AB-100 Deploy AI solutions explained
Deploy AI-powered business solutions is the largest AB-100 domain at 40–45%, and it is mostly about running AI responsibly after it is built. It has four parts: analysing, monitoring and tuning; managing testing; designing the ALM process; and designing responsible AI, security, governance, risk management and compliance. Questions ask what to measure, what to gate, how to promote changes and how to protect data.
Analyse, monitor and tune
The outline asks you to recommend the process and tools for monitoring agents, analyse the backlog and user feedback, use AI-based tools to find issues and tune, monitor performance and metrics, and interpret telemetry for performance and model tuning.
| Question | Source |
|---|---|
| Is the agent being used? | Usage and adoption metrics |
| Is it answering well? | Quality signals, user feedback, escalation rates |
| Is it fast and reliable? | Latency, failures, telemetry |
| What should change? | Backlog of unresolved questions and feedback |
Tune from evidence. If the telemetry shows retrieval failing on a topic, fix the knowledge source before touching the model.
Manage testing
Test agents with defined metrics rather than impressions. The outline names metrics and process for testing agents, validation criteria for custom AI models, best practice for Copilot prompts, end-to-end scenarios across several Dynamics 365 apps, and creating test cases with Copilot.
Because model output varies, a good test set has fixed inputs, expected behaviours and pass criteria, and is re-run after every change to prompts, knowledge or model. End-to-end tests matter where an agent hands work between apps, because the failure usually sits in the handoff.
Design the ALM process
The outline lists six objects that need their own lifecycle:
- Data used in AI models and agents
- Copilot Studio agents, connectors and actions
- Microsoft Foundry Agents Service
- Custom AI models
- AI in Dynamics 365 finance and supply chain apps
- AI in Dynamics 365 customer experience and service apps
The shared principles are: separate development, test and production environments; move changes as versioned packages rather than editing production; test before promotion; keep configuration such as connections and endpoints separate from the solution; and be able to roll back. For models and data, add versioning and record which model version and data produced which behaviour.
Responsible AI, security, governance, risk and compliance
This part is long and practical. Group it as follows:
- Agent security and governance. Who can create, share and change agents; what each agent may access; approvals before publishing.
- Model security. Protecting model endpoints, tuning data and deployments.
- Vulnerabilities and prompt manipulation. Analyse attacks where input tries to override instructions or extract data, and design mitigations such as input and output filtering, least-privilege tools and not trusting retrieved content as instructions.
- Responsible AI review. Check the solution against Microsoft’s responsible AI principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability.
- Data residency and movement. Confirm where data is processed and stored and whether it crosses boundaries.
- Access controls on grounding data and tuning. An agent should retrieve only what the requesting user may see.
- Audit trails. Record changes to models and data so you can answer who changed what, and when.
Example questions
Question 1. Telemetry for a service agent shows that unresolved questions cluster around one product line, and escalations to humans are high for it. What is the most appropriate tuning step?
- A. Retrain a custom model on all products
- B. Disable the agent for all products
- C. Make escalation to humans harder
- D. Improve the knowledge content and retrieval for that product line
Show answer
Answer: D
The evidence points to missing or weak grounding for that product line, so improving its knowledge content or retrieval is the targeted fix. Retraining the whole model is heavy and unevidenced. Disabling the agent removes value and leaves the cause. Raising the escalation threshold hides the problem rather than fixing it.
Want more questions like this? Full AB-100 practice tests →
Question 2. A team edits a production Copilot Studio agent directly to fix urgent wording issues. Two edits have already broken a connector reference. What should the ALM process require?
- A. Test all changes only in production
- B. Make changes in development and promote tested, versioned packages to production
- C. Disable all connectors in production
- D. Document each production edit after it is made
Show answer
Answer: B
Changes should be made in a development environment, tested, and promoted to production as a versioned package, which prevents direct edits breaking connections. Restricting testing to production is the opposite of safe. Disabling all connectors removes needed function. Documenting edits afterwards records problems without preventing them.
Want more questions like this? Full AB-100 practice tests →
Question 3. A legal assistant agent retrieves from a shared document library. A user without access to the merger folder asks about the merger and receives details from it. Which design change addresses the cause?
- A. Add a disclaimer to every answer
- B. Use a larger model
- C. Enforce the user’s permissions on the grounding data at retrieval time
- D. Delete the conversation logs
Show answer
Answer: C
Access to grounding data should follow the requesting user’s permissions, so the agent only retrieves what that person may see. Adding a disclaimer does not stop disclosure. Moving to a larger model does not change access. Deleting conversation logs removes evidence rather than the exposure.
Want more questions like this? Full AB-100 practice tests →
How to revise this domain
Build a checklist: monitor, test, promote, secure, audit. For each agent you design, state what you would measure, which test set gates release, how a change moves to production, and what is logged. Then try the practice test, which puts nine of its twenty questions here.