Raiffeisen Bank Ukraine: AI Made Engineers Different

Key takeaways

1. Date / source. May 2026 (relative "2 days ago" at the time of consultation, i.e. ~2026-05-05), Medium @milhibisidek. URL: https://medium.com/@milhibisidek/ai-didnt-make-our-engineers-just-faster-it-made-them-different-95f1c1d4efd0
2. Format. CTO op-ed, organizational case study (~17 min read).
3. Author. Hryhorii Tatsyi, CTO of Raiffeisen Bank Ukraine — a regulated systemic bank, ~900 IT engineers.
4. Study period. May 2025 → April 2026 (12 full months).
5. Title aphorism. "AI didn't make our engineers just faster. It made them different."
6. Strategic pivot phrase. "AI expanded our production possibility frontier, and we deliberately allocated the freed capacity." ### Organizational framing — the observed transformation | Dimension | Before (May 2025) | After (April 2026) | Change | |-----------|------------------|--------------------|-----------| | IT headcount | ~900 engineers | ~825 (−75 people, including 64 engineers) | −8% | | AI adoption | 62% | 83% | +21 points | | Engineers ≥50% code via AI | — | 68% | new metric | | Onboarding (1st PR) | 60-90 days | ~40 days | −~50% | | Code volume shipped | baseline | increased | ↑ | | Blocking incidents | baseline | −70% | ↓ | | Critical resolution time | baseline | −68% | ↓ | | High-severity security alerts resolved | baseline | +155% | ↑ | Reading: the headcount contraction (−8%) coincides with improvement on every axis — output, quality, security, onboarding. This is the central data point Tatsyi highlights, and he does not dodge it politically. ### Three emerging engineer archetypes | Archetype | Tools | Productivity effect | Scope effect | |-----------|--------|---------------------|------------------| | Copilot-only | GitHub Copilot alone | +10-25% on PRs | stable scope | | Multi-tool | combination of several AI assistants | story points ×1.5-3 | cross-repo scope +50-80% | | Claude on corporate stack | Claude Code (or equivalent) integrated into the internal stack | code volume ×4.5 | scope radically expanded | Convergence: senior architects return to active development after years away from it (consistent with Karpathy's claim that "agents reduce friction to creation" and Cherny's "best accountant writes accounting software"). Counterintuitive insight: AI lifts underperformers to baseline rather than mainly accelerating top performers. A position opposite to the "elite 10×+ tail" reading (Cherny / Curran top 5% / Karpathy) — Tatsyi describes a bottom-up catch-up effect that tightens the distribution. The two readings are compatible: the distribution tightens from the bottom AND widens at the top (top performers who stay at 10×+). ### Seven AI products built that did not exist before | # | Product | Description | Key metrics | |---|---------|-------------|----------------| | 1 | Service Knowledge Hub | Auto-generated microservice documentation via Kubernetes parsing | 57 microservices, 83 releases/month | | 2 | Mobile Android workflow CI | Automated plan / implementation / test pipeline for mobile | complete redesign of the mobile SDLC | | 3 | AI Agent Portal | Internal portal for automatic MCP agent generation from OpenAPI specs | 2,085 users, 649 MAU, 87 days to reach this adoption | | 4 | Shift-left Security Plugin | In-IDE vulnerability detection before commit | −82% exposed secrets | | 5 | DevPortal | Backstage + AI Kubernetes diagnostics agents | −68% critical incident resolution time | | 6 | DRAIF MCP | Text-to-SQL over a Data Lake | 10,000 tables, embedding fine-tuned ×2 OpenAI models | | 7 | Call Evaluation | Audio transcription analysis + script redesign | >97% accuracy, voted best product in the Raiffeisen group (RBI) | Strategic reading: this is not a list of experiments, it is a product portfolio deployed in production, with measurable internal adoption, and one product (Call Evaluation) that spans the group's subsidiaries (moving from local to group-wide RBI). ### The pivot thesis — production possibility frontier > "AI expanded our production possibility frontier, and we deliberately allocated the freed capacity."
7. Keyword 1 — "expanded". AI does not do the same thing better, it enlarges the set of what's possible.
8. Keyword 2 — "deliberately allocated". the freed capacity is redirected by managerial decision, not mechanically absorbed into more of the same work.
9. Three reallocation directions. 1. Features (new products — the 7 listed above). 2. Stability (incidents −70%, resolution −68%). 3. Technical-debt repayment (rare and capitalizable from a CTO's standpoint). ### The reframed evaluation question Wrong question: "By how much % did our existing KPIs increase?" Right question: "What did your engineers build that didn't exist before?" Why it matters: percentages on existing metrics miss the main transformation — not speed but type of work. Optimizing for legacy KPIs means missing the strategic window where AI makes it possible to build what was previously unaddressable. ### Tie-in to the watch corpus #### Numeric convergences (committed median 3-5×)
10. Frizzo. (LinkedIn 2026-05-05): 3-5× productivity multiplier over 1 year of daily use.
11. Wescale Usine Logicielle Augmentée. (2026-05-03): "realistic X3-X4".
12. Curran/Intercom. (2026-04-16): 3× R&D productivity over 16 months (entire organization, 500 R&D staff).
13. DORA Report 2025. (2025-09-23) and Stanford Denisov-Blanch (2025-11-23).
14. Tatsyi/Raiffeisen. (2026-05-05): story points ×1.5-3 (multi-tool), ×4.5 code volume (Claude corporate stack). #### Onboarding convergence (60-90 days → ~40 days)
15. Tatsyi/Raiffeisen. 60-90 days → ~40 days.
16. Anthropic internal studies. (cited by Sun NYT 2026-04-30 and others): 82 days → 40 days.
17. → Correct reading: independent convergence between a European bank and a Silicon Valley AI player on the same target figure (~40 days). This is a robust stylized fact for 2026. #### "Seven new products" / creative capacity convergence
18. Tatsyi. 7 AI products in 12 months.
19. Cherny. (2026-05): 100% of code generated, "a few dozen PRs/day, 150 PRs in a single day record", multiple Anthropic Labs products.
20. Curran/Intercom. (2026-04-16): Skills-Based Plugin Architecture (153 contributors, 267 skills).
21. Wescale. (2026-05-03): "long-standing needs that remained too costly can finally be addressed".
22. → Strong convergence: the relevant measure is no longer speed on existing work but the portfolio of new products / newly addressable spaces. #### "The job is changing shape" convergence
23. Frizzo. (2026-05-05): "the new bottleneck is supervision", "writing muscle atrophy".
24. Tatsyi. (2026-05-05): "AI didn't make our engineers just faster. It made them different", three emerging archetypes, senior architects returning to active development.
25. Karpathy. (2026-04-29): Software 1.0/2.0/3.0, "outsource thinking but not understanding".
26. Mornati. (2026-03-14): What is a Developer When We Use Coding Agents?
27. Habert PROJ-AI. (2026-05-05): "agent directives + Decision Records + five validation dimensions".
28. → Cross-cutting convergence: the nature of the work has changed, not just its speed. Tatsyi contributes the banking-sector organizational data point that was missing from the corpus. #### Productive tension with "AI lifts underperformers"
29. Cherny. (2026-05): elite 10×+ tail (150 PRs/day record).
30. Curran/Intercom. (2026-04-16): top 5% at 6× median PR throughput (≈ 18× pre-AI baseline).
31. Karpathy. (2026-04-29): "10× is not the speed up — people who are very good at this peak a lot more than 10×".
32. Tatsyi/Raiffeisen. (2026-05-05): "AI lifts underperformers to baseline" — a bottom-up catch-up effect.
33. → A compatible and productive reading: the distribution tightens from the bottom (Tatsyi) AND widens at the top (Cherny, Karpathy, Curran top 5%). Both phenomena coexist. Useful for executive-committee presentations that want to both reassure (catch-up) and inspire (top performers). #### FR / Central-European vs. Anglo-Saxon positioning
34. Tatsyi is European (Ukraine, Austrian RBI group), a regulated bank, in a wartime context.
35. His methodological rigor (12 months, internal figures, per-archetype granularity) aligns more with French caution (Wescale, Habert) than with American optimism (Cherny, Curran).
36. To be used in FR presentations as a European banking case study complementing the Wescale/consulting and Curran/SaaS figures.
37. Rare advantage. a systemic bank CTO publishing his internal figures — a type of testimony almost absent from the 2026 corpus. ### Limitations to flag
38. No detailed quantitative methodology. Tatsyi reports percentages without precisely documenting how they are measured (e.g., "story points ×1.5-3" — what baseline? what attribution? what bias adjustment?).
39. Author with low public visibility. (25 Medium followers) — authority rests on institutional position rather than personal brand. Should be weighed against other sources if the stakes involve citing this in an executive committee.
40. No explicit discussion of the cognitive costs. (FOMO, deskilling, ownership erosion) that Frizzo names. Tatsyi is organizational, Frizzo is individual — the two notes complement each other.
41. No discussion of banking-sector-specific regulatory risks. (GDPR, EU DORA, ECB/EBA supervision) — surprising for a systemic bank CTO. The topic was possibly omitted to preserve the readability of the public article.
42. Onboarding convergence 82→40 days with Anthropic. needs verification that the cited Anthropic data point is indeed the study Tatsyi refers to (risk of reverse cherry-picking — an apparent convergence on a round number).
43. Sample of n=1 organization. a single case, even if quantified. Should be replicated at other European banks to conclude a sector-wide stylized fact. ### To be used for
44. Banking/insurance/regulated-sector executive committee presentations. a rare, quantified European case study, a counterweight to the Silicon Valley over-representation in the corpus.
45. HR / transformation strategy. justifying the reallocation of freed capacity rather than simple cost reduction.
46. Boards / transformation sponsors. reframing the evaluation question ("what did your engineers build that didn't exist before") — a tool for defusing the % productivity debate.
47. Convergent quantitative sourcing. 60-90 days → ~40 days onboarding (Anthropic + Tatsyi), 3-5× committed median, headcount contraction compatible with improved output/quality/security.
48. Equity debate / productivity-distribution skew. to be used alongside Cherny/Karpathy/Curran top 5% to present a nuanced reading — AI tightens from the bottom AND widens at the top simultaneously.
49. Engineering Management community. a rare CTO publication from a large regulated European IT organization on the subject.