Hackathon Exercise: Governance Review and Go-Live
Welcome to the final stage of your hackathon. You have designed, built, and tested an AI agent in Microsoft Copilot Studio (Microsoft’s low-code platform for designing, grounding, testing, and governing custom AI agents with restricted knowledge sources and action boundaries.) using synthetic financial data. Before any new technology system goes live in a UK financial services firm, it must pass a rigorous governance review. This lesson simulates that critical milestone. You will prepare a presentation for a mock AI Governance Board to prove your agent is safe, compliant, and ready for publication to your colleagues.
Refresher Core Mental Model: Hackathon Exercise: Governance Review and Go-Live 1 min
In Copilot Studio and AI Agents for Non-Developers: Building Safe Agents in Regulated Financial Services, Hackathon Exercise: Governance Review and Go-Live focuses on turning The Mock Governance Board Scenario into a repeatable, governed practice. Rather than treating tools as black-box oracles, anchor every step to three questions: What approved data goes in? What structured draft comes out? And who verifies the result against primary evidence before action is taken?
In the highly regulated UK financial sector, building a functional tool is only half the task. You must also prove that the tool operates within strict regulatory boundaries. The governance board will look for evidence that your agent aligns with the Financial Conduct Authority (FCA) Consumer Duty (The UK Financial Conduct Authority (FCA) regulatory standard requiring firms to act to deliver demonstrably good outcomes for retail customers across products and support.) by delivering reliable outcomes, even if it is an internal staff-facing tool. They will require proof that it respects UK General Data Protection Regulation (UK GDPR) principles by protecting personal data, and that it adheres to operational resilience guidelines through robust error handling and audit trails (Tamper-evident chronological records capturing who triggered, reviewed, approved, or modified an automated workflow step and when.).
The Mock Governance Board Scenario
For this exercise, imagine you are presenting your completed synthetic data (Artificially created test records that mirror production data structure without containing any real customer PII or confidential business information.) agent to your firm’s internal AI Governance Committee. This committee includes stakeholders from compliance, risk management, IT security, and business leadership. You must convince them that your agent is ready for production.

To pass this review, you cannot simply show that the agent answers questions correctly. You must demonstrate exactly how the agent is governed. The board expects clear answers on human ownership, data protection, security permissions, auditability, and the rigour of your testing phase. If you cannot evidence these controls, the agent will not be approved for release.
Proving Named Human Ownership and Accountability
Under the Senior Managers and Certification Regime (The UK financial regulatory framework establishing personal, non-delegable accountability for senior managers over the operational processes and automations in their area.) (SM&CR), financial firms must maintain clear lines of accountability for all business processes. AI agents are no exception. An agent cannot exist in a vacuum without a named individual responsible for its actions.
You must document exactly who owns this agent. In Microsoft Copilot Studio, the person who creates the agent is logged as the maker. However, business ownership goes beyond the person who built it. Your governance pitch must name the specific business lead who is accountable for reviewing the agent’s performance, updating its knowledge sources when policies change, and approving any modifications to its instructions.
In your presentation, you must confirm that this named owner will actively monitor the agent. You must also detail the process for transferring ownership if the current owner leaves the firm or changes roles, ensuring the agent never operates without human oversight.
Demonstrating Data Protection and Least-Privilege
This is arguably the most critical section of your governance review. You must prove that your agent cannot overreach its intended purpose or expose sensitive information. You will demonstrate this through the principle of least-privilege permissions.
When your agent uses SharePoint as a knowledge source, whether you built it in Copilot Studio or in Agent Builder in Microsoft Copilot, it respects the signed-in user’s existing permissions. If an employee using the agent does not have permission to view a secure SharePoint folder containing confidential merger documents, the agent cannot access or summarise those documents for them. You must explain this security model to the board, showing that the agent only retrieves SharePoint data the user is already permitted to see, and noting the exceptions: files uploaded directly to the agent, and tools that run with maker-provided credentials, are available to everyone who can use the agent.
Furthermore, you must explain how your agent is restricted by data policies (formerly data loss prevention or DLP policies) configured in the Power Platform admin center. These policies, set by your IT administrators, can prevent makers from connecting agents to high-risk external websites or unapproved third-party databases, and Copilot Studio enforces them in real time with no exemptions for individual agents. You must state clearly that you have adhered to these policies, that no real customer data was used during development, and that your agent relies entirely on approved, synthetic datasets or public regulatory documents for its knowledge.
| Governance Requirement | Poor Presentation Example | Strong, Compliant Presentation Example |
|---|---|---|
| Accountability | The IT department owns the agent because they gave us the Copilot Studio licences. | Jane Doe, Head of Retail Credit, is the named business owner. She is accountable for the agent outputs and reviews its performance monthly. |
| Data Protection | We uploaded a spreadsheet of recent customer complaints to see if the agent could summarise them. | We used only synthetic, artificially generated data for testing. Real customer data is protected and never uploaded to the test environment. |
| Least-Privilege | The agent can search the whole company network to find the best answers. | The agent uses each user’s own Microsoft 365 permissions. It can only surface information from SharePoint sites the specific user already has permission to read. |
| Testing Evidence | We asked the agent a few questions and it gave good answers, so it is ready. | We executed 50 test cases, including prompt injection attempts and out-of-scope questions, to show the agent fails safely and declines rather than inventing answers. |
Plain-text version
Governance Requirement | Poor Presentation Example | Strong, Compliant Presentation Example Accountability | The IT department owns the agent because they gave us the Copilot Studio licences. | Jane Doe, Head of Retail Credit, is the named business owner. She is accountable for the agent outputs and reviews its performance monthly. Data Protection | We uploaded a spreadsheet of recent customer complaints to see if the agent could summarise them. | We used only synthetic, artificially generated data for testing. Real customer data is protected and never uploaded to the test environment. Least-Privilege | The agent can search the whole company network to find the best answers. | The agent uses each user's own Microsoft 365 permissions. It can only surface information from SharePoint sites the specific user already has permission to read. Testing Evidence | We asked the agent a few questions and it gave good answers, so it is ready. | We executed 50 test cases, including prompt injection attempts and out-of-scope questions, to show the agent fails safely and declines rather than inventing answers.
Evidencing Comprehensive Logging and Audit
FCA rules (SYSC 9) require firms to keep orderly records of their business that are sufficient for the FCA to monitor compliance. Your governance board will want to know that your agent does not operate in a black box. If an employee makes a poor financial decision based on information provided by the agent, the firm must be able to audit exactly what the agent said and what knowledge it relied upon.
You must prove that comprehensive logging is in place. In Microsoft Copilot Studio, you have access to built-in analytics and the ability to download conversation transcripts. Explain to the board how these tools allow the named business owner to monitor interactions and audit the agent decisions. You must highlight how this audit trail helps identify silent failures, which occur when the agent provides a plausible but factually incorrect answer without triggering an error message.
Confirm to the board that conversation transcripts will be reviewed on a regular schedule, and that these records will be retained in accordance with your firm’s corporate record-keeping policies, which usually means extending the default 30-day transcript retention.
Validating Pre-Release Testing
No software system should transition to a live environment without passing exhaustive testing. You must present a summary of your testing phase to the governance board. Detail how you tested the agent using exclusively synthetic data that mimics realistic financial scenarios.
You must prove that you actively tested for AI-specific risks. Explain how you conducted prompt injection testing, where you deliberately tried to trick the agent into ignoring its core instructions or revealing system prompts. Show the board how the agent successfully resisted these attempts.
Additionally, present your checks for hallucination. Show how you intentionally asked the agent questions completely unrelated to its designated topic (such as asking a mortgage policy agent for personal investment advice) to prove that it responds with a safe, polite refusal rather than inventing an answer. Demonstrate that when the agent encounters a complex vulnerability scenario, it follows instructions to escalate the query to a human-in-the-loop rather than attempting to resolve it autonomously.
Hackathon Instructions: Prepare Your Pitch
It is now time to prepare your final governance pitch. Work with your hackathon team to compile your evidence and structure your presentation. You will need to address the mock governance board (played by your instructors or peers) and defend your agent’s readiness.
- Step 1: Document Ownership. Write a formal statement naming the business owner of the agent. Detail their responsibilities, including how often they will review Copilot Studio analytics and transcript logs. Outline the process for updating the agent instructions if underlying financial regulations change.
- Step 2: Map the Permissions. Create a simple map showing exactly what knowledge sources your agent connects to. Explicitly state that all data is synthetic or public. Write a paragraph explaining how Microsoft 365 permissions support least privilege, showing that users cannot bypass their standard security permissions by using the agent.
- Step 3: Export Test Evidence. Summarise your test log. You must present at least five documented test cases. Include one successful standard query, one successfully deflected prompt injection attempt, one out-of-scope query that the agent declined rather than inventing an answer, and one complex scenario that triggered a mandatory human-in-the-loop handover.
- Step 4: Draft the Go-Live Request. Combine your ownership documentation, permission mapping, and test evidence into a concise go-live request. Practice delivering this request aloud. Focus on plain language and clear regulatory alignment. Be prepared to answer challenging questions from the board regarding data leakage and silent failures.
By completing this exercise, you demonstrate that you understand the fundamental difference between simply building an AI tool and responsibly deploying an AI agent within a regulated financial institution. Governance is not a barrier to innovation; it is the framework that makes safe innovation possible.

Keep going with Tesseract Academy
Get personalised feedback from the Tesseract Academy AI Coach
Paste your work below and the AI Coach will review it against a rubric and suggest concrete improvements.
Two tools to take this further:
- Tesseract Academy AI Coach: a personal AI coach that reviews your work, runs practice scenarios and helps you apply what you learned.
- LearnGen: build your own AI-powered courses and training with Tesseract Academy’s course engine.
Go deeper Architectural & Governance Trade-Offs in Hackathon Exercise: Governance Review and Go-Live 1 min · optional
When scaling Proving Named Human Ownership and Accountability across a team or regulated workflow, the primary bottleneck is rarely generation speed—it is verification overhead and drift. High-performing teams track two operational metrics alongside throughput: Edit-to-Approval Time (how many minutes a human reviewer spends correcting outputs) and Source-Traceability Rate (whether 100% of material numbers and claims link back to an audited record). If edit time exceeds manual drafting baseline across two cycles, tighten your input schema and negative constraints before expanding scope.
AI and Data Science for Decision Makers
Equip leadership and operations teams to evaluate, scope, and govern production AI initiatives.
