AI use-case playbooks
Choose the business workflow first. For each scenario, identify which execution path Talon actually intercepts, trigger an allowed action and a forbidden action, then inspect the evidence. A model's proposed tool call is not proof that the tool executed.
The scenarios below use documented Talon capabilities, but they are not all equally mature demos. Consult the linked case page for the current reproduction status before describing one as a completed integration.
Customer support
Business question: Can the assistant draft a useful response without leaking customer data or issuing an unauthorized refund?
| Workflow | n8n support workflow processes synthetic customer text and prepares a reply |
| Boundary | Governed model calls for data handling and fallback; an intercepted MCP/tool call for a prohibited financial action |
| Allow case | Draft response with customer data appropriately handled |
| Deny case | A forbidden refund tool does not reach the synthetic downstream handler |
| Evidence | PII action, provider/fallback decisions, tool-call denial and signed history |
| Do not infer | Talon itself performs a human refund-approval workflow; or that a live Zendesk account has been tested |
How to evaluate: Begin with customer-support case details and first-line support guide. For the shared control-plane product demo run three-use-case proof. Verify the blocked call separately; a drafted reply is not authority to execute a refund.
n8n refund approvals
Business question: Can an existing workflow request a financial action without deciding its own approval thresholds or dispatching the action directly?
| Workflow | Importable n8n refund workflow with an authenticated webhook caller and a separate Talon use-case credential |
| Boundary | Talon HTTP Action Gateway: exact argument policy, separate human approval and claim-driven dispatch |
| Allow case | €30 action allowed; one Talon dispatch to the deterministic counting adapter |
| Approval case | €75 operation waits for a support-lead reviewer; approval itself dispatches nothing |
| Deny case | €900 action blocked by the company policy; no adapter call |
| Evidence | Effective action policy and rule provenance, approval status, attempt outcome and signed operation lifecycle |
| Do not infer | The counting fixture processes real payments, or MCP tool calls inherit the HTTP Action Gateway approval/claim lifecycle |
How to evaluate: Start with the n8n refund approval walkthrough or business use-case page. Import the reference workflow, run all three inputs and compare the signed action lifecycle with the destination call counter. This is distinct from the customer-support demo above, whose n8n approval step is a separate synthetic operator gate.
Coding agents
Business question: Can an assistant inspect release state and prepare work, but not publish through a forbidden tool?
| Workflow | GitHub Copilot CLI or a coding-assistant use case |
| Boundary | Model requests routed through Talon and MCP release calls traversing its proxy |
| Allow case | Read release_status and prepare a release using allowed tools |
| Deny case | release_publish or a prohibited admin_* tool is blocked on the intercepted path |
| Evidence | Authenticated agent key, asserted session, model/tool timeline, denied call, signature verification |
| Do not infer | Talon watches local file edits or shell commands that bypass it; a benchmark proves savings for every coding agent |
How to evaluate: Follow GitHub Copilot CLI governance for the bounded Copilot/MCP scenario and coding agents demo for a no-provider-key session and budget test. Read coding-agent fleet controls for identity and session limitations. The website case reports the optimization experiment's bounded result rather than claiming a universal cost reduction.
People Analytics
Business question: Can an AI analyst answer aggregate workforce questions while refusing to export individual records?
| Workflow | Synthetic People Analytics requests using aggregated and individual-record tools |
| Boundary | MCP or exact HTTP action path deliberately routed through Talon; existing warehouse/HRIS authorization remains mandatory |
| Allow case | Aggregated headcount or trends within authorized data-access rules |
| Deny case | Attempted employee-record export blocked before the intercepted upstream call |
| Evidence | Tool/action identifier, authenticated use case, effective policy decision and signed denial |
| Do not infer | A production Workday/HRIS connector is shipped or tested; Talon prevents a direct warehouse query that bypasses it |
How to evaluate: Read the People Analytics scenario and current integration status. Use the MCP governance reference or exact HTTP action proof to test an equivalent supported authorization boundary on a mock service. The website scenario is synthetic and its end-to-end integration has been tracked separately; do not label it as a verified production HR deployment.
Vendor contract review
Business question: Can procurement use AI on confidential documents without routing them to an unapproved model or provider?
| Workflow | Synthetic n8n vendor document review |
| Boundary | Provider/model/data-egress policy on traffic routed through Talon's LLM gateway |
| Allow case | Approved provider path handles the sanitized or permitted request |
| Deny case | Unapproved provider path is rejected before egress |
| Evidence | Effective policy snapshot, provider decision, data handling and signed record |
| Do not infer | Complete contract exfiltration prevention for integrations that bypass Talon, or regulatory compliance certification |
How to evaluate: See the vendor-review case, integration path decision and policy cookbook. Test a forbidden provider before testing the approved path, then verify the denial and allow records.
Which demo should I actually run?
| Goal | Best first proof | Prerequisites |
|---|---|---|
| Inspect the simplest governed LLM request | 60-second demo | Docker; no provider key |
| See support, coding and document-summary controls together | Product demo | Local Talon checkout; real paid provider keys |
| Prove exact-action denial, human approval and attempt control | Action proof | Go/test tooling; no provider key; HTTP Action Gateway |
| Run an importable n8n workflow with human approval | n8n refund approvals | n8n, Talon, a test adapter and separated credentials |
| Test a coding session with dual model APIs | Coding-agents demo | Integration test or Docker; no provider key |
| Test a real Copilot MCP release boundary | GitHub Copilot guide | Specific documented demo setup |
For every evaluation, record the path, the test input, the expected upstream effect, the observed effect and the evidence ID. The evaluation checklist gives a reproducible template.