Service
AI agents and software orchestration
Handling a request can involve reading documents, checking several systems and preparing an action. An AI agent combines a model with software tools to carry out selected steps. The service defines the task, develops the integrations and evaluates the results before planning use in your working environment.
Agents built around specific tasks
Gather the information needed for a decision
Preparing a response or reviewing a case may require both documents and the status of an operation in an application. The agent connects to authorised sources and relevant query interfaces. Its output retains source references and identifies missing information so a person can check the result.
Prepare actions in existing software
A request may lead to creating a task, updating a record or drafting a message. Development covers the tools that perform these operations, their inputs and their access permissions. Actions requiring approval are presented to an authorised person before execution, together with their content and destination.
Evaluate a use case before deployment
A demonstration needs to be tested against the requests and data it will handle in practice. A representative set of cases makes it possible to assess result quality, errors, response time and execution cost. These observations inform the selected tasks, review steps and conditions for deployment.
The proposed engagement
Scoping establishes the requests to handle, the sources and the possible actions. Development connects the model to the required tools, with software rules for permissions, approvals and execution tracking. Tests on representative cases guide adjustments to instructions and integrations before defining how the service will be used.
An example scope
To help handle customer requests, an initial agent could consult a CRM record and authorised documentation, then propose a response and a follow-up task. The responsible person would review the proposal before the message is sent or the task is created.
An agreed scopeThe scope defines the operations to build, the required approvals and the results expected during evaluation.
What the delivery includes
- Task definition, permitted sources and actions, and acceptance criteria.
- Application tools, step orchestration and approval controls integrated into the user workflow.
- An evaluation set and documented results for the selected model and instruction versions.
- A deployable service, action logs, error monitoring and procedures for recovery or human intervention.
Integration with your environment
- Available databases, documents and APIs determine what the service can read and which operations it can perform.
- Application identities and permissions define the access granted to the service.
- A local model or a model accessed through an API is selected against the data involved, measured results, cost and response time.
Service lineArtificial intelligence →
Plan the first stage
Start with a bounded task, representative requests and the systems used to handle them. Scoping distinguishes steps that benefit from model analysis from those governed by defined rules. A first delivery can include a prototype connected to selected sources and a documented evaluation, providing observable results on which to base the next decision.
Related projects
The related reference concerns agents used in a software-development environment, with code access, tools and verification of changes. Its scope is the development work described in that project.
To go deeper
Four questions about the EU AI Act for a Swiss company
The company’s role, its EU connection and the system’s purpose determine the assessment. Those conclusions must lead to design decisions the team can verify.
Choosing agent memory and managing its lifecycle
Conversation history, reference documents and persistent memory serve different purposes. Their value depends on provenance, permissions and the ability to correct information.
Four AI agent workflows and their acceptance criteria
An agent connects information analysis with an organisation’s tools. Four examples show the integrations, expected results and checks needed to evaluate that work.
Discuss your project
A discussion of your project’s design, development and production deployment.