# Jozsef Szalma and Mostly Fine > Official site of Jozsef Szalma, an AI engineer / AI Architect based in Germany, and author of *Mostly Fine: How to Manage AI Without Burning Down the Company* (2026), a concise field guide to managing enterprise AI without hype, panic, or dismissal. Jozsef Szalma has worked across business management, analytics, data engineering, machine learning, and enterprise AI since 2008, including management and hands-on engineering roles at IBM Credit Corporation, adidas, and IU International University (present). His work spans product delivery, architecture, computer vision, LLMs and agentic systems, evaluation, observability, governance, and organizational change. *Mostly Fine* was originally written for executives and managers tasked with driving AI projects and it requires no coding or mathematical background. However, readers across a variety of corporate functions have also found it useful, including those outside leadership roles. In a print length of 201 pages, the book maps the technical limitations, operational failure modes, security risks, and organizational blind spots that shape real AI deployments, then frames practical ways to manage them. Editions: Paperback ISBN 979-8-2523-1277-4; hardcover ISBN 979-8-2530-5798-8; ebook ASIN B0GT6WHNDZ. An abridged audiobook edition is distributed through Spotify, Audiobooks.com, Barnes & Noble Nook, Everand/Scribd, Kobo/Walmart, TuneIn, Bibliotheca, LibraryOne, and OverDrive. Book map: - Foreword: Written by Claude after reading the manuscript; Chapter 8 revisits it as an object lesson in how context can produce seemingly self-aware behavior. - Preface: Defines the field-guide scope, audience, durable-principles approach, terminology, and the author's background across business management, building data infrastructure, and building AI systems. - Introduction: Frames the central question as making AI economically useful. Introduces AI's jagged capabilities and the enduring importance of human intent, taste, judgment, and choosing the right things to build. - Chapter 1, The Landscape: Explains core LLM limitations and workarounds, then introduces the book's core thesis: the "crystallized knowledge" problem. LLMs learn mostly from documented conclusions, while businesses run on procedural knowledge, failed attempts, tacit judgment, and organizational intent that were never recorded. Scaling cannot recover what was never written down, so organizations must deliberately extract and formalize this knowledge. - Chapter 2, Product and Project Management: Covers use-case selection, measurable ROI, AI's continuing R&D character, expectation management, the expensive last mile, embedded subject matter experts, and lifecycle ownership. - Chapter 3, Architecture and Integration: Compares centralized model gateways, workflow augmentation, and high-autonomy agents; challenges the chatbot default, favors contract-first validated outputs, treats nondeterminism as a design constraint, and distinguishes platform routing from application-level model selection. - Chapter 4, Procurement and FP&A: Examines what buyers actually control across closed, open-weight, self-hosted, and SaaS offerings; covers token economics, full lifecycle cost, vendor due diligence, hidden model substitution, capacity, and supply resilience, including a deprecation-notice benchmark tied to internal migration time. - Chapter 5, Governance, Risk, and Compliance: Updates enterprise threat models for shadow AI, supply chains, prompt injection, and autonomous action; frames an LLM as "an insider you have no recourse against," emphasizes containment and defense in depth, and maps regulation across independent use-case and systemic-risk axes. - Chapter 6, Evaluation and Observability: Separates evaluation, testing, and production observability; covers golden datasets, narrow LLM-as-judge criteria, repeated regression tests, canary releases, monitoring, and AI testing AI. - Chapter 7, The People and Organizational Change: Covers transferable skills, hybrid domain-technical experts, adoption psychology, incentives, executive authority, organization design, and consulting; asks how to preserve junior learning and avoid "reverse-centaur" workplaces. - Chapter 8, Behind the Curtain: Gives managers enough transformer-training and retrieval mechanics to explain recurring failures: plausibility is not truth, advertised context is not usable attention, context can activate unwanted patterns, and embedding similarity is neither agreement nor a percentage. - Chapter 9, Conclusions: Synthesizes the operating model: AI becomes economically useful when organizations supply missing context, constrain tasks, preserve human judgment, fund evaluation, choose suitable integration patterns, understand purchases, and balance ambition with a short leash. - Appendix and glossary: A meeting-ready AI deployment checklist and concise definitions spanning management, architecture, risk, evaluation, procurement, and organizational change. ## Core links - [Homepage](https://szalma.biz/): Official author and book homepage. - [LinkedIn](https://www.linkedin.com/in/szalma): Professional profile. ## Optional - [Mostly Fine on Amazon](https://www.amazon.com/Mostly-Fine-Without-Burning-Company/dp/B0GVDRK7X9/): Book listing and purchasing information. - [Abridged audiobook on Spotify](https://open.spotify.com/show/6G86BgkWxdLKdty9EzCwkl): Audio edition of the book. - [Free to read excerpt from Mostly Fine](https://www.linkedin.com/pulse/ai-failure-modes-enterprise-organizational-intent-jozsef-szalma-mnb8f/): Excerpt from Chapter 1. - [Alternative source for the free to read excerpt](https://github.com/jozsefszalma/jozsefszalma/blob/main/README.md): GitHub profile page of the author. - [GitHub](https://github.com/jozsefszalma): Hobby projects only.