:::tipStages 0-5 were covered in the first part of this article. Read it here: The AI Adoption Curve Nobody Warned Me About - Part 1:::Stage 6 — Hackathons as a distribution channelWe ran a series of internal hackathons. The output was pretty strict: it should be shipped skills and agents that landed in the shared action layer immediately. Production-grade work from solo engineers in 48 hours: a custom course generator running an 8-step safety pipeline behind it, a reflection feature with a 5-layer defense and an internal MCP shipped Day 1.Hackathons at Promova stopped being "team-building" and became the primary mechanism for putting new abstractions into shared infrastructure.Stage 7 — Variance discovery + coachingTelemetry from Stage 2 finally paid back here. We could see who routed tasks well (mixed Sonnet/Opus/Haiku) and who defaulted to the largest model for everything. Our spend distribution sat around 75% Opus when the right mix for our task distribution is closer to 40 / 50 / 10. Correcting that mix alone could cut 40–55% of spend with no headcount change and no adoption hit.What we did: monthly per-user reports surfaced through the dashboards, model routing rules in the org-level CLAUDE.md, optional personal DMs to outliers when patterns looked extreme. No public shaming, no policy-by-spreadsheet. The note to one engineer running 93% Opus on tasks that fit Sonnet was a five-line nudge, not a memo about "AI fluency."This is the part most "AI-ready team" frameworks miss. It's one-on-ones plus visibility, and you only get to that conversation if the telemetry was there from day one.Stage 8 — Multi-provider routingThe current frontier. As Anthropic restructures Enterprise billing (seat fee separate from token billing), OpenAI moves Codex to token metering, GitHub tightens Copilot limits, and Claude Managed Agents add $0.08/session-hour on top of tokens — every model call is a financial decision.We're moving toward a proxy/routing layer (LiteLLM, Bifrost, etc.) that sanitizes keys across providers and lets us route by task type, not by historical default. Same pattern as Stage 1 — but one level up the stack.The parallel track that surprised meEngineers weren't the only ones on this curve. Non-technical people went through the same stages, on roughly the same primitives, at roughly the same pace.Their path: Personal ChatGPT → corporate Claude subscription → discovering skills exist → copy-paste skills, plugins, and workflows from the internet → building their own personalized skills → sharing skills cross-team → integrating with our knowledge and action layers for project memory and shareable artifacts → moving into CLI and contributing to operational flows alongside engineers.Engineering path: Same stages, same primitives, similar pace. Engineers moved faster on three things specifically — writing CLAUDE.md and AGENTS.md, building their own MCP servers, and embedding agents into the dev loop. R&D teams jumped further ahead, going from Claude desktop straight into Claude Code CLI and multi-agent terminals with cross-workspace orchestration. That's the inside track, not the average path.What really surprised me: a technical background didn't accelerate base primitive adoption. Skills, sharing, agentic workflows — these are new abstractions. Everyone started with the same learning curve. The convergence on base primitives happened at roughly the same time across the org.Where the tracks still diverge: building your own agents, designing agentic systems, deciding when a workflow becomes an agent. That stays on the engineering side, probably permanently.A small detour: how everyone learned to promptA year ago, the prompting pattern looked the same for engineers and non-engineers. Imperative, single-context: "Fix this." "Why doesn't it work." "You don't understand." Everything dumped into one chat session, context window stretched until output quality collapsed.Today, the dominant pattern is different. People front-load context in the first prompt. They switch sessions when topic changes. They ask differently — not "do this," but "teach me so I don't repeat this," "where am I making a systematic mistake," "how do I stop this routine from coming back."That shift — from AI as executor to AI as teacher — is the most stage-agnostic signal I've seen of someone moving up the curve. It happens to engineers and non-engineers at the same rate.What I'd tell another company or engineering leaderDifferent roles adopt this differently. Three things apply across all of them.One: Day-one telemetry. Not month-three telemetry, not "once we hit scale" telemetry. Day one. Cost variance and routing patterns are invisible until you can measure them, and you can't fix what you can't see.Two: Split knowledge and action into separate systems. Knowledge layer (graph, search, memory) and action layer (skills, MCPs, automations) have different access patterns and different lifecycles. Treating them as one is the architectural decision most "AI didn't scale" stories quietly retell.Three: Stop measuring AI adoption as a single percentage. Adoption is the start. The interesting numbers are downstream — cost variance per task type, model routing distribution, edit acceptance rate, ratio of shared skills to personal skills, number of non-engineers shipping production code. None of those existed as a job two years ago. All of them matter now.What's nextWe're working on three things in parallel: defining where agentic workflows end and true agents begin, token economics (caching, prompt reuse, eval), and cross-team monitoring of CLAUDE.md / AGENTS.md setups across our repos.If you're somewhere on this curve and seeing something different — or the same — I'd like to hear it.\n \