Leading AI adoption
Everyone at Mangrove, from the board to the store managers, wants to know what the company is doing with AI. This lesson shows you how to lead it as CTO: find where AI moves a business number, run a small portfolio of bets with owners, write a use policy people can follow, govern the risks, and measure value honestly.
By the end you can
- Map where AI could change Mangrove's business and its technology department, and pick the few that move a business number
- Run a portfolio of AI bets with owners, baselines, cost caps and stop rules
- Write an AI use policy built on data classes, approved tools and human review
- Set up governance with an AI register and risk tiers, using the NIST AI Risk Management Framework
- Measure value with a baseline and a controlled rollout, and lead the skills and change work
The everyday picture
When pocket calculators became cheap, schools had to decide what to do about them. A school that banned them fell behind. A school that handed one to every student and reported "100% calculator adoption" learned nothing about whether its students got better at maths.
The schools that did it well made a few clear choices. Calculators were allowed in some exams and not in others. Students still had to show their working, so a teacher could check the reasoning behind the answer. Teachers learned how to teach with them. And the school judged the change by exam results and by whether students could still estimate an answer in their heads, never by how many calculators it had bought. Leading AI adoption is the same job: decide where it helps, set rules for its use, keep a human check where mistakes matter, train people, and measure results instead of usage.
Where AI changes the business and the department
Start where the strategy does: theme 5 of Technology strategy is "AI where it moves a business number". So the first step is a map of business numbers that AI might move, not a list of AI tools. The method in Picking the right problems for AI applies at company level: look for tasks that are frequent, take real time or money, involve reading or writing text, images or patterns in data, and where a wrong answer can be caught or is cheap.
At Mangrove, four business areas stand out, plus the department itself (the business figures marked "illustrative" are for the course):
- Customer support. About 12,000 support chats a week (illustrative), with a median handling time of about 9 minutes. A model can draft replies from the order history for an agent to check and send. Number: handling time and cost per chat at the same quality.
- Demand forecasting. About 9% of app orders have an item cut for being out of stock, and the 80 new-city stores will have no sales history. Machine-learning forecasts are not new, but better models and data can help. Number: item cuts per order.
- Store operations. Receiving one supplier delivery takes about 51 minutes, about 21 of them on workarounds. A model that reads a photo of the delivery note and fills in the goods received note could cut that. Number: minutes per delivery and corrections later.
- Partner ordering. About 1,500 partner orders a week still come by phone, and about 1 in 12 is corrected after delivery. Reorder suggestions in the Partner app could move orders into the app. Number: the app's share of partner orders and the correction rate.
- The technology department. AI coding assistants are already in daily use by most engineers (Working with AI coding assistants). Number: delivery and stability measures, not lines of code (Quality, debt and AI tools).
For each candidate, write four lines: the business number and its baseline, the size of the prize (what moving it is worth a year), the risk if the AI is wrong (who is harmed and how badly), and readiness (is the data there, is there an owner who wants it). Then pick three to five. Prefer the ones where the prize is large, a wrong answer is caught by a human or cheap to fix, and the data exists today.
Quick check
Which candidate should most likely go first at Mangrove?
Reply drafts are frequent, have a clear baseline, and a human catches mistakes before a customer sees them, so the risk is low and the value is measurable. Automatic credit decisions carry high risk if wrong and need much more governance first. A chatbot "for innovation" has no number to move. Licences for everyone measure usage, not value.