A WEBSITE DIDN’T MAKE THE MALL AN INTERNET COMPANY.

THE SHOPPING-MALL TEST

Andrew Ng wrote the AI Transformation Playbook from his experience leading Google Brain and Baidu's AI Group, and he often makes its central point with an analogy that has held up well: in the internet era, a shopping mall that built a website did not become an internet company. An internet company is organized to do what the internet does well, with constant A/B testing, short release cycles and product decisions pushed down to engineers and product managers. The same test now applies to AI: buying ChatGPT subscriptions doesn't make a company an AI company. What counts is whether the company is organized around what AI does well, and that gets decided in the org chart. Procurement can't settle it.

FIVE MOVES, PARAPHRASED

• First projects are for momentum. Ng's advice is to pick early projects for their odds of success over their size: meaningful enough to convince the organization, feasible enough to show traction fast, with a measurable business objective set up front. Early wins start the flywheel, and the first project's size matters less.
• Capability moves in-house over time. External partners accelerate the start; a durable advantage needs internal capability — a centralized AI function that serves every division and sets standards, rather than scattered experiments per department.
• Training is set per role. The playbook prescribes hours by audience: 4+ hours for executives (what AI can and can't do, strategy, resourcing), 12+ for division leaders (setting direction and tracking AI projects) and 100+ for engineers retraining for AI work. A lunch-and-learn is not an adoption strategy.
• Data needs a strategy. Unifying the silos helps, but Ng's sharper warning is about the opposite mistake: a big pile of data is not automatically valuable, and companies waste money collecting low-value data, or even buying other companies for data that turns out to be useless. Bring AI expertise in before the data spend.
• Communicate on purpose. Staff who fear automation and outsiders fed on hype both slow adoption down. Ng's last step is a deliberate communication plan for employees, investors, customers and regulators, because silence leaves the field to fear and hype.

THE TIMELINES AGED; THE ORG LOGIC DID NOT

The playbook dates from December 2018, before the generative-AI wave, and one number shows its age: Ng budgeted 6–12 months for a first project to show traction. With today's models and integration tools a first agent can reach production on live data in weeks, and the hard part has moved from ML engineering to process design. Stanford's 2026 study of 51 deployments found that similar use cases took weeks at one company and years at another, depending on executive sponsorship, existing processes and how willing end users were. Take out the timelines and the organizational logic holds: momentum first, capability in-house, training per role, data as strategy. Our five-phase protocol (Audit → Design → Implement → Operate → Optimize, run as a loop) applies the same logic to agents, with training built in as its own service line. Stanford's numbers show why: in 77% of those deployments the hardest problems were process, data and change management.