AIGP Quest · cheat sheet · BoK v2.1

IV.A Deciding to deploy

IV.A Deciding to deploy (6 to 8 exam questions)

Evaluate key factors and risks relevant to the decision to deploy the AI system.

IV.A.1 · Is AI even the right tool for this job?

Domain IV takes the deployer's view: the organization putting AI to work, bought or built.
Five context checks: business objectives, performance requirements, data availability, ethics and workforce readiness.
Set performance requirements for your own context; vendor benchmarks aren't your requirements.
Data availability includes the legal right to use the data, not just having it.
Workforce readiness means training, role design and calibrated trust for everyone affected, not just IT; organizational readiness adds budget, data and security maturity, and risk tolerance.

Deployer:: The organization that uses an AI system under its own authority, whether it bought the system or built it.
Use-case context:: The business objectives, performance requirements, data availability, ethical considerations and workforce readiness that decide whether an AI deployment makes sense.
Performance requirements:: The accuracy, speed, explainability, availability and error tolerance a specific use demands, set by the deployer for its own context.
Data availability:: Whether suitable, good-quality data exists and the organization has the right to use it for the intended purpose.
Workforce readiness:: Whether the people who will use, oversee or be affected by an AI system have the skills, training, roles and trust to work with it well.
Organizational readiness:: Whether the organization has the leadership, budget, data governance, security and privacy maturity, risk tolerance and operating capacity to support an AI deployment.

Exam tips: When a stem lists facts a confident sponsor is ignoring (unlabeled data, untrained staff, no success metric), those ignored facts usually point to the answer. · FIRST questions about a new deployment usually reward defining the use case and its context before pilots, vendors or models. · Benchmarks answer only the performance question, and only on the vendor's data. They never settle a deploy decision on their own. · 'Don't use AI here' can be the right answer. Take options seriously when they question whether AI fits at all.

IV.A.2 · Model menu: four ways to tell models apart

Four axes: classic vs. generative, proprietary vs. open, small vs. large, language vs. multimodal.
Generative models add confabulation, IP and harmful-content risks.
Open weights give control and transparency but put hosting, security and safety on you, under the license's terms.
Small models are easier to test, explain and host privately; large ones are broader but harder to evaluate.
Multimodal capability adds new attack channels and sensitive-data risks.

Classic (predictive) AI:: Models that classify, score or forecast from inputs rather than create new content.
Generative AI:: Models that create new content such as text, images, code or audio.
Proprietary (closed) model:: A model whose code and weights are not published; you use it under license or through an API.
Open-weights model:: A model whose trained parameters are published for download, usually under a license that may still carry conditions.
Multimodal model:: A model that processes more than one type of data, such as text together with images, audio or video.

Exam tips: Model-type questions often describe a model in one sentence and flip one axis in each wrong option. Check all four axes. · 'Open' never means 'no conditions'. Look for license limits on commercial use, attribution and acceptable use. · Bigger is rarely the governance answer. Labeled history plus fixed categories points to a classic model. · Closed models shift evidence-gathering to vendor diligence; they never shift accountability off the deployer.

IV.A.3 · Where should it run? Cloud, on-prem or edge

Cloud = scale and convenience, with data leaving your walls and shared responsibility.
On-premise = maximum control and inspectability, and all the security work is yours.
Edge = low latency, offline operation and local data, but hard fleet updates and monitoring.
Hybrid mixes environments (for example, edge inference with cloud training): more fit, more integration work.
Choose by data sensitivity, control needs and connectivity, not by hype or headline cost.

Cloud deployment:: Running an AI system on a provider's remote infrastructure, typically as a managed service.
On-premise deployment:: Running an AI system on servers the organization owns and operates itself.
Edge deployment:: Running an AI model directly on an end device such as a phone, camera, vehicle or sensor.
Shared responsibility model:: The split of security and compliance duties between a cloud provider (infrastructure) and its customer (data, access and use).
Data residency:: A requirement that data be stored or processed within a particular country or region.
Hybrid deployment:: Combining environments, such as edge or on-premise inference with cloud training and updates, so each workload runs where it fits.

Exam tips: When several options satisfy the loud constraint (such as residency), the answer turns on the quiet one: offline operation, physical control or auditability. · Residency is not control. An in-country cloud region keeps data local, but the provider still operates the hardware. · Moving to the cloud moves the work, never the accountability. · Edge is the answer when latency, offline operation or on-device privacy is a hard requirement. Expect follow-ups about updating and rolling back a fleet, and about slimming models (quantization, distillation) at some cost in accuracy.

IV.A.4 · Make it fit: prompts, RAG, fine-tuning, agents

As is → prompting → RAG → fine-tuning → agents: each adds fit and power, and shifts responsibility to you.
RAG keeps knowledge fresh and citable, but the document store needs access control, quality checks and injection defenses.
Fine-tuning changes the model: govern the training data, re-test safety, and check whether you've become a provider.
Agents need least privilege, human approval for high-impact actions, action logs and a way to stop and roll back.

Prompt engineering:: Shaping a model's behavior through the instructions and examples sent with each request, without changing the model.
Retrieval-augmented generation (RAG):: A technique where the system retrieves relevant documents at query time and adds them to the prompt, leaving the model's weights unchanged.
Fine-tuning:: Further training a pre-trained model on your own examples, which permanently changes its weights.
Agentic architecture:: A design in which a model plans steps and calls tools to take actions toward a goal with limited human input.
Excessive agency:: OWASP's term for an AI system with more functionality, permissions or autonomy than its task needs.
Indirect prompt injection:: Malicious instructions hidden in content the model reads, such as a retrieved document or web page, that hijack its behavior.

Exam tips: Changing content + citations + no training budget = RAG. Stable format + huge volume + cost pressure = fine-tuning. · Upgrading from answers to actions raises permissions, logging, human approval and rollback. Training-data risks are unchanged by that upgrade. · Fine-tuning or repurposing can make a deployer the provider under EU AI Act Art. 25. Watch for 'substantial modification' and 'changes the intended purpose'. · A system prompt is guidance, not a security control.

Independent study aid, not affiliated with the IAPP. Written with AI from primary sources; laws change, so check the sources. Last updated 7 Nov 2026.