Appendix A

Discovery question bank

Use these questions to prepare the workshops in Appendix B, picking the ones the data request has not already answered. The security questions by layer are in Section 4. Several questions draw on the review frameworks in Section 5: the AWS Well-Architected Agentic AI Lens, Microsoft’s Cloud Adoption Framework and Google’s 20 questions for the agentic enterprise.^14,47,76

Goals and value

• What are the business goals and pressures for the next 12 to 24 months?
• Which process, if it ran twice as fast with half the errors, would matter most?
• What starts the process, and what is true when it is done?
• Who owns the outcome, and who would sponsor the change?
• What does a delay or an error cost today, in hours, revenue, risk or client experience?
• What is out of scope for the first phase?
• How will finance want to see the value measured, and who signs it off?

Process and systems

• Can you walk through a recent case that went well and one that went badly?
• Where is information re-keyed, chased or checked by hand?
• Which systems of record hold the data, and who owns each?
• Which interfaces exist today (APIs, events, files), and which are missing?
• Which steps should stay human-led, which could be AI-assisted and which AI-led, and why?

AI estate

• Which agents and copilots are in production, pilot and planning, on which platforms, and who owns each?
• How do agents reach models and tools today: directly, through a gateway or through MCP servers?
• Which models and providers are in use, under which terms (data retention, regions)? What did AI cost last quarter, by use case?
• What share of AI spend can you attribute to an owner and a use case today? Can your providers export billing data in the FinOps FOCUS format?
• Is every agent registered somewhere with an owner and a risk tier?
• How would you find an agent that someone built without telling IT?
• How are agents evaluated before they go live, and monitored afterwards? Do you evaluate the tools an agent chose, or only its final answer?

Data and master data

• Which master data domains does the process depend on, and is there a golden record for each?
• What are the known quality issues: duplicates, missing identifiers, stale hierarchies?
• Is there a glossary, lineage and classification for the data involved?
• Which data is sensitive, and which policies apply (masking, information barriers, residency)?

Security and compliance

• How do people sign in, and which actions need step-up authentication?
• Does each agent have its own identity? Can agents act on behalf of a person, and with what scope?
• Which approvals are required today, and how are they evidenced?
• Which regulations and audit findings apply, and what evidence must the process produce?
• How would you stop an agent quickly, and has anyone tried?
• If you switched off your most important agent today, who would take over its work, and with what information?
• Does any agent use a static API key or a shared service account?
• Which agentic threats (the OWASP agentic top 10) has your red team tested?
• What are the rules for data leaving your environment, including to model providers?

Architecture and operations

• Which runtimes are approved: managed cloud, customer-managed or on premises?
• Which integration and API management platforms are in place?
• How are services observed today: logs, traces and correlation across systems?
• What are the availability and recovery requirements?
• Which standards (API-led connectivity, MCP, A2A) are enterprise policy today, and which MCP and A2A versions do your platforms and gateways support?

Governance and lifecycle

• Who sponsors each agent, and what happens to an agent when its sponsor changes roles?
• How often is each agent recertified, and how is an agent retired?
• Which agent control planes are already in use, such as Agent 365, Amazon Bedrock AgentCore, Gemini Enterprise Agent Platform or Agentforce, and which is authoritative for inventory, identity and policy?
• Does each agent’s access to data match the access of the user it acts for?
• Which review framework does architecture use (AWS Well-Architected, Microsoft’s Cloud Adoption Framework, Google’s 20 questions), and who runs the review?
• Which standards does compliance report against, such as ISO/IEC 42001, the NIST AI RMF or the EU AI Act?
• Who owns AI skills and training, and which roles will change?

Phasing and decisions

• Which slice could go live in 90 days with value people can see?
• Which volumes, rates and costs should the baseline use, and who signs them off?
• What would make you pause or stop the program?
• Who reports value, and at which reviews? How will you measure value, rather than usage?
• What decisions do you need to take, and who takes them?