration services. Shared foundations support three branches: language and knowledge applications, image and video generation, and machine vision.

Entry sources reviewed: 2026-09-29

Choose a branch from shared foundations

All three branches can use cloud or local computation, and language models can include image understanding. Categories can overlap.

Data centers connect upstream to GPUs, servers, networks, power and cooling. Agent platforms support development and deployment; agent applications deliver user-facing tasks.

A network of roles rather than one assembly line

Customers can buy finished services, combine models with knowledge sources, or hire an integrator. Evaluation and governance span development and operation.

Graph connections describe general functions, not confirmed purchasing or partnerships between listed companies.

Cloud AI Compute Language Models and APIs Enterprise Data and Knowledge Retrieval AI Development and Agent Platforms AI Applications and Agent Services Content and Data Licensing Image and Video Generation Models AI Creative Tools and Workflows Machine Vision Image Capture Vision Models and Platforms Vision Inspection and Automation Integration Enterprise AI Integration AI Evaluation, Monitoring and Governance

Distinction: Generating images, analyzing images and language applications are different roles. The words video or LLM alone cannot establish which workload needs more compute.

Sources

Explore this role on the map Suggest a correction Back to contents

Cloud AI Compute

Training and inference resources offered as a service

Cloud providers offer compute, storage and networking to model developers and application operators. Training builds or adjusts models; inference runs deployed models. Enterprises may also use their own infrastructure or edge equipment.

Entry sources reviewed: 2026-09-29

Different from a data center or model API

A data center is the equipment and facility environment. Cloud compute exposes resources as a service, while a model API provides a higher-level capability.

Training emphasizes coordinated resources and data movement; inference deployment must balance latency, throughput, scaling and cost.

Distinction: Shared infrastructure requirements do not mean every business purchases from the same cloud provider.

Company roles

Scope of these company examples

CoreWeave · United States · Example role · AI cloud computing
GPU computing, storage, networking and cloud environments.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

Enterprise AI Integration

Deploying models and tools within business operations

Integrators help define requirements, prepare data, choose models and tools, connect existing systems and deploy AI applications. Their role includes acceptance testing and ongoing improvement, not only installing technology.

Entry sources reviewed: 2026-09-29

Examples of the integration role

SYSTEX AI BOX combines language models, GPU management and no-code applications. Accenture offers enterprise generative-AI implementation services.

An integrator may use outside platforms and also offer its own products; both roles can coexist.

Distinction: Offering an AI solution does not mean a company trained every underlying foundation model itself.

Company roles

Scope of these company examples

SYSTEX · Taiwan · Example role · Enterprise AI deployment
AI BOX integrates models, GPU management and enterprise knowledge applications.
Product / role source
Sources reviewed: 2026-09-29
Accenture · Ireland · Example role · Generative AI implementation services
Enterprise data, model and workflow implementation support.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI Evaluation, Monitoring and Governance

Quality, operation and policy across the application lifecycle

Evaluation checks responses and task results; monitoring tracks operational change; governance manages systems, policies and risks. These activities span development, deployment and continuing operation, with metrics adapted to language, visual generation or industrial recognition.

Entry sources reviewed: 2026-09-29

Tools cover different responsibilities

AWS AgentCore Evaluations focuses on agent quality and task evaluation. IBM watsonx.governance provides monitoring and governance capabilities.

One product should not be treated as complete governance for every AI system or workload.

Distinction: Evaluation scores and governance tools cannot guarantee every answer is correct. Metrics and thresholds must fit the actual use.

Company roles

Scope of these company examples

Amazon (including AWS) · United States · Example role · AgentCore Evaluations
Evaluation of agent responses, task completion and tool use.
Product / role source
Sources reviewed: 2026-09-29
IBM · United States · Example role · watsonx.governance
AI system inventory, monitoring and governance.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI language & knowledge

Language Models and APIs

Text and code capabilities for applications

Model developers provide capabilities for understanding and generating text, code and related outputs. Applications call APIs or, where permitted and supported, deploy models themselves. Some models also accept images or other inputs; capabilities vary by model.

Entry sources reviewed: 2026-09-29

A model is one part of an application

Enterprise support also needs knowledge sources, access permissions, interfaces and business-system connections. OpenAI and Anthropic illustrate model/API roles rather than a complete customer-service deployment.

Training data, prompt context and retrieved evidence are different inputs with different roles.

Distinction: Fluent answers do not guarantee accuracy or knowledge of the latest private company information.

Company roles

Scope of these company examples

OpenAI · United States · Example role · Large language models and APIs
Text generation, summarization, conversation, code generation and other model capabilities through the OpenAI API.
Product / role source
Sources reviewed: 2026-09-29
Anthropic · United States · Example role · Claude models and API
Language, coding and selected multimodal model capabilities.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

Enterprise Data and Knowledge Retrieval

Finding relevant business information for model responses

Enterprise documents and data can be organized into searchable knowledge sources that applications retrieve when answering. Retrieval-augmented generation supplies found information to a model; it is not the same as retraining that model.

Entry sources reviewed: 2026-09-29

Choose retrieval separately from the model

Databricks AI Search provides vector, keyword and related search capabilities; the application decides how to use retrieved results.

Official documentation renamed Vector Search to AI Search. Data freshness, access controls and source quality still need management.

Distinction: RAG still needs answer validation. Indexing documents does not grant every user permission to read them.

Company roles

Scope of these company examples

Databricks · United States · Example role · AI Search retrieval platform
Enterprise data indexing with vector and keyword retrieval and RAG support.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI Development and Agent Platforms

Infrastructure for connecting models, data and tools

Agent platforms help developers combine models, data and business tools into task-performing systems. They support deployment, tool connections, identity, permissions and monitoring. They are the foundation for applications rather than the final support or office service a user buys.

Entry sources reviewed: 2026-09-29

Separate platforms from finished applications

AWS AgentCore illustrates a platform for developing, deploying and operating agents. Business applications package those capabilities into specific functions.

Not every AI application needs autonomous multi-step execution. Permissions and monitoring should fit the actual workflow.

Distinction: An agent that can call tools should not automatically have unrestricted authority to perform every action.

Company roles

Scope of these company examples

Amazon (including AWS) · United States · Example role · AgentCore development and operations
Agent deployment, tool connections, identity management and observability.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI Applications and Agent Services

User-facing features built from models and data

Application providers package models and data into services such as customer support, knowledge answers, coding and workflow automation. The same model can power different products, whose value also depends on workflow integration.

Entry sources reviewed: 2026-09-29

Customers have several delivery options

Enterprises can buy a ready-made application, build through development platforms or hire an integrator.

Salesforce Agentforce is an example combining agents, enterprise data and application workflows. A functional example does not establish a universal customer configuration.

Distinction: An application brand, model developer and systems integrator may be separate companies or roles covered by the same business.

Company roles

Scope of these company examples

Salesforce · United States · Example role · Agentforce enterprise agents
Connects enterprise agents to data, applications and business processes.
Product / role source
Sources reviewed: 2026-09-29
SYSTEX · Taiwan · Example role · Enterprise knowledge applications
AI BOX provides enterprise knowledge search, Q&A and no-code applications.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI image & video generation

Content and Data Licensing

Rights for source materials and their intended uses

Image and video generation involves images, footage, captions and other data. Content suppliers and rights holders provide materials with usage permissions. Training-material terms must be read separately from terms governing generated outputs.

Entry sources reviewed: 2026-09-29

One documented training-data example

Adobe states that Firefly is trained on licensed content, including Adobe Stock, and public-domain content.

This example does not establish that every generation model uses the same sources or licensing model.

Distinction: A description of data sources does not automatically grant the same rights for every input, output or use case.

Company roles

Scope of these company examples

Adobe · United States · Example role · Adobe Stock licensed content
Adobe identifies it as one source of Firefly training material.
Product / role source
Sources reviewed: 2026-09-29

Sources

Adobe | Firefly training materials and generative AI approach

Explore this role on the map Suggest a correction Back to contents

Image and Video Generation Models

Models that create or transform visual content

Generation models turn prompts, reference images or other inputs into new images and videos, or modify existing content. Models and APIs supply the capability; creative tools place it within design and media workflows.

Entry sources reviewed: 2026-09-29

Generation and recognition serve different purposes

Generation creates content; recognition analyzes existing images. They can be combined while remaining distinct roles.

Compute demand depends on the model, resolution, video duration, generation steps and usage volume.

Distinction: Video generation cannot automatically be described as more compute-intensive than every language-model workload.

Company roles

Scope of these company examples

Runway · United States · Example role · Image and video generation models
Text-to-video, image-to-video and related generation models.
Product / role source
Sources reviewed: 2026-09-29
Adobe · United States · Example role · Firefly generation capabilities
Generative AI capabilities for creative work.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI Creative Tools and Workflows

Generation and editing within a creative process

Creative tools bring generation, editing and asset management into design and media workflows. They let creators control inputs, iterate and deliver work. Model developers can also offer tools directly to users.

Entry sources reviewed: 2026-09-29

One company can participate at several layers

Adobe spans assets, generation and creative tools; Runway offers models and creative capabilities.

This node describes the tool role. Advertising or film customers are not treated as confirmed purchasers without evidence.

Distinction: Providing generation models and providing creative tools are overlapping, not mutually exclusive, roles.

Company roles

Scope of these company examples

Adobe · United States · Example role · Firefly creative tools
Generation and editing capabilities within creative workflows.
Product / role source
Sources reviewed: 2026-09-29
Runway · United States · Example role · Generative creative capabilities
Image and video generation and transformation.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

AI machine vision

Machine Vision Image Capture

Optics, lighting and cameras before model analysis

Machine vision starts with usable images of the real scene. Cameras, sensors, lenses and lighting jointly determine image quality and must fit production speed, field of view and the inspected object.

Entry sources reviewed: 2026-09-29

Capture quality comes before model judgment

Lighting, lens choice and camera configuration affect whether the needed information exists in the image. A new model cannot automatically recover information never captured.

Camera manufacturers supply capture equipment, a different role from recognition software or production-line integration.

Distinction: Not every industrial camera contains AI, and vision systems do not all use the same sensing equipment.

Company roles

Scope of these company examples

Basler · Germany · Example role · Machine-vision cameras
Area-scan, line-scan, 3D and embedded-vision cameras.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

Vision Models and Platforms

Understanding existing images and video

Vision models analyze images for objects, defects, actions or events. Development platforms support models and processing pipelines, then deployment at the edge or in the cloud according to the use case.

Entry sources reviewed: 2026-09-29

Data and deployment conditions interact

Model development must consider real-world examples and evaluation methods. Operation still needs monitoring as environments change.

NVIDIA Metropolis covers vision models, tools and an edge-to-cloud deployment ecosystem; it is not one universal production-line configuration.

Distinction: Recognizing existing images and generating new ones serve different purposes. Some multimodal systems combine image understanding with language.

Company roles

Scope of these company examples

NVIDIA · United States · Example role · Metropolis vision AI platform
Vision AI models, tools and deployment capabilities from edge to cloud.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

Vision Inspection and Automation Integration

Connecting captured images to actions on the production floor

Integration combines image capture, recognition and equipment actions into usable systems for defect checks, measurement, identification or robot guidance. The object, cycle time and acceptance criteria shape the design.

Entry sources reviewed: 2026-09-29

Verify functions for the actual product

Cognex In-Sight systems support inspection, measurement, identification and guidance. AI and conventional vision tools can work together.

Functional graph links do not prove a specific supply relationship between Basler, NVIDIA and Cognex.

Distinction: Industrial vision is not entirely AI. Cameras, rule-based algorithms and practical site engineering remain important.

Company roles

Scope of these company examples

Cognex · United States · Example role · In-Sight industrial vision systems
Automated inspection, measurement, identification and robot guidance.
Product / role source
Sources reviewed: 2026-09-29

Sources

Explore this role on the map Suggest a correction Back to contents

How the roles connect

Each connection appears once in this handbook. Classification describes a branch; technology links compare or relate approaches; service arrows point to the supported activity. Company examples do not establish a supply contract.

Connections, handoffs and sources
From / to	Handoff / relationship	Evidence
AI Data Centers Cloud AI Compute	Equipment resources offered as services Cloud compute uses data-center computing, storage and networking resources; this link represents functional roles.	CoreWeave \| AI cloud-computing platform 2026-09-29
Cloud AI Compute Language Models and APIs	Training and inference compute These workloads may use cloud resources, self-hosted systems or edge deployment; no cloud-to-model-company purchase is implied.	CoreWeave \| AI cloud-computing platform 2026-09-29
Cloud AI Compute Image and Video Generation Models	Training and inference compute These workloads may use cloud resources, self-hosted systems or edge deployment; no cloud-to-model-company purchase is implied.	CoreWeave \| AI cloud-computing platform 2026-09-29
Cloud AI Compute Vision Models and Platforms	Training and inference compute These workloads may use cloud resources, self-hosted systems or edge deployment; no cloud-to-model-company purchase is implied.	CoreWeave \| AI cloud-computing platform 2026-09-29
Language Models and APIs AI Development and Agent Platforms	Access to model capabilities Agent platforms can combine compatible models and tools; listed representatives are not thereby designated partners.	AWS \| Amazon Bedrock AgentCore 2026-09-29
Enterprise Data and Knowledge Retrieval AI Development and Agent Platforms	Retrieval and knowledge inputs Agents or applications can use enterprise retrieval results as context, with data updates and access permissions enforced.	Databricks \| AI Search and enterprise data retrieval 2026-09-29
AI Development and Agent Platforms AI Applications and Agent Services	Application development and operation Platforms support building and deploying agents; ready-made applications may also use their own architectures.	AWS \| Amazon Bedrock AgentCore 2026-09-29
Language Models and APIs AI Applications and Agent Services	Applications call models Applications can call models directly without passing through every platform layer listed here.	Anthropic \| Claude models and capabilities 2026-09-29
Enterprise Data and Knowledge Retrieval AI Applications and Agent Services	Enterprise knowledge inputs Knowledge search and enterprise Q&A need data-layer support; applications need not share the same RAG architecture.	SYSTEX \| SYSTEX AI BOX 2026-09-29
Content and Data Licensing Image and Video Generation Models	Licensed training content Adobe states that Firefly training includes licensed and public-domain content; this does not establish other models' content sources.	Adobe \| Firefly training content and generative AI approach 2026-09-29
Image and Video Generation Models AI Creative Tools and Workflows	Generation and editing capabilities Tools bring model capabilities into creative workflows; one company may provide both models and tools.	Runway \| Image- and video-generation models 2026-09-29
Machine Vision Image Capture Vision Models and Platforms	Images from the operating environment Captured images provide analysis inputs; cameras and vision platforms require interface and workload integration, not a designated supplier relationship.	NVIDIA \| Metropolis vision AI platform 2026-09-29
Vision Models and Platforms Vision Inspection and Automation Integration	Recognition and analysis Model outputs enter inspection and automation workflows, which still require on-site validation and system integration.	NVIDIA \| Metropolis vision AI platform 2026-09-29
Machine Vision Image Capture Vision Inspection and Automation Integration	Image-capture equipment Industrial vision needs image capture; integrated and external-camera approaches vary by product.	Cognex \| In-Sight industrial vision systems 2026-09-29
Edge AI Devices Vision Models and Platforms	On-site deployment support Vision AI can run on edge devices or work with cloud services; recognition does not always require the cloud.	NVIDIA \| Metropolis vision AI platform 2026-09-29
Vision Inspection and Automation Integration Robots and Intelligent Equipment	Vision guidance and on-site coordination Industrial vision can guide robots; robots do not all use the same vision solution.	Cognex \| In-Sight industrial vision systems 2026-09-29
Language Models and APIs Vision Models and Platforms	Multimodal understanding overlap Vision-language models combine image understanding and language processing, so these categories are not mutually exclusive.	NVIDIA \| Metropolis vision AI platform 2026-09-29
Enterprise AI Integration AI Applications and Agent Services	System connections and deployment Enterprise implementation spans data, models and application workflows; an integrator need not supply every component.	Accenture \| Enterprise generative AI implementation services 2026-09-29
AI Evaluation, Monitoring and Governance AI Development and Agent Platforms	Quality evaluation and ongoing monitoring Evaluation should cover development testing and production operation, with metrics defined for the task and tools.	AWS \| AgentCore Evaluations 2026-09-29
AI Evaluation, Monitoring and Governance AI Applications and Agent Services	Quality evaluation and ongoing monitoring Evaluation should cover development testing and production operation, with metrics defined for the task and tools.	AWS \| AgentCore Evaluations 2026-09-29
AI Evaluation, Monitoring and Governance AI Creative Tools and Workflows	Cross-layer governance and tracking AI governance spans the application lifecycle; this functional category does not imply identical evaluation metrics across tools.	IBM \| watsonx.governance 2026-09-29
AI Evaluation, Monitoring and Governance Vision Inspection and Automation Integration	Cross-layer governance and tracking AI governance spans the application lifecycle; this functional category does not imply identical evaluation metrics across tools.	IBM \| watsonx.governance 2026-09-29
AI Software and Services AI Applications and Agent Services	Branches by service use Reading branches organized by use case, not procurement sequence; branches may overlap.	AWS \| Amazon Bedrock AgentCore 2026-09-29
AI Software and Services AI Creative Tools and Workflows	Branches by service use Reading branches organized by use case, not procurement sequence; branches may overlap.	AWS \| Amazon Bedrock AgentCore 2026-09-29
AI Software and Services Vision Inspection and Automation Integration	Branches by service use Reading branches organized by use case, not procurement sequence; branches may overlap.	AWS \| Amazon Bedrock AgentCore 2026-09-29
Vision Models and Platforms Drone Software and Operations Services	Image-data analysis support Inspection can use vision analysis for data processing; this is a functional connection, not a specific model contract.	Skydio \| Asset and infrastructure inspection 2026-09-29

Scope of company examples

Functional roles follow official product descriptions; this is neither an exhaustive market list nor a list of company partnerships.

What must a business add after acquiring model capability?

Enterprise documents and retrieval → Knowledge applications

Content and generative models → Creative workflows

Captured images and recognition models → Validated inspection systems

What changes the choice?

Explore the technologies and changes

OpenAI’s Annualized Revenue Falls Short of Earlier Reports: What $50 Billion Versus $70 Billion Says About AI Bubble Risk

Anthropic Turns an Operating Profit in Q2: Can High Gross Margins Fund Continued AI Research?

AI software & services

Choose a branch from shared foundations

A network of roles rather than one assembly line

Sources

Different from a data center or model API

Company roles

Sources

Examples of the integration role

Company roles

Sources

Tools cover different responsibilities

Company roles

Sources

AI language & knowledge

A model is one part of an application

Company roles

Sources

Choose retrieval separately from the model

Company roles

Sources

Separate platforms from finished applications

Company roles

Sources

Customers have several delivery options

Company roles

Sources

AI image & video generation

One documented training-data example

Company roles

Sources

Generation and recognition serve different purposes

Company roles

Sources

One company can participate at several layers

Company roles

Sources

AI machine vision

Capture quality comes before model judgment

Company roles

Sources

Data and deployment conditions interact

Company roles

Sources

Verify functions for the actual product

Company roles

Sources

How the roles connect

Scope of company examples

Long-form guides (Chinese)

AI 供應鏈：語言模型、生成影像、機器視覺與企業導入