Third review: evidence and editorial decisions
28 October 2026 · private working record · supports html-v15
This record reconciles the founders’ third review transcript [local reference omitted] with primary sources and retained project records. The transcript is a direction-setting summary, not a verified measurement log. Figures and source locators are in the source register and chart data.
Decisions for the main deck
| Review question | Resolution | What the evidence does not establish |
|---|---|---|
| How is the whole AI budget divided? | Four cost branches, then allocate shared costs to a workflow. The graphic is a cost-boundary schematic informed by FinOps. | No verified universal percentages or Knauf cost split. Equal-sized branches do not mean equal spending. |
| Which newer financial examples work? | Klarna’s reported 2025 assistant savings and DBS’s reported 2025 analytics/AI value, disclosed in 2026. | Neither is an independently estimated net ROI or a promise for Knauf. Their currencies and metric definitions differ. |
| What is a concrete underdelivery example? | Deloitte’s AI-assisted government report triggered a request to repay its final instalment after quality and disclosure failures. | The cited brief verifies a request, not settlement, total fee refund or a negative project ROI. |
| Can Care Access have absolute bars? | No absolute dollar or per-record duration baseline is disclosed. Show the reported reductions directly, before the mechanism. | Normalized bars would be mathematically possible but would not answer the speakers’ request for underlying absolute values. |
| Does the Opus change preserve the old cache ratio? | No. Opus 5.5 ordinary input is $4/M; qualifying cache reads are $0.20/M: a 95% reduction for those read tokens. | This is not a 95% reduction in the full request or workflow. Cache writes, misses, output and other costs remain. |
| What should the future slide show? | Three separate observations: token unit price, a dated GPU instance price cut, and historical cost of reaching a fixed benchmark score. | They have different denominators, periods and populations. They cannot be compounded or drawn as one market time series. |
| What is the evaluation slide’s message? | Define the accepted task, test capability, establish a complete baseline, change one constraint and remeasure. | No assumed saving before the comparison; acceptance and human review stay inside the measurement boundary. |
Financial evidence and scope
Klarna (T04). The 2025 Form 20-F reports approximately US$59 million in 2025 customer-service assistant cost savings and 80% of service chats handled. This is a company estimate. Its report retains human service as an option; do not retell this as a simple AI reversal. Locator: PDF page 188, printed page 184, operating-leverage discussion. The earlier cost-per-transaction series has a different denominator and remains separately sourced.
DBS (T05). The 2025 annual report attributes approximately SGD1 billion of economic value to data analytics and AI/ML, across more than 430 use cases and 2,000 models. This is broader than generative AI, and economic value is not net profit. Use it as a reported scale example, not a directly comparable bar beside Klarna.
Deloitte / DEWR (T06). A government brief cleared 17 September 2025 records a requested repayment of A$97,587.11 including GST following report-quality issues and failure to follow AI-use disclosure policy. The report had citation/reference problems and required correction. The defensible leadership lesson is that verification affects whether output is accepted and paid for. Do not claim that every finding was wrong, that all fees were refunded, or that AI alone caused the whole loss. Locator: PDF page 66, printed page 108. Parliamentary correspondence provides additional detail on the AI-assisted citation process; direct retrieval returned 403, so the accessible DEWR brief anchors the monetary claim.
Care Access (T03). The co-authored AWS case reports 86% lower Bedrock service cost and 66% less processing time per record. It does not disclose absolute spend or duration baselines. Processing hours saved are not employee hours saved. Preserve this component boundary and do not imply that Opus 5.5 was the model used in that November 2025 case.
Price and outlook verification
Opus 5.5 (T02, T07). Standard USD per million tokens: input 4; output 20; five-minute cache write 5; one-hour cache write 8; cache read 0.20. The September 2026 launch compares Opus 5 to 5.5: input 5→4 and output 25→20, both 20% lower. Its vendor estimate of 40% lower typical-workload cost is a different measure and is not used as the main graph.
GPU instance pricing (T08). AWS announced a 44% reduction for P5/H100 On-Demand instances, effective 1 June 2025 against 31 May prices. This is one dated provider/product change, not annual compute deflation or a current all-GPU market index. A general 2026 compute-price trend was not verified. A current NVIDIA benchmark page describes software throughput gains, but throughput is not the requested rental-price measure; it is not substituted into the price chart.
Technical task cost (T09). Epoch’s September 2026 study estimates an average historical 47% quarterly decline since 2023 in frontier inference cost at fixed performance across five benchmarks. It accounts for model inference consumption, not human review, integration or operations. Benchmarks, model switching and noisy frontier estimation limit transfer to business workflows. It is historical evidence, not a 2027–28 forecast.
Batching (T10). The documented async service discount is usable evidence. A claim that a provider necessarily waits for idle GPUs is not. Talk about deadline flexibility and the contracted completion window, with data terms checked for the chosen provider.
Internal demo evidence
[Private project provenance and founder rehearsal notes omitted from this public copy.]
Transcript ambiguities and readiness
[Private speaker preparation omitted from this public copy.]
Verification method
Two research agents independently checked financial examples and deployment/pricing evidence. The parent opened the cited primary pages, verified the relevant PDF passages and reconciled a live NVIDIA page change before implementation. The implementation agent maintains the slide/source/narration contracts and renders both themes. Evidence access date: 28 October 2026. Source excerpts are not reproduced wholesale; the register records claim scope and exact locators.