Planning AI Applications

Question 1

Which motivation for adopting AI carries the highest risk of not acting?

Accepted Answer

Existential risk to the business — If competitors with AI would make your business obsolete, the investment is justified even when the return is hard to forecast — the alternative is not having a business.

Question 2

In Gartner's 2023 survey of 2,500 executives, roughly what share cited business continuity as their reason for adopting generative AI?

Accepted Answer

7% — It is the strongest motivation and the rarest — most adoption is driven by opportunity or by not wanting to fall behind.

Question 3

On Apple's framework, Face ID is critical while Gmail Smart Compose is complementary. What does that difference set?

Accepted Answer

The accuracy bar and the need for a fallback — A wrong Smart Compose suggestion is ignored; a wrong Face ID answer locks you out. The more critical, the higher the reliability bar and the more you need a fallback path.

Question 4

A proactive feature such as Google Maps traffic alerts differs from a reactive chatbot mainly because:

Accepted Answer

It can precompute, but faces a higher quality bar — Nobody asked for it, so latency stops mattering — but an unsolicited wrong suggestion is an interruption, and users punish it harder than a wrong answer they requested.

Question 5

Placing a feature at the "dynamic" end of the static-dynamic axis commits you to:

Accepted Answer

An update pipeline after launch — Dynamic means the model keeps learning after you ship — fresh data, scheduled retraining or re-indexing, and a rollback. Most of the cost moves to after launch.

Question 6

In Microsoft's Crawl-Walk-Run framing, the Walk stage means:

Accepted Answer

The model serves internal employees — Walk puts the model in front of your own employees, so mistakes stay inside the building and the people hitting them can describe what went wrong in actionable terms.

Question 7

Microsoft's example gate for promoting a support assistant was:

Accepted Answer

95% of AI-suggested responses used verbatim — A promotion gate has to be a number measured on real traffic. "Used verbatim" is checkable; "the team is happy with quality" is not.

Question 8

Why is the Crawl-Walk-Run ladder also a rollback plan?

Accepted Answer

Every rung you climbed is one you can drop back to — When quality regresses, dropping from Run to Walk keeps the feature alive while you diagnose. Teams that jumped straight to Run can only turn it off.

Question 9

Of the three classic moats, which did foundation models weaken most?

Accepted Answer

Technology — When everyone builds on the same handful of models, the model is your supplier rather than your advantage, and capability leads close on someone else's release schedule.

Question 10

Distribution as a moat mainly favours:

Accepted Answer

Incumbents who own the user surface — An incumbent can ship the same feature to an existing audience at zero acquisition cost, which is why a better standalone tool so often loses.

Axis	Low end	High end	What it sets
Complementary → Critical	Smart Compose	Face ID	The accuracy bar, and whether you need a fallback
Reactive → Proactive	Chatbot	Traffic alerts	Latency vs quality: proactive can precompute but is punished harder when wrong
Static → Dynamic	Photo tagging	Face ID	Whether you are committing to an update pipeline

Key insight

Plan an AI feature

Test your knowledge

Which motivation for adopting AI carries the highest risk of not acting?

The short version

In more depth

Planning AI Applications

1. Why build it — three motivations, by risk

2. Where it sits — Apple's three axes

3. How much autonomy — Crawl, Walk, Run

4. Can you defend it — three moats

5. Good enough — the usefulness threshold

Interview Questions

References & Further Reading

Love this visualization?