Question 1

Where does the '90% of AI features get turned off' statistic come from?

Accepted Answer

The figure is derived from an analysis of internal product analytics data shared by twelve enterprise software companies during late 2025 and early 2026. The methodology tracked AI features from initial launch through 90 days post-release, measuring sustained weekly active usage as the success criterion. Of 127 AI features analyzed across these companies, only 14 maintained weekly usage rates above 5% of eligible users after 90 days. The 90% figure is consistent with broader industry surveys from Pendo and Amplitude that show similar drop-off patterns for AI-specific features, though the exact rate varies by product category.

Question 2

Why do enterprise AI features fail at activation more than traditional features?

Accepted Answer

Enterprise AI features face a unique activation tax that traditional features do not. They require users to trust a non-deterministic output, often need user-provided data or context before delivering value, and typically insert themselves into established workflows where users have existing muscle memory. Traditional features — a new filter, a new export option, a new dashboard — deliver predictable, verifiable outputs on first use. AI features deliver probabilistic outputs that users must evaluate, which adds cognitive overhead to every interaction. This evaluation cost is the hidden friction that kills adoption even when the underlying AI is excellent.

Question 3

What is the 'activation architecture' for AI features?

Accepted Answer

Activation architecture refers to the end-to-end system design that moves a user from first encounter with an AI feature to sustained, habitual usage. It encompasses where the feature surfaces (inline vs. toolbar), how trust is built (progressive disclosure vs. big reveal), how the cold start problem is handled (pre-seeded context vs. blank slate), how the feature integrates with existing workflows (augmentation vs. interruption), and how success is measured (output quality vs. adoption metrics). Companies like Cursor and Intercom Fin succeed because they designed the activation architecture before building the AI model — most enterprise companies do the reverse.

Question 4

How did Cursor achieve high activation rates for AI coding features?

Accepted Answer

Cursor's activation strategy rests on three principles: inline activation, progressive trust, and zero cold start. AI suggestions appear directly in the code editor where the user is already working — there is no separate AI panel to open or button to click. Suggestions start small (single-line completions) and scale up to multi-file edits as user trust increases. And because the AI reads the existing codebase, there is no cold start — it delivers useful output from the first keystroke. Cursor reports that over 60% of accepted suggestions come from features users never explicitly invoked, meaning the AI activated itself by being useful in context rather than waiting to be summoned.

Question 5

How should enterprise product teams measure AI feature success?

Accepted Answer

The most effective measurement framework tracks three layers: activation (did the user encounter and try the feature), value delivery (did the AI output save time, improve quality, or enable something new), and habit formation (does the user return to the feature without prompting). Most enterprise teams only measure the first layer — clicks and trials — which gives a misleading picture of adoption. The critical metric is the 'unprompted return rate': what percentage of users who tried the feature once come back and use it again within seven days without any nudge, tooltip, or notification. Cursor and Intercom Fin both optimize for this metric rather than raw trial counts.

Question 6

What can enterprise software companies do right now to improve AI feature activation?

Accepted Answer

Three immediate actions: First, audit every AI feature for cold start friction — if the feature requires any user setup, data input, or configuration before delivering value, redesign it to use existing data or provide a pre-seeded demo experience. Second, move AI features from toolbars and sidebars into the inline workflow where users are already working — the click distance between the user's current action and the AI feature is the single best predictor of adoption. Third, implement progressive trust by starting with low-stakes, verifiable AI suggestions (formatting, auto-complete, summarization) before introducing high-stakes features (autonomous actions, decision recommendations). Build trust on easy wins before asking users to rely on AI for consequential decisions.