CopilotForgeDraft

Templates › Microsoft Foundry solution design

Microsoft Foundry solution design

Template id: foundry-solution-design

Responsible skill: ai-foundry:foundry-solution-assessment. Sections are detailed with ai-foundry:model-selection, ai-foundry:foundry-agent-architecture, ai-foundry:foundry-tool-design, ai-foundry:mcp-integration, ai-foundry:foundry-grounding-strategy, ai-foundry:retrieval-design, ai-foundry:foundry-prompt-engineering, ai-foundry:model-evaluation and ai-foundry:foundry-safety-evaluation.

Use this design from the moment a need reaches the Foundry rung until the solution is released, and keep it current after every material change. Every fact the organisation must supply is marked [To confirm: ...]; replace the marker when the fact is confirmed, and never fill it with a guess. Example rows are fictional and are there to be replaced or deleted. Where a design choice depends on a product feature, model, region, limit or price, record the source record checked and the date, because these change often. Licensing questions go to the Microsoft licensing guide or a licensing partner; data protection questions go to the DPO.

1. Summary

FieldEntry
Solution name[To confirm: working name]
Problem statement and suitability record reference[To confirm: reference and date]
Business owner (role)[To confirm]
Product owner (role)[To confirm]
Engineering team and technical owner[To confirm]
Operations and support route[To confirm]
Risk tier (0-4) and date classified[To confirm]
Azure subscription, landing zone and Foundry project[To confirm]
Design statusDraft / Agreed / Reviewed / Superseded
Version and date[To confirm]

2. Assessment

2.1 Capability drivers

DriverApplies?Evidence
Model choice beyond what configured agents exposeYes / No[To confirm]
Retrieval engineeringYes / No[To confirm]
Complex orchestrationYes / No[To confirm]
AI embedded in a product, API or batch processYes / No[To confirm]
Production-grade evaluation and observabilityYes / No[To confirm]

2.2 Lower rungs tested

Option tried (Microsoft 365 agent, Copilot Studio agent, automation, hybrid)What was triedWhy it fell shortSource record checked and date
Fictional example: Copilot Studio agent with SharePoint knowledgeTwo-week spike on the claims libraryNeeded metadata filters by effective date that configured knowledge could not applyFictional example only
[To confirm][To confirm][To confirm][To confirm]

2.3 Ownership conditions

ConditionStatus (met / not met / partly)Evidence or action
Product owner and engineering team named[To confirm][To confirm]
Azure foundation in place[To confirm][To confirm]
Running cost funded beyond the first year[To confirm][To confirm]
Operations route agreed[To confirm][To confirm]
Reviewer capacity available at this risk tier[To confirm][To confirm]

2.4 Decision

FieldEntry
Decision (proceed / proceed with preconditions / use a lower rung / stop)[To confirm]
Preconditions and owners[To confirm]
What this assessment cannot determine[To confirm]
Decision owner and date[To confirm]

3. Model

FieldEntry
Model selection record reference (model-selection-record)[To confirm]
Chosen model and version[To confirm]
Fallback model and version[To confirm]
Lifecycle review date and owner[To confirm]

4. Agent architecture

Complete if the solution includes one or more agents.

FieldEntry
Responsibility statement (one sentence)[To confirm]
What the agent must refuse or hand over, and to whom[To confirm]
Shape (single agent / coordinating agent with specialists) and reason[To confirm]
Conversation state and cross-session memory, with retention agreed[To confirm]
Launch surface[To confirm]
Architecture diagram reference[To confirm]
ActionIdentity it runs under (user-delegated / managed identity)Reason if not user-delegatedHuman confirmation
[To confirm][To confirm][To confirm]Yes / No

5. Tool register

One row per function or tool the organisation hosts. MCP servers are recorded in section 6.

ToolPurpose and model-facing descriptionEffect class (read low / read sensitive / reversible write / irreversible)IdentityControls (limits, idempotency, confirmation, audit)Failure behaviourSystem owner agreedTest evidence
Fictional example: get_policy_sectionReturns one section of a named policy by identifierRead lowUser-delegatedResult size limitStructured "not found"Yes, fictionalFictional example only
[To confirm][To confirm][To confirm][To confirm][To confirm][To confirm]Yes / No[To confirm]

6. MCP server register

ServerOwner and hostingTerms and data handling reviewed byAllowed toolsEffect classIdentity and authenticationApproval ruleVersion pin or tool-list fingerprintChange check rhythm
[To confirm, or "none"][To confirm][To confirm][To confirm][To confirm][To confirm][To confirm][To confirm][To confirm]

7. Grounding and retrieval

FieldEntry
Grounding approach and reason[To confirm]
Content left in place rather than indexed, and why[To confirm]
SourceOwnerReadiness verdictClassificationPermission trimming (per user / open to all users, confirmed by owner)Refresh and deletion route
[To confirm][To confirm][To confirm][To confirm][To confirm][To confirm]
Retrieval design itemEntry
Cleaning rules[To confirm]
Chunking approach[To confirm]
Metadata kept on every chunk[To confirm]
Search method and ranking[To confirm]
Embedding model and version[To confirm]
Passages passed to the model and citation format[To confirm]
Retrieval test set reference and baseline[To confirm]

8. Prompts and output contract

FieldEntry
Prompt location in source control and current version[To confirm]
Prompt owner (role)[To confirm]
Output schema and version, and consuming systems[To confirm, or "free text only"]
How untrusted content is delimited[To confirm]
Evidence required per change type[To confirm]

9. Evaluation and safety

FieldEntry
Evaluation plan reference (evaluation-plan)[To confirm]
Quality thresholds and date agreed[To confirm]
Harm profile summary[To confirm]
Safety controlSettingReasonBehaviour when it firesLast changed, by whom, with what evidence
Input and output content filtering[To confirm][To confirm][To confirm][To confirm]
Prompt attack detection[To confirm][To confirm][To confirm][To confirm]
Groundedness or protected material checks[To confirm, or not used with reason][To confirm][To confirm][To confirm]
Blocklists[To confirm, or "none"][To confirm][To confirm][To confirm]

Untested harms and residual risks: [To confirm]

10. Deployment, observability and cost

FieldEntry
Environments and deployment topology[To confirm]
Tracing and monitoring owner[To confirm]
Running cost estimate, budget owner and source records used[To confirm]
Pipeline and release process reference[To confirm]

11. Reviewer gates

ReviewerRequired from tierFinding summaryBlocking / advisoryResolved?
Design authority[To confirm][To confirm][To confirm]Yes / No
Security and privacy[To confirm][To confirm][To confirm]Yes / No
Responsible AI[To confirm][To confirm][To confirm]Yes / No
Evaluation and quality[To confirm][To confirm][To confirm]Yes / No

12. Open questions and preconditions

ItemOwnerDue
[To confirm][To confirm][To confirm]

13. Agreement

RoleNameDate
Business owner[To confirm][To confirm]
Product owner[To confirm][To confirm]
Technical owner[To confirm][To confirm]

Credit: Template from CopilotForge by Simon Foote (https://github.com/SimonFoote-greensparx/CopilotForge), licensed under CC BY 4.0. Keep this credit when you share or adapt the template; in your own completed document you may keep it as a footer or remove it.

Download the template (Markdown)