Most corporate AI strategies fail in the same two places: the data covenant was never renegotiated, and the program was scoped against the org chart instead of a P&L line. This guide is the working playbook I use with operators in steel, chemicals, logistics, and cross-border trade — the order of operations that moves a company from strategy theatre to industrial reality.
Before a single model is fine-tuned, the operator has to renegotiate the implicit contract between IT, OT, and the line. Who owns the historian? Who can write to the MES? Which sensors are trusted, which are folklore? Most enterprise AI programs fail not on the model — they fail because nobody had legal, technical, or political authority over the data the model needed.
A corporate AI strategy that tries to serve every function ends up serving none. Pick the single P&L line where a 3–5% move changes the company's trajectory — yield, uptime, working capital, landed cost — and route every AI investment through that line for the first 18 months.
Frontier models are a buy. Vector search is a buy. The judgment of which exception to escalate to a 30-year planner — that is build, and it's the only place defensibility lives. Most enterprises invert this: they build commodities and buy their judgment from a vendor.
Federated, centralized, or hub-and-spoke — pick one and write down the swim lanes. Who is accountable for the model in production? Who funds GPU spend? Who carries the pager when accuracy drifts? Ambiguity here is the single largest predictor of program death by month nine.
A model accuracy number is not a business metric. Translate every initiative into hard-currency: tons of throughput, basis points of margin, days of cash. The CFO must be able to defend the AI line in an analyst call without flinching.
Every quarter, two things must happen: one pilot crosses into production with an SLA and a budget, and one pilot is killed publicly. Programs that cannot kill their own pilots accumulate technical debt until the next CEO inherits the cleanup.
A 90-slide AI strategy deck without a named accountable owner, a P&L line, or a kill criterion. It survives one board meeting and dies in the next budget cycle.
Fourteen proofs of concept, zero in production. The org has learned the vocabulary of AI without ever rewiring a single operational decision.
A Center of Excellence with a headcount, a logo, and no P&L. It writes guidelines, runs Lunch & Learns, and is the first to be cut in a downturn.
The strategy is whatever the largest hyperscaler's account team last pitched. The operator no longer owns the roadmap — the vendor does.
Map the data covenant. Interview the CEO, plant managers, the controller, and the two engineers who actually know how the line works. Output: a single page naming the P&L line, the constraint, and the 3 candidate use cases.
Score candidates on value-at-stake × feasibility × data readiness. Pick 2 to fund, 1 to kill loudly. Output: a board-ready memo and a 90-day SLA for each surviving bet.
Run pilots on the real line — not a sandbox. Instrument for hard-currency metrics. Output: one pilot in production behind an SLA, one pilot killed with a written postmortem.
Lock the build/buy/partner posture, the federation model, and the pager rotation. Output: a written operating model the CFO will defend and the COO will run.
I take a small number of advisory mandates each quarter — diagnosis, bet selection, and the first industrial pilot. If the shape above matches the room you're in, write.
Hey — I'm Aarya. Ask me about Vik's background, the consulting practice, AI Pulse, or anything on this site. I'll keep it punchy.