Get your systems to define the story of your organization – Kovács & Hermesz (Talk Outline)

The day’s closing (5 p.m.) talk by Kovács (chief AI strategist & co-founder, AI Enablement Academy; ex-cloud company that grew from 86k → 1.6M people while he was there) and Clara Hermesz (ex-Facebook/Meta and Uber; built/scaled global upskilling, a metaverse tech-education program for 3,000+, and Uber’s global mentorship/coaching). The framing question: getting your systems to define your company’s story — “surveillance, or hyper-optimized operational excellence?”

1. The Problem Space (opening polls)

• Poll 1 — have an internal knowledge base? A few hands. Updated last week? One hand. (“Didn’t your company learn anything new last week?“)
• Poll 2 — could any given employee describe exactly what they do? Zero hands. Nor the standardized processes they follow, nor how their day-to-day connects to the company’s overall goal (and would that answer even be correct?). “Welcome to the problem space.”
• Thesis: “Knowledge is not the superpower — sharing knowledge is.” You can’t talk about systems that “remember everything” while being dishonest about what was never captured in the first place.

2. Speakers & Community

• Kovács — chief AI strategist, co-founder AI Enablement Academy; rode a growth wave from 86,000 → 1.6M people.
• Clara Hermesz — ex-Facebook/Meta and Uber; scaled global upskilling & social learning; ran a metaverse education program (3,000+); redesigned Uber’s mentorship/coaching. Lesson: tools aren’t the problem — getting people to change how they work is.
• Experience spans early-stage startups to thousand-person scale-ups, Europe and the Middle East — the problem exists across geographies and sizes. Community ties: accelerators/innovation studios (360 Social Impact, a Seattle nonprofit on the future of work) and the Agentics Foundation (co-founding members; Seattle chapter organizers).

3. Why Scale Matters — Volume Kills, Impact = Blast Radius

• Volume kills in an unorganized, uncontrolled company.
• Impact radius = blast radius: hire 10 people, or drop an agent / an agentic swarm running hundreds of agents into the system — ask “what could possibly go wrong?”
• Knowing how work is actually done day-to-day, relative to the bigger goal, becomes a question of existence and survival.

4. Old Capture Is Broken

• Institutional knowledge is “captured” in wikis, playbooks, PRs, issues, codebases — much already outdated.
• Three broken realities:
  • Hired for a role that’s completely different by the time you’re through the door.
  • Your best people already left with the know-how, uncaptured.
  • Interviewing to capture knowledge is a broken way.

5. The New Way — Ontologies, Temporal Graphs, Provenance

5.1 Broken “new” experiments

• Mark Zuckerberg / Meta — surveillance-flavored keylogging spun as “you’re the best person to teach our AI to do the job well” → employees revolt.
• Marc Benioff / Slack — reads the company’s DMs and channels every morning hunting intervention opportunities → people hate it. New, but already broken.

5.2 Ontologies

• Poll: familiar with ontologies? ~1%.
• Think of it as your organizational-level type system — for processes/business facts instead of code.
• Biggest value: prevents semantic drift — an operational/survival risk when humans and agents make judgment calls on info that’s missing (so they assume), stale, or nonexistent.

5.3 Humans vs. agents tension

• For agents, observability is fine and constant modification is a given; for humans, that reads as surveillance and change (what humans hate most).
• Resolution: radical transparency from the top — leadership open about how they use AI, what they built, lessons learned, and a clear strategy → shows opportunity instead of resistance/fear.

5.4 Don’t be hypocrites — it’s already logged

• At work, everything is on record — Slack, email, network logs, CrowdStrike Falcon on laptops. Assume total recording.
• If you run the company: organize that information. Collect/log and draw lines — not just what happens but the metadata and knowledge graphs around it.
• Learn temporal knowledge bases/graphs (time matters for changing knowledge and system behavior over time; impact on productivity/performance).
• Establish provenance — answer the hard, real meeting questions: “What do we know? What does it mean? Who created it? Can we use it? Is it true? Outdated? Still relevant/applicable?”
• It’s old technical standards applied to business operations — e.g., FAIR (Findable, Accessible, Interoperable, Reusable), provenance, ensuring answers are true/correct/trustworthy.

6. Surviving the Change — Skills & Spaces

• Future/soft skills (hardest to master): systems thinking, critical thinking, growth mindset (“I just don’t know it yet”), experimental mindset (test, iterate, be okay to fail), judgment, giving clear directions, asking the right questions, accountability.
• Build them in spaces — humans have an innate need to come together, share, give/get feedback: learning circles, show-and-tells, lunch-and-learns, and the Agentics Foundation’s Friday Hacker Space.

6.1 Fighting resistance

• These spaces already exist, but people attend mandatory trainings with zero interest.
• Blunt analogy: hire someone used to Teams/Outlook/Docs into a Google/Slack shop and they refuse the company’s tools — how fast would you fire them? AI is also just a tool. Brute force is part of the picture but not the best solution.
• Better: practical, relevant enablement. AI is currently ”extra on top of work” — until you show practical, role-relevant examples in intentional spaces to experiment/fail/map processes and find time savings, resistance persists.

6.2 From human-vs-AI to human-with-AI

• Augment humans first so they can orchestrate agents: give them judgment, governance, accountability (human-in-the-loop), then automate and go agentic.
• Once people see the opportunity (higher-value tasks, what to automate away), they share their knowledge → feed it into chatbots, skills, plugins from a centralized place.
• Access to tools ≠ enablement. Anti-pattern: everyone gets Claude access and five engineers build the same skill. Instead, centralize the skills everyone will use — build once, with a strategy, from what already works.

7. Audit → Eliminate → Simplify → Automate

• Centralization is a chance for a full audit and to trim the fat — old RPA/process-automation wisdom: eliminate → simplify → automate (only automate what survives).
• Tough truth: sometimes it’s humans you don’t need — have answers ready for those situations or employees will revolt.
• Deciding what to eliminate isn’t rocket science — quantify/qualify input and output metrics, define what “good”/value creation looks like (which will shift week over week as augmented employees go faster).
• Design-thinking flywheel (5 steps: empathize → define → ideate → build small MVPs → test → repeat). With vibe coding, run dozens of tiny experiments per day to optimize operations.

8. “Memory Is All You Need” — Building the Engine

• Riff on “Attention is all you need” → today memory is all you need. Coding data is clean (PRs, issues, traceability, trace logs); business data is messy/multimodal — audio, video, images, meeting recordings, water-cooler conversations, tables, calculations. Build an ingest engine and an egest engine.

8.1 Multi-layer, multi-speed memory

• Attributes many attributions at the right time, applies recency decay (new truth vs. old truth). Design-thinking creativity applies. Reference: Karpathy’s auto-researcher running nightly to find optimizations; or use Claude’s skill-builder skill that runs evals/benchmarks to find what to eliminate.

8.2 The build list

• Small-time-model ingestion system — bad ingestion poisons the pool; each memory type/source needs its own parsing, chunking, metadata, quality checks.
• Hybrid response generation — BM25 (best match 25) for exact terms (IDs, error codes, product/policy names, phrases) + dense embeddings for meaning/paraphrase/fuzzy intent → hybrid search combining lexical + semantic recall.
• Fusion + re-ranking — one question yields many candidates (like Google’s hundreds of pages); fuse candidate sets, then re-rank for deep relevance so the model works on a small set, not the “big ocean.”
• Token economics (“27× on GitHub last week”) — cast a wide net with cheap models first, then use expensive models on the narrowed set.
• Permission-first memory — always multiple users (≥1 human orchestrator + ≥1 agent, up to hundreds of thousands of humans running millions of agents): who’s asking, what can they see vs. act on, prompt-injection prevention, when to require human-in-the-loop/approval — built at the system level.
• Freshness / temporal validity / decay — some facts stay true, some won’t be true next week; you know that, but can you trust a hallucinating agent? Build the egest system to apply this at system level.
• Garbage collection (like web-dev) to control entropy/disorder — dedupe via hashes, canonicalize entity IDs, prune regularly, and run evals/benchmarks against a defined “good memory result.”
• They’ve run this for >half a month on themselves and at clients; open-source the spec (QR sign-up).

9. Redesigning the Workforce

• “Move fast and break things” needs resilience and flexibility. At scale, communication breaks first → smaller teams return.
• The startup “Three H” — hipster, hustler, hacker; recurring now as Jack Dorsey’s Block: product person, builder, player-coach (the player-coach replacing middle management — a generalist who enables the team to move fast).
• Real-world examples:
  • Bad — Klarna: fired ~700 people in 2023 too early; the cost was worse than the saving.
  • Good — IKEA: AI agents handled ~50% of support tickets; they found what AI couldn’t do (interior design) and reskilled support staff into interior-design support → allegedly +$1B revenue.
  • Their own: a 5,000-person enterprise-tech client — identified 40 AI champions across 13 functions, ran a 4-day practical boot camp, and structured product/builder/change-agent triads (the all-in-one “unicorn” who has all three is rare → you need small teams).

9.1 Adoption fails at bottlenecks

• Engineering (Claude Code, Codex) races ahead for months; product, support, go-to-market don’t → friction points/bottlenecks force AI-enablement one onion layer at a time. It works eventually but introduces friction/frustration that hurts the bottom line.
• Identify the real opportunities (people + org) to intervene in time with the highest ROI/impact — via people who care about process, change management, and identifying what needs building, not just the tech.
• Kovács the pessimist: ”18 good months, maybe 24,” but humans stay part of the picture in some shape.
• The real bottleneck isn’t technology — it’s the human in charge deciding not to bring people along.
• Answer to the opening question: not surveillance — observability. “Don’t be a car on a dark road with no headlights.” Reframe: treat your system as a customer and listen to it; AI is not the goal — it’s a tool toward business excellence.

10. Q&A

• Can a company become too optimized? Yes — seen in “workforce 3.0”/centralization waves (single-threaded leader → matrix → over-optimized, losing the big picture). High-frequency design thinking catches it earlier.
• Should employees be able to opt out of monitoring? Yes — with consequences. With radical transparency they must know what’s monitored, when, why, how it’s used; they can opt out, but if it’s business-critical it may cost the job. (“You can always opt out — there are just consequences.“)
• Is radical transparency good for innovation? Innovation can’t happen in a bubble (needs spaces to bounce ideas). Kovács (ISO 56000 innovation-management trained; ran an Amazon innovation hub): the key to innovation is seeing opportunities, so transparency doesn’t hurt and is good long-term — but seeing ≠ having to act.
• More resistance: fear of change or of being measured? Fear of the unknown — until people understand what AI is / can do. (Amazon context: the problem isn’t measurement, it’s non-transparent standards — measured against what to be “good”?)
• Organizations in 10 years if AI succeeds? “Not to be too dark” — likely forced change in capitalism/society, redefining value, purpose, and how we organize; won’t look anything like today.
• More dangerous: too little or too much data? Not understanding the data. With too much you can at least choose what not to use; with too little you run on assumptions.
• Where’s the line between operational excellence and surveillance? “Ask your legal department.” Depends on culture, appetite, packaging/branding. If you’ve defined goals and “good,” reached operational excellence with no friction, you can stop collecting for its own sake — but you don’t know what you don’t know, and seeing patterns/opportunities requires visibility. “Is that surveillance or observability? I like to say observability.”

People, Companies & References Cited

• Kovács — chief AI strategist, co-founder AI Enablement Academy (learn@aienablement.academy).
• Clara Hermesz — ex-Facebook/Meta, ex-Uber.
• 360 Social Impact (Seattle nonprofit), Agentics Foundation (Friday Hacker Space).
• Meta/Zuckerberg, Salesforce/Benioff (Slack), CrowdStrike Falcon, Klarna, IKEA, Jack Dorsey / Block, Amazon.
• Concepts/tech: ontologies, temporal knowledge graphs, provenance, FAIR, BM25, dense embeddings, hybrid search, fusion + re-ranking, permission-first memory, garbage collection, Karpathy’s auto-researcher, design-thinking flywheel, hipster/hustler/hacker, ISO 56000.

Video: https://www.youtube.com/watch?v=G2PdKKYzrH8 — Transcript via yt-transcript.sh; outline generated from the transcript.