The AI-Native Enterprise

1 · The pilot problem comes first

The single most important step happens before any training: leadership — not delegates — chooses one real, recurring, cross-department problem that costs money every cycle, writes it down with an honest cost estimate, and names the one metric that would prove it solved. Everything in the rollout aims at that pilot.

A program aimed at a named, costed problem lands as transformation. The same program without one lands as a workshop people politely forget. The other Phase-0 move that decides everything: the staff announcement is written as a promotion, not a threat — judgment, relationships and accountability stay human; what gets systematised is the repetitive weight around them. Companies that say this out loud get adoption; companies that don't get quiet resistance no training fixes.

Do not proceed past phase one without the pilot problem written down and filed.

2 · Foundation before network — nodes are made of people

The architecture is three layers. First, every staff member builds a personal AI operating system: their recurring tasks written as plain-language processes, then converted into .md skill files an AI executes to their standard. The exam is the cold handoff — a colleague runs your skill with no explanation. If it runs for someone else, it's real, and the company just gained documentation, onboarding material and continuity insurance in one artifact.

Second, champions — identified by observation during training, never appointed by seniority — consolidate each department's skills into a governed node: deduplicated, one source of truth, indexed, with a review gate. Third, nodes link into a company work network through written handoff protocols: what marketing hands finance, in what format, to what standard, captured as files.

The payoff is adaptability: reorganise a team and you re-point references. The knowledge no longer leaves with anyone. And everyone builds inside one company-standard folder skeleton from day one — the small administrative discipline that makes all the consolidation possible.

A work network with nothing trained underneath it stays a diagram. Foundation first, every time.

3 · Privacy is a routing decision, not a blocker

Most enterprise AI fear dissolves once data is explicitly sorted into three lanes. GREEN: public or harmless — any model. YELLOW: internal, needs contractual cover — cloud models under business terms. RED: personal data, payroll, health, anything regulated — never leaves the building, local models only.

The lane is written into each skill file as a DATA LANE line, so the routing travels with the skill instead of living in someone's memory. And the RED lane is workable because free tools like Ollama run quantised models on hardware you already own — an 8 GB laptop runs a small model that is genuinely sufficient for structured skill execution. The proof ritual: run a RED-lane skill with the network cable out.

The rule that never bends: "just this once" to the cloud with RED data is how companies end up in the news.

Local models are sufficient, not frontier. Their job is privacy and free precompute — and they do that job completely.

4 · The decomposition method — the keystone

The reframe the whole method exists to teach: don't bring AI to your problems — bring your problems to the shape AI can eat. The worked example: searching 10,000 video files. The naive design runs an expensive vision model over video every time anyone searches. The systems design extracts one frame every 30 seconds, labels the frames once with a free local vision model, and from then on every search is a cheap text query. One preprocessing pass turns an expensive recurring cost into a free one.

The shape to memorise: expensive-per-query → cheap-once-then-free. It generalises through a one-page sheet: name the expensive operation, ask what can be precomputed, pick the cheapest sufficient model per step, decide where the data should live, and cost the before/after with a payback period.

People who learn this shape stop asking "which AI tool should we buy?" and start asking "what's the expensive operation, and can it be done once?" That question, asked quarterly, is the culture change.

"Sufficient" is the standard when picking models, not "best". The savings live in that word.

5 · Ship the pilot, then let the network do the onboarding

The graduation is a live demo: the pilot problem from day one, decomposed with the sheet, built across at least two linked nodes, privacy-routed correctly, and run end to end in front of leadership with the before/after metric. The pilot's design document becomes the company's first internal case study — the template for every problem after it.

Sustaining it is deliberately light: fix skills the moment an output disappoints (in the file, never just the chat), keep the review gate alive, re-run one decomposition per quarter, and onboard every new hire by having them read their department's node and run two of its skills. The network becomes the onboarding.

The honest annual question that keeps it alive: if our best person left tomorrow, what walks out the door with them? File whatever the answer surfaces.

What matters is not speed. What matters is that the pilot ships.