HALBACH

· MD

Review-transcript evidence audit — 28 October 2026

Purpose: support the new draft while separating founder observations, measured examples and external evidence. URLs were read on 28 October 2026; dynamic pricing should be checked again before reuse.

R01 — Generative AI at Work

Source: Generative AI at Work.

Supported: Final published study: 5,172 customer-support agents; average 15% more issues resolved per hour.

Boundary: Single-firm staggered introduction; heterogeneous outcomes; productivity is not net ROI. Supersedes working-paper 14% / 5,179 numbers.

R02 — METR developer productivity update

Source: METR developer productivity update.

Supported: 2025 experiment found 19% longer task times; 2026 follow-up cannot reliably quantify current effects because of selection and measurement problems.

Boundary: Do not present the 2025 slowdown as a universal or current finding. METR believes developers likely benefit more now; no defensible current universal percentage.

R03 — 2025 DORA report announcement

Source: 2025 DORA report announcement.

Supported: AI adoption positively associated with delivery throughput and product performance, but negatively with delivery stability.

Boundary: Survey associations, not causal estimates. Older 2024 negative-throughput results do not describe the latest report.

R04 — Claude API pricing

Source: Claude API pricing.

Supported: Sonnet 4.6 example: standard input $3/million tokens versus cache reads $0.30/million tokens.

Boundary: 90% discount applies to qualifying cached input tokens only. Cache writes cost extra. Rates vary by model, date and platform; not a whole-workflow saving.

R05 — Prompt caching

Source: Prompt caching.

Supported: Reuse an unchanged common prefix, followed by task-specific content.

Boundary: Minimum length, expiry, invalidation and actual cache hits matter; monitor observed usage rather than assuming every request hits.

R06 — Message Batches API

Source: Message Batches API.

Supported: Anthropic documents 50% lower batch API pricing for asynchronous requests.

Boundary: Suitable when immediate response is unnecessary; not a promise of 50% lower total application cost.

R07 — FinOps Unit Economics

Source: FinOps Unit Economics.

Supported: Connect technology unit metrics to business outcomes and document denominators, data sources and allocation.

Boundary: Our accepted-task metric is an application-specific synthesis: full attributable workflow cost divided by accepted tasks; define acceptance, period and cost boundary.

R08 — Knauf commercial excellence, digital and product management

Source: Knauf commercial excellence, digital and product management.

Supported: Public positioning includes customer-focused digital solutions and improving customer and employee experiences.

Boundary: Does not establish a rigid Digital=R&D / IT=operations organizational split.

Transcript corrections and presentation decisions

  • No verified basis for “75–80% organizational costs” or “coding is 30% of software work.” Omit percentages; show the complete workflow without invented allocations.
  • Avoid “AI brings zero net benefit to software teams.” DORA 2025 and the METR update require a more nuanced, time-bound claim: evaluate end-to-end results, including stability and review.
  • Founder conversations about feasibility/performance taking priority over cost are qualitative observations from their own conversations, not a representative European market survey.
  • Klarna 32¢ → 25¢ is a 21.875% reduction, not 33%. The retained 32¢ → 25¢ → 19¢ series is a 40.625% reduction overall in adjusted customer-service cost per payment transaction; do not relabel it cost per chat or all-company AI savings. Retain the existing filing citation.
  • Demo latency, fully offline operation, native 4K and smaller-GPU same-quality claims require the actual demo configuration and measurements. Do not transfer a separate fixed-input memory benchmark to the video demo.
  • Do not infer degrees from research experience. Founder biographies should say mathematics / ML research and applied mathematics / real-time GPU systems unless more specific credentials are verified.
  • Model and hosting provider are distinct choices. Precision, region, data terms and operating responsibility vary by deployment; quality is a measured acceptance gate, not guaranteed by a precision label.
  • Local AI is a conditional deployment choice, not a universal compliance guarantee or automatic saving. Include utilization, maintenance, hardware and operational ownership.
  • Time released is capacity, not automatically cash saved. Keep implementation investment separate from recurring operating costs and show the volume/time horizon when evaluating payback.

Suggested restrained evidence use

One main evidence slide: +15% support issues/hour (QJE) beside DORA’s mixed delivery/stability finding, clearly marked different contexts. Narration: there is evidence of value, but the relevant unit is the completed workflow. Put METR’s dated experiment and 2026 caveat in backup/notes, not a misleading universal negative headline.

Caching slide: stable prefix + variable task, with a clearly scoped 90% cache-read input-price example. Practical levers: remove unnecessary repeated work, select models against acceptance tests, batch when deadlines allow. All cost reductions must pass the same quality/performance gate.