How to lead AI development inside large corporations

Executive overview

Most corporations treat AI as just another technology to bolt onto existing systems. It is not. The shift from deterministic to probabilistic programming is a deeper paradigm change than the move to the cloud — it changes how software is built, not just where it runs.

The path through this is to operate like a pioneer: carve out a protected space inside the corporate structure, start with tiny constrained bets, and build confidence through early wins before asking for scale.

Tight constraints, a portfolio of small bets, and values-aligned teams produce better AI outcomes than big budgets and long timelines.

The paradigm shift: deterministic vs probabilistic programming

• Traditional code is deterministic: intelligence is baked in at the point of writing.
• Probabilistic programming lets you inject intelligence at the point of execution, through prompting.
• This unlocks what feels like infinite use cases — the brittle tendrils of legacy code collapse into a flexible layer.
• The tradeoff: you give up certainty of output, but gain human-level cognition in decision making.
• LLMs have a zone of competence — keep tasks within it and hallucination risk drops to near zero.
• That zone expands over time; widen your usage as the models improve, not before.

Operating as a pioneer inside a corporate

• Negotiate a protected space with cybersecurity, legal, and governance before starting.
• Act as the mayor of a patch of land: decide what shops (modules) are needed, then source founders (startup leads) to run each one.
• Behave like a VC toward internal startups: fund them, align on method, then step back.
• Walk the track before racing it — co-develop pseudocode with each team, agree on it, then hand over execution authority.
• Once the pseudocode is signed off, own the risk of success or failure; don't revisit implementation decisions.

Portfolio approach to managing uncertainty

• Structure the program as a basket of objectives, not one big bet.
• Non-negotiable foundations (e.g. the data layer) are run like the US Army — structured, well-resourced, predictable.
• Pioneer teams are run like startups — independent, flexible, fast.
• Measure portfolio ROI over six-month windows, not quarterly; the basket compounds.
• If a single drill finds no water, it doesn't kill the colony.

Generating buy-in and starting small

• Spend the first six months understanding how people actually feel about AI — not just what executives say the strategy is.
• Unwritten culture and individual biases matter more than printed strategy documents.
• Start with proof of concepts, not platform pitches; a smaller ask buys more patience.
• Pick the highest-value part of the business first — early wins in a visible area unlock the next round of funding.
• Constrain timelines deliberately: if the company expects 18-month projects, propose three months.
• Don't propose incremental gains; articulate transformatively large outcomes or go back to the drawing board.

Culture and team structure

• Drop role labels — in the frontier, everyone solves whatever the frontier demands.
• Raise the bar across the divide: business people must be computer-literate; engineers must understand the domain.
• Teach deterministic programmers probabilistic thinking, and vice versa.
• Agile in this context means agile in the English sense — not the methodology; pioneer project managers cannot run Gantt charts three years out.
• Constraints forge better teams: the best work consistently comes under pressure and limited resources.

Three pieces of advice for corporate AI leaders

1. Be humble and empathetic — learn the organisation deeply before trying to change it.
2. Be the bravest person in the room — people will look to you as a prophet; small visions signal fear.
3. Accept that this time really is different — probabilistic programming is a generational shift, not another cloud migration.