Agent role and operating rules
We define the task, inputs, expected result, stop conditions, escalation route and actions that always require a person.
Service IT / AI agent engineering
AI agents that work inside your business process
We design agents that can keep task context, use approved tools, work with files and business systems, and return a traceable result. When a conversational interface is useful, we embed ChatKit into your website, customer portal or internal application.
What we build
Not just a chatbot, but a controlled digital operator
A useful agent has a specific role, durable task context, limited tools, clear approval boundaries and an observable result. The interface may be a ChatKit conversation, while the same workflow can also start from email, CRM, a webhook or a schedule.
Official OpenAI building blocks
OpenAI DocsChatKit: the embedded interface
A customizable chat surface for messages, widgets, actions, files, progress and tool interaction inside an existing product.
OpenAI DocsAgents API: the managed agent runtimeDurable sessions, orchestration, context management and recovery, with tools, MCP connections, sandboxes and generated artifacts.
We connect these components through a server-side integration that keeps authentication, authorization, business rules and audit under control.
What is included in the service
From a focused pilot to a supported agent workflow integrated with real systems.
Durable sessions and orchestration
We configure Agents API or another suitable runtime so work can continue across turns, interruptions and long-running steps.
Embedded conversational UI
We add an on-brand chat to a website, portal or application with files, widgets, actions, task progress and accessible responsive states.
Tools, functions and MCP
The agent receives narrow typed operations for CRM, ERP, documents, databases, websites and internal or external APIs.
Access, approvals and audit
User isolation, least privilege, human approval for significant actions, idempotency, logs and rollback paths are designed from the start.
Tests, evals and operations
We test normal and failure paths, monitor quality and cost, document the system and support controlled iteration after launch.
Reference architecture
A clear route from user request to verified action
The model never receives unrestricted system access. Every transition is bounded by an interface, a session, a typed tool and a business rule.
Request
A user writes in ChatKit, or an approved email, CRM event, webhook or schedule starts the workflow.
Identity and scope
The backend authenticates the user, selects the workspace and applies the permissions for this exact process.
Agent session
The agent keeps task context, plans the next step and can continue after a pause or new input.
Controlled tool
A function or MCP tool reads or changes only the fields and objects defined in its contract.
Business gate
Validation, duplicate protection and human approval run before a significant action is committed.
Result and trace
The user receives the answer, artifact or status, while technical identifiers and audit events remain available for support.
Anonymized scenario examples
Where an agent can remove handoffs and routine work
These are representative, invented scenarios without company names, people, domains, account identifiers or operational data. The exact scope is designed for each organization.
Service and support agent
Finds the relevant product and order context, searches the approved knowledge base, prepares an answer and asks for confirmation before changing a CRM status or creating a return request.
Incoming-document agent
Reads an attachment, extracts structured fields, checks completeness, creates a draft in ERP and routes exceptions to the responsible employee instead of guessing missing data.
Infrastructure operations agent
Receives an alert, gathers logs and health signals, runs diagnostics in a sandbox, proposes a recovery plan and performs only pre-approved or explicitly confirmed steps.
Analysis and research agent
Uses read-only data tools and selected documents, explains the calculation, produces a reusable report and separates facts from assumptions and missing evidence.
Systems and channels
The agent works where the process already lives
We connect only the sources required for the selected use case and keep development, test and production access separated.
Safety and data
Architecture before autonomy
Before launch we review data routes, retention, residency, permissions and the consequence of every write action. Sensitive cases may require a separate or self-hosted contour.
- Authentication and workspace isolation
- Least-privilege tool scopes
- Human approval for consequential actions
- Input validation and prompt-injection boundaries
- Idempotency and duplicate protection
- Audit events, tracing and incident rollback
- Data minimization and retention review
- AI Act and GDPR assessment when applicable
Deliverables
A maintainable product, not a demo prompt
- Use-case and risk map
- Agent instructions and tool contracts
- Backend integration and authentication
- ChatKit interface when needed
- Connectors, MCP servers and event routes
- Test and evaluation scenarios
- Monitoring and technical documentation
- Deployment, handover and support
How the work starts
Select one process
We choose a repeatable task with an owner, known inputs, a useful output and measurable acceptance criteria.
Build a safe pilot
We test the conversation, tools, permissions and failure cases on controlled data without broad production access.
Integrate and verify
We connect the required systems, add audit and approvals, and verify the complete user-visible route.
Operate and improve
We monitor quality, latency and cost, review failures and extend the agent only when evidence supports it.
Questions before development
Is ChatKit the agent itself?
No. ChatKit is the embedded user interface. The agent logic, sessions, tools, permissions and business rules live in the server-side architecture behind it.
Does every agent need a chat?
No. A workflow may start from email, CRM, a webhook, a file or a schedule. ChatKit is useful when a person needs dialogue, progress, files, review or approval.
Can the agent change data in CRM or ERP?
Yes, when the write action has a narrow contract, access is authorized, validation passes and the selected risk level allows automatic execution or human approval.
Can we use confidential or personal data?
Only after a data-processing review. We verify current provider retention and residency controls, minimize the data and choose a different architecture when the requirements are not compatible.
How long does development take?
It depends mainly on tool count, authentication, data quality, approval logic and test coverage. We estimate the first controlled scope after a short process and risk review.
Next step
Show us one process where people repeatedly search, copy, check and hand work over
We will map the data and approval route, decide whether ChatKit is needed and propose the smallest agent version that produces a verifiable business result.