Outcome

An implementation-ready definition of who will use the system, what it must know, what it may access, how answers and actions will be evaluated, and what the first pilot must prove.

Typical scope

• An employee group and an illustrative set of priority questions or tasks
• Existing chatbot or Copilot implementation issues
• Knowledge-source inventory, ownership, and freshness
• RAG, citation, interface, authentication, and permission requirements
• Evaluation dataset, security requirements, and pilot acceptance criteria
• Implementation schedule and final pilot proposal

Example scope

An example starting scope covers one primary use case and employee group, selected knowledge sources, stakeholder interviews, a pilot recommendation, a findings presentation, and a revision. We agree the source count, interview plan, and deliverables together. Production development, data remediation, custom integrations, security certification, and third-party fees are scoped separately.

Outcome

A complete architecture for an operational agent or multi-system assistant, including its knowledge, tools, authority, security, evaluations, deployment model, and implementation milestones.

Typical scope

• RAG architecture and live API or database access
• MCP tool definitions and input-output schemas
• User, department, organization, and agent identity
• Read, draft, approval, and transaction authority
• Audit, monitoring, failure, and recovery behavior
• Model policy, stage gates, hosting, support, and milestone proposal

Example scope

A possible scope covers selected employee groups and connected systems, stakeholder discovery, one agent or workflow family, an architecture readout, and a revision. Together we set the boundaries and deliverables. Production development, source-system changes, data remediation, penetration testing, and external provider costs are scoped separately.

Employee experience

• Web-based employee chat and conversation history
• Source citations and supporting-document view
• Unsupported-answer refusal and user feedback
• Controlled pilot deployment

Governed system

• Permission-aware retrieval and administrative controls
• Evaluation dataset and answer-quality testing
• Usage analytics
• Knowledge freshness and synchronization

Examples & boundary

Product and technical support, SOP and policy assistance, employee training, sales support, purchasing and supplier-document assistance, claims and returns guidance, and internal company knowledge.

The assistant answers and explains. It does not automatically change operational systems.

Capabilities

• RAG retrieval
• Live API or governed database reads
• Fixed MCP or tool registry
• Authentication and role scoping
• Tool-call audit history

Operational controls

• Tenant, organization, user, and role boundaries
• Current-state responses
• Evaluation and monitoring
• Bounded calls, runtime, tokens, and cost

Examples & boundary

Retrieve current inventory, order or shipment status, open purchase orders, product comparisons, and customer or supplier history—while combining policy documents with current facts.

No writes, destructive actions, or changes to authoritative systems.

Workflow

• Draft creation and action preview
• Human approval and role checks
• Material-change summaries
• Reapproval after changed price, quantity, dates, or substitutions

Controls

• Duplicate-action prevention
• Complete audit history
• Read-back after approved actions
• Failure and recovery handling

Examples & boundary

Draft customer responses, quotes, purchase orders, product records, catalog corrections, claim or credit requests, substitutions, and exception cases.

Each write or approval capability requires a separately reviewed and validated capability profile; it is not assumed to be part of the current standard Forge profile.

Authority controls

• Fine-grained authorization and explicit action policies
• Idempotency, replay protection, and duplicate prevention
• Security testing and independent QA

Operational proof

• Transaction reconciliation and read-back verification
• Rollback or compensating procedures
• Detailed audit history and written acceptance tests
• Incident and recovery procedures

Examples & boundary

Submit or update an order, post a shipment notice, update a product or account record, create a refund or credit, or change an allocation.

These are not generic chatbot projects. They are production transaction and workflow systems with an AI interface, and they require dedicated architecture, security, testing, and operating controls.

Possible agents

• Product-information, sales-support, and customer-service agents
• Purchasing, catalog, and receiving-discrepancy agents
• Training and executive-reporting agents

Shared foundation

• Identity, tenant isolation, knowledge services, and tool registry
• Permission model, approval engine, routing, and agent handoffs
• Monitoring, evaluation, administration, and emergency disabling

Boundary

Multi-agent orchestration is separately designed and validated. It is not the current default Forge generation profile.