AI applications, AI agents and private AI
These terms describe different things. An AI application gives people an AI capability for a defined job. An AI agent can use information and tools to carry work through several steps. Private AI describes how the technology and data are hosted and controlled.
Understanding those differences makes it easier to choose the platform, access and level of human involvement the work requires.
The terms describe different things
AI feature
An AI feature sits inside software people already use. It may summarise a document, draft text, classify a request, extract information or suggest the next action.
Examples include Copilot features in Microsoft 365 and AI Builder capabilities within Power Platform.
AI application
An AI application is built for a defined business purpose. It combines a user interface, business information, a model and sometimes a workflow or system connection.
Examples include a document review application, an internal knowledge tool or an application that classifies and routes incoming requests.
AI agent
An AI agent can work towards a goal, use approved information and tools, and complete several steps. Depending on its access, it may retrieve information, update a system, trigger a workflow or pass work to another agent.
The important difference is not whether the interface looks like a chatbot. It is what the agent can access and what actions it can take.
Private AI
Private AI is not a separate type of application or model. It describes the architecture around the AI, including where the model runs, how information reaches it, who can access it and whether network traffic uses public or private paths.
An AI application or agent can use a private architecture, but the term alone does not confirm data residency, local processing or the absence of an external provider.
Assistive AI prepares the work, agentic AI can move it forward
Assistive and agentic AI are not completely separate categories. They sit on a spectrum based on what the technology is allowed to do.
- 01
Retrieve, summarise and draft
The AI finds relevant information, summarises material or prepares a draft. A person decides what to do with the output.
ControlsTypical controls include identity, permissions and approved information sources.
- 02
Classify and recommend
The AI classifies a request, identifies an exception or recommends a next action. A person still makes the decision.
ControlsThis adds a need for repeatable evaluation, clear escalation and a record of the recommendation.
- 03
Prepare an action for approval
The AI completes the preparatory work, such as populating a record or assembling an approval request, but waits for a person before taking the next step.
ControlsThe approval point needs to be explicit and recorded.
- 04
Act within defined limits
The agent updates records, triggers workflows or completes approved steps within agreed boundaries.
ControlsAt this level, identity, access, logging, monitoring, exception handling and human escalation become part of the production design.
Where applications and agents are being used
Document review
Review tenders, contracts, reports, submissions and supporting material together, then identify relevant clauses, exclusions, differences or missing information.
Knowledge retrieval
Retrieve answers and supporting material from approved SharePoint sites, documents, policies and other business information.
Request triage
Read an incoming request, classify it, gather the relevant context and direct it to the correct person or workflow.
Reporting and analysis
Bring together information from several sources, identify patterns or exceptions and prepare a current view for review.
Multi-step work
Use information and tools across several stages, such as checking a request, retrieving a record, updating a system and triggering the next action.
The same use case can begin as assistive AI and become more agentic later. The change should follow evidence from the work, not pressure to give the agent more autonomy.
Which platform fits which type of work
Microsoft 365 Copilot
Microsoft 365 Copilot supports work across Outlook, Teams, Word, Excel and other Microsoft 365 applications. It is primarily assistive and works within the information each user is already permitted to access.
Agent Builder
Agent Builder provides a simple way to create declarative agents within Microsoft 365 Copilot using natural language. It suits lighter agents grounded in Microsoft 365 information.
Microsoft Copilot Studio
Copilot Studio is a low-code environment for building and managing agents and workflows. It adds connections, actions, agent flows and more control over how the agent behaves.
Microsoft Foundry
Microsoft Foundry is used for custom production agents that need model choice, substantial multi-step logic, evaluation, connections to several systems, their own identity or private networking.
A Foundry agent can still be made available through Teams or Microsoft 365 Copilot, so the user experience and the production platform do not need to be the same thing.
Microsoft Power Platform
Power Apps, Power Automate and Dataverse provide structured applications, workflows and data. AI Builder, Copilot features and agent flows can add AI within those processes.
Many implementations use more than one platform. A Power App may provide the interface, Power Automate may run defined workflow steps and a Foundry agent may handle document analysis or reasoning.
What private AI means in practice
Private AI should be described in specific architectural terms rather than used as a general reassurance.
Controlled cloud environment
The application runs in a managed cloud environment with defined identities, permissions, information sources, logging and administrative controls.
Private networking
Microsoft Foundry supports private endpoints and virtual network options. These can keep traffic between the agent and connected Azure services on private network paths, depending on the architecture used.
Local model execution
Foundry Local can run supported models on a device or local infrastructure. Foundry Local remains in preview, so its capabilities and production suitability need to be checked for the proposed workload.
Data residency
Private networking does not automatically mean information is processed or stored in Australia. Data location depends on the service, region, model and features being used and needs to be checked for each implementation.
Access and action determine the controls
The controls should reflect what the AI can reach and the consequences of what it can do.
Identity and permissions
Microsoft 365 Copilot works within existing permissions. Custom agents need their own access model based on the information, tools and systems their role requires.
Information boundaries
Approved information sources, external sharing, sensitivity labels and data loss prevention rules determine what information the AI can use and where it may move.
Human approval
A person should approve actions involving money, legal commitments, safety, employment, reputation or other material consequences.
Logging and evaluation
The organisation needs a record of what the AI did and repeatable evaluation against examples of the work, particularly when the agent can take action.
Ownership and lifecycle
Every application or agent needs an owner, support arrangements and a process for changing or retiring it when the business need, model or connected systems change.
AI near operational technology needs tighter limits
AI can help operational teams retrieve technical information, summarise maintenance records, classify exceptions and prepare recommendations from site or production data.
It should not be given uncontrolled access to change a live operational state. Any connection near operational technology needs clear separation between business IT and OT, explicit approval points, logging, change control and agreement with the people responsible for the operational system.
AI does not replace the underlying OT security controls. Asset visibility, segmentation, controlled remote access and vendor requirements still apply.
Automation follows rules, AI interprets information
Automation works well when the steps and rules are known. It can route requests, manage approvals, update records, send notifications and produce consistent outputs.
AI becomes useful when the work also involves interpreting documents, classifying information, retrieving context, drafting or making a recommendation.
The two can work together. An agent might read an incoming document and identify what it contains, while Power Automate routes it through the agreed approval process and records the outcome.
Explore business applications and automationAI applications and agents questions
Is an AI application the same as an AI agent?
No. An AI application provides a defined AI capability. An AI agent can also use tools and complete steps towards a goal. Some AI applications contain agents, but many do not.
Is Microsoft 365 Copilot an AI agent?
Microsoft 365 Copilot is primarily an assistive AI product across Microsoft 365. Organisations can extend it with agents built through Agent Builder, Copilot Studio or custom development.
Does private AI mean our data stays in Australia?
No, not automatically. Private networking, tenant controls, local model execution and Australian data residency are different things. The service, region, model and processing arrangement must be checked before making a residency claim.
Can an AI agent access everything in our environment?
No. It can access only the information, systems and tools granted to its identity. Microsoft 365 Copilot works within existing permissions, while a custom agent receives access according to its defined role.
Does every agent need human approval?
Not for every step. Retrieval, classification and low-consequence workflow actions may run within agreed limits. A person should approve consequential actions involving money, legal commitments, safety, employment or reputation.
Can an application use our own business information?
Yes. It can retrieve information from approved Microsoft 365 locations, databases, applications and APIs. Giving an application access to business information does not necessarily mean training a model on that information.
Should we automate the process before adding AI?
Sometimes. If the work has no clear stages, ownership or exceptions, structure those parts first. If the workflow is already clear and the difficult part is interpreting information, AI may be useful immediately.
Move from a defined use case to an agent in production
Inlight IT implements AI agents across Microsoft Foundry, Copilot Studio and Agent Builder, connects them to the required information and systems, and can look after them once they are running.