What you need to know
A technically sound design still fails if nobody can show it is worth paying for. The exam expects you to connect architecture choices to named business outcomes and to defend them with numbers the business recognizes.
The value pillars
Most cases rest on one or two of these:
| Pillar | The business question | Typical measure |
|---|---|---|
| Efficiency | Do we do the same work with less effort? | Handling time per case, cases per analyst per day |
| Productivity | Can our people do more, or more valuable, work? | Output per person, time freed for higher-value tasks |
| Transformation | Can we do something we could not do before? | New service, new channel, coverage that was not possible manually |
| Cost | Does it cost less to run than the alternative? | Cost per transaction, total run cost against today's cost |
| Performance and SLAs | Does it meet the service level we promise? | 95th-percentile response time, availability, accuracy threshold |
Name the pillar the stakeholder cares about first. A contact-centre director may care most about handling time; a compliance head may care about consistency and audit evidence; a product owner may want a new capability.
Build the value case in four steps
- State today's baseline in a business unit. "Analysts spend 40 minutes per claim", measured from real operational data.
- Predict the future state in the same unit. "15 minutes per claim, including review of the draft."
- Subtract the run cost, taken from the sizing model rather than from a demo.
- State the payback period and how sensitive it is to volume, adoption and accuracy.
Size cost and latency properly
The sizing model takes call volume (from the business owner), tokens per request as a distribution, the model tier, and sensitivity scenarios. Then look at the levers:
- Prompt caching cuts cost and latency when a long, stable prefix (system prompt, reference documents) repeats across many requests. Cache writes cost more than normal input, and cached content expires after a short time without use, so infrequent traffic gains little.
- Batch processing lowers the price substantially for work that can wait, such as overnight document runs. It does not fit interactive, real-time use. For regulated data, confirm the batch route is covered by the same compliance arrangements.
- A smaller model tier for simpler steps, with an evaluation proving quality holds.
- Output length limits where long answers add cost without adding value.
In the exam
Value questions usually give you a stakeholder, a goal and some numbers. The best answer states the benefit in the stakeholder's unit, builds on a measured baseline, includes the real run cost and the human review the design requires, and chooses the cost or latency lever that matches the workload's shape. Be wary of options that promise savings from full automation, quote average latency for an SLA, or pick the most capable model without a reason tied to the stated pillar.
Check your understanding
A value case assumes analysts will stop spending time on claim summaries, but the design keeps a reviewer on every summary. What is wrong with the case?
Show answer
Answer: A. When the design keeps a person in the loop, the value case must count the review time that remains. Modeling full automation promises savings that will not appear.
Which SLA statement fits a customer-facing assistant best?
Show answer
Answer: D. SLAs are about the slow tail under real load. Medians, monthly averages and single-request tests all hide the moments when users wait longest.
A nightly job reconciles 200,000 documents and nobody waits for the results. Which lever cuts cost most directly?
Show answer
Answer: C. Work that can wait is what batch processing is for, and it is priced well below interactive calls. Streaming changes perceived speed, not cost, and a larger model raises the price per token.
Apply it: Build the support value case
Corvane's Finance partner asks for a value case for the routing design from lesson 1.3, scoped to order-status and product questions. Build it in the four steps from this lesson, using the assumed prices on the cases page.
- Baseline. Agent hours and cost spent today on the in-scope contact types.
- Future state. Assume Claude fully resolves 70% of in-scope contacts (to be confirmed in a pilot) and that each escalated contact saves 2 minutes because the agent receives a summary.
Why: modelling escalations keeps the case honest about the work that remains. - Run cost. Estimate the monthly model cost at Tier M, including peak season, and check it against the $18,000 budget. Remember that conversation tokens accumulate across turns.
Done when: you know whether caching is optional or required. - Payback and sensitivity. Payback on the $250,000 build, and the result if resolution is only 40%.
Show model answer
1. Baseline. In scope: 38% + 21% = 59% of 95,000 = 56,050 contacts a month. Blended handling time 0.55 × 7.5 + 0.45 × 11 = 9.1 minutes. That is about 8,480 agent hours, or $271,300 a month at $32 an hour.
2. Future state. Resolved: 70% × 56,050 = 39,235 contacts × 9.1 min = 5,930 hours saved. Escalated: 16,815 × 2 min = 560 hours. Total 6,490 hours, about $207,800 a month of agent capacity.
3. Run cost (per in-scope contact, four turns):
| Item | Tokens | Cost |
|---|---|---|
| Stable prefix, cached, read 4 times | 24,000 | $0.007 |
| Conversation input (2,500 + 5,000 + 7,500 + 10,000) | 25,000 | $0.075 |
| Output (4 × 300) | 1,200 | $0.018 |
| Per contact | about $0.10 |
56,050 contacts × $0.10 = about $5,600 a month; at peak (2.5 times) about $14,000. The classifier on all 95,000 contacts at Tier S adds under $100. Within budget. Without caching the prefix, each contact costs about $0.17, and peak reaches about $24,000: caching is required, not optional.
4. Payback and sensitivity. Net monthly benefit: $207,800 minus $5,600 run cost minus $6,000 platform cost = about $196,000, so the $250,000 build pays back in under 2 months. At 40% resolution, capacity saved falls to about 4,510 hours ($144,400), still paying back in about 2 months.
The caveat Finance will raise: this is agent capacity, not cash. It becomes savings only if Corvane changes staffing, avoids peak-season temporary hires, or redeploys agents to work it values. The case should say which.