Module 4 of 11
Evaluate costs and benefits of AI solutions
Build a defensible ROI case, decide build versus buy versus extend, and control the variable cost of generative AI.
Exam questions Q31–Q40 7 units on Learn
Open module 4 on Microsoft Learn ↗ Watch the episode (46 min)
Why this module matters
AI economics differ from traditional software economics in one crucial way: a large share of cost is variable and usage-driven. Licences are predictable; tokens, tool calls, and retrieval are not. Architecture decisions therefore have direct, ongoing financial consequences.
The exam tests whether you can construct an honest business case — including the costs teams habitually forget, and the discipline of comparing against a do-nothing baseline.
The one big ideaROI is net quantified benefit over total cost of ownership across a defined period — always compared against the cost of doing nothing.
What Microsoft Learn covers
Units in this module, as published on Microsoft Learn:
- Introduction
- Evaluate ROI criteria for AI-powered solutions
- Create ROI analysis for a proposed AI solution
- Analyze whether to build, buy, or extend AI components
- Implement a model router to intelligently route requests to the most suitable model
- Module assessment
- Summary
Watch this module
Microsoft Learn publishes a free video course for AB-100 on YouTube. This module is covered by one 46 min episode. An instructor walking through the module is often faster than reading it twice.
Key concepts and definitions
| Total cost of ownership (TCO) | Build cost plus licensing, consumption, integration, monitoring, content maintenance, support, and change management, over the solution's life. |
| ROI | Net quantified benefit — time saved, revenue gained, cost avoided — relative to TCO over a defined period. |
| Hard benefit | Directly measurable in currency: reduced handling time, avoided headcount, lower error rework, retired licences. |
| Soft benefit | Real but harder to price: employee experience, reduced cognitive load, faster onboarding, improved consistency. Still belongs in the case, stated honestly. |
| Do-nothing baseline | The quantified cost of the status quo. Always include it — it is what the proposal is actually competing against. |
| Variable cost drivers | Request volume, input and output tokens, retrieval scope, tool and connector calls, and model choice. |
| Model router | Directs each request to the most suitable model, so cheap tasks do not pay frontier-model prices. |
| Build / buy / extend | Build for competitive differentiation not available off the shelf; buy for commodity capability; extend when a platform capability is close and configuration or a connector closes the gap. |
| Opportunity cost | The value forgone by spending scarce delivery capacity on this initiative rather than the next best one. |
| Staged funding | Releasing budget in tranches tied to value checkpoints, with go/no-go gates after pilot and after scaled rollout. |
Design guidance
Costs teams routinely forget
- Ongoing operations — monitoring, evaluation runs, incident response. The single most under-estimated category.
- Content maintenance — someone must keep grounding sources accurate and current, forever.
- Change management and training — adoption does not happen by itself, and unadopted agents have infinite cost per unit of value.
- Re-evaluation after model updates — platform model changes force regression testing.
Making the pilot honest
- Before extrapolating pilot savings enterprise-wide, confirm the pilot population, task mix, and measurement method are representative.
- Model an adoption ramp rather than assuming instant full uptake.
- State assumptions explicitly and show the case at conservative, expected, and optimistic adoption.
Reducing cost without losing quality
- Right-size the model per task and route intelligently; do not send every request to the largest model.
- Constrain retrieval scope — narrower, curated sources cost less and answer better.
- Cache repeated work and cap output length where long answers add no value.
- Never cut cost by removing logging, monitoring, or grounding. Those are controls, not overheads.
Good practice vs. anti-pattern
Do
- Include a do-nothing baseline in every business case.
- Fund in stages against value checkpoints.
- Quantify soft benefits explicitly rather than silently.
- Use a model router and right-size per task.
Avoid
- Extrapolate pilot results without validating representativeness.
- Express ROI in tokens, agents deployed, or model accuracy.
- Treat operations and content upkeep as one-off costs.
- Save money by disabling observability.
Exam traps to watch for
TrapThe most under-estimated cost is ongoing operations — monitoring, evaluation, content maintenance, change management. Not licences, not hardware.
TrapROI must be expressed as net benefit against TCO. Answers phrased as token counts, deployment counts, or accuracy percentages are distractors.
TrapThe do-nothing baseline is always included, not only in regulated industries and not only on request.
TrapOpportunity cost still applies when the budget is fixed — arguably it matters most then, because capacity is the binding constraint.
Readiness checklist
You are ready to move on when you can do all of these without notes.
- Construct an ROI statement with benefit, TCO, period, and baseline.
- Name four cost categories teams commonly omit.
- Decide build, buy, or extend for a given capability and defend it.
- Explain what a model router does and when it pays for itself.
- Describe three ways to reduce cost that do not degrade quality or control.