Beyond the Chatbot: People and AI Agents in the Same Application

· elbtech

Updated:

Integrating AI into software can mean more than placing a chat window next to an existing interface. The interesting part is people and agents working with the same application: seeing relevant state, proposing changes, and using defined business operations. For us, that is a software architecture and usability question, not just a prompting exercise.

One Workflow, Several Ways In

Imagine an employee asking an application to prepare a quotation from an inquiry. An agent proposes line items. The application shows quantities, prices, and changes in an ordinary form. The employee corrects an item and approves sending. Natural language makes it easier to start; structured views make the result reviewable. This describes an interaction pattern, not an already delivered elbtech application.

Agentic UI Is Not the Same as Chat

Terms such as agentic UI and Agent UX describe these interfaces. They concern the interaction between user, agent, and application: what is running, what is only a proposal, and what has actually changed? A dashboard with reviewable suggestions may be more useful than a long conversation. Generated interfaces are one possible technique, not a prerequisite.

Shared Business Rules Do Not Mean Identical Permissions

Whether a person clicks or an agent calls a tool, the business operation must enforce its requirements on the server. An agent needs an appropriate identity and narrowly scoped permissions. Direct database access must not let it bypass the rules. An API or MCP tool makes an action reachable; it does not automatically establish whether that action is authorized or appropriate in context.

Proposal, Approval, Execution, Outcome

  • Proposed changes must be distinguishable from committed changes.
  • Consequential actions need risk-appropriate approvals and clearly bounded authority.
  • When people edit in parallel, the application must handle stale context and conflicts.
  • Cancelling a conversation does not automatically cancel or reverse an external action.
  • A confirmed operation result is stronger evidence than a model-generated success message.

Choosing the Right Action Is a Separate Verification Problem

An action can be technically authorized and still be the wrong response to a user's intent. Model evaluation, business-operation tests, and interaction review therefore belong together. Documents and external tool results remain untrusted input; they must not grant additional permissions. An approval button alone is not a security architecture.

The Application Remains the Product

Our focus is the whole system: interface, agent integration, business logic, data, tests, and operations. People should not have to shuttle between an AI portal and their actual work. They need understandable state and ways to correct the system. That connects accountable AI-assisted development with applications that integrate AI.

Further Reading