AI agents and custom AI apps for specific business workflows

From chat to work tool

A generic AI chat answers questions. An automated workflow performs known steps. A specialised AI agent can interpret input and choose between limited tools. A custom app brings together user interface, process status, rules, integrations and approvals.

Multi-agent architecture is only used when a clear division of roles provides a real advantage. More agents are not in themselves a better solution.

Custom apps don't have to be a chat

The user interface is chosen based on the task and the employee's work situation.

• Internal portals and case summaries
• Form-based workflows
• Document and knowledge tools
• Dashboards and action queues
• Approval views
• Assistants integrated into existing systems

Agent roles and architecture

A solution can consist of an intake agent, a search agent, an agent that prepares a proposal, a control agent and an integration component. Each role can have its own data sources, tools, instructions, rights, context, quality requirements, escalation rules and logging.

Access controls, approvals and final write operations should be enforced by deterministic control logic, not left to the agent’s discretion.

AI Meeting & Note Agent: internal pilot

Engineering Autonomous has an internal, Teams-based pilot with a visible browser guest. The documented solution uses a restricted invitation and approval workflow, camera and microphone off, Teams live captions, explicit start/stop and post-processing to a structured minutes draft.

The pilot is not a finished standard product and cannot be described as guaranteed to be compatible with all Teams environments. Deployment in a customer environment requires a separate pilot, consent procedures, retention rules, access controls and acceptance testing of the minutes.

Project or CRM update, search across meetings and other meeting platforms are possible future extensions, not documented features in the current pilot.

Nine use cases that can be developed and adapted

Only the meeting assistant above has a concrete internal pilot. The others are solution patterns, not finished standard products.

Four levels of action

The agent’s level of agency is defined for each workflow: read, propose, act with approval or act within fixed, reversible limits. A solution does not have to begin with automatic action.

• Data minimisation and restricted sources
• Role-based access and least privilege
• Retention, deletion and audit trail
• Allowed and prohibited tools
• Human approval before critical actions
• Monitoring, fallback and testing against data leaks or wrong actions

From clarification to operation

• Define workflow, users and success criteria
• Clarify data, systems, risk and approvals
• Build a read-only or draft-based pilot
• Test quality, error modes and user flow
• Implement with documented operation, monitoring and change management

Typical questions