Worked examples and teaching boundaries
Current direction, 24 October: these are retained research workpapers, not required stage cases. Existing synthetic calculations are assumptions, not measured effects. No measured own-work case or fabricated replacement is required. Prefer a qualitative decision walkthrough; see current direction [local reference omitted].
Date: 2026-10-22. All figures in this memo are synthetic teaching assumptions, except where explicitly linked to measured internal evidence. They are not Knauf data, wage estimates, vendor quotes or savings forecasts.
Example A: Product-data workflow
Public Knauf context makes product records a relevant example. It does not establish the following workflow or numbers. See K02/K05 and the audience research.
Suppose a team processes 100,000 product records per month. For comparison, 98,000 records meet the final acceptance rule in each scenario. The acceptance rule includes correct values, source/version and required domain approval. The percentages below are chosen to teach the economics, not measured model performance.
| Monthly assumption | Baseline | Cheaper model alone | Improved workflow |
|---|---|---|---|
| Attempted records | 100,000 | 100,000 | 100,000 |
| Accepted records after review | 98,000 | 98,000 | 98,000 |
| Model/tool cost per attempted record | €0.04 | €0.01 | €0.02 |
| Share requiring human review | 20% | 30% | 15% |
| Minutes per reviewed record | 2 | 2 | 2 |
| Value of review capacity per hour | €60 | €60 | €60 |
| Attributable platform/runtime per month | €1,000 | €1,000 | €1,500 |
| Model/tool charges | €4,000 | €1,000 | €2,000 |
| Review capacity value | €40,000 | €60,000 | €30,000 |
| Total attributed service cost | €45,000 | €62,000 | €33,500 |
| Cost per accepted record | €0.459 | €0.633 | €0.342 |
Equations:
- Review capacity value = attempted records × review share × review minutes / 60 × hourly capacity value.
- Attributed total = model/tool charges + review capacity value + platform/runtime.
- Unit cost = attributed total / accepted records.
- Count all attempts and rejected records in costs. Do not divide by all submitted records when claiming accepted-output economics.
Interpretation: the cheaper model saves €3,000 of usage charges but creates €20,000 more valued review work in this constructed example. The improved workflow is a separate hypothetical candidate with a higher platform cost and less review. It must earn its acceptance and review rates in an experiment.
Possible changes to investigate are processing only changed fields, validating deterministic values, narrowing the model’s job, routing routine cases and scheduling non-urgent work. The numerical result does not prove any of those changes works.
Cash savings versus capacity released
Assume the improved workflow requires €12,000 of incremental cash implementation/evaluation spend. If existing staff perform the work without additional cash outflow, classify it as an economic opportunity cost and label the payback accordingly.
Baseline runtime cash charges are €5,000/month: €4,000 model/tools plus €1,000 platform. Candidate charges are €3,500/month. Conditional runtime cash reduction is therefore €1,500/month, if all those charges are actually variable or cancellable.
Review falls by 166.7 hours/month in the illustration. Valuing those hours at €60 adds €10,000 of monthly capacity value. That becomes cash savings only if actual spending falls. It may instead allow a team to process more work or address a backlog.
- Runtime-cash-only simple payback: €12,000 / €1,500 = 8 months.
- Including the entire review-capacity value: €12,000 / €11,500 = 1.04 months.
- The second number is an economic-capacity scenario, not an invoice-savings promise.
- Neither result includes ramp-up, tax, discounting or uncertainty. Additional continuing optimisation support must be deducted if not already included.
- If only part of the capacity has a useful redeployment, value only that part. If cash charges are committed, revisit avoidability.
This distinction is likely more valuable to leadership than a large percentage reduction. In the talk use a cohort of 1,000 attempted records, of which 980 are accepted. The total attributed cohort costs are €450, €620 and €335. These are not costs per 1,000 accepted records. The latter are €459.18, €632.65 and €341.84. Keep the monthly-volume assumptions in the workpaper rather than implying a Knauf workload forecast.
Example B: Different service requirements change eligible levers
For an overnight record update, waiting for a vendor’s eligible batch tier may be acceptable. An interactive employee query needs an agreed response-time requirement, current source evidence and a clear escalation route. Both can use AI, but neither the best model nor the best hosting choice follows from the label “AI assistant.”
A provider discount applies only to eligible billing components. The cost-economics memo contains verified terms and a separate illustrative calculation showing why a 50% API discount does not imply 50% lower service cost. See C05/O06 and cost economics.
Historical example C
[Private experiment evidence omitted from this public copy.]
Sensitivity worth measuring before presenting real numbers
For the main example, review intensity and the value/avoidability of human time dominate the result because of the deliberately chosen assumptions. Real work must measure them. Repeat the calculation across task complexity, language, record size and critical-error classes. Averages can conceal a poor result in a commercially important segment.
Published benchmark scores cannot supply the review probabilities in this table. A held-out test and a representative production pilot must establish them.