What happened

NN/G's July 10 piece names five design dimensions for site-specific AI chatbots. First and least glamorous: handoff willingness. Never gatekeep when a user explicitly asks for a human, and escalate when the bot clearly can't help or the user shows frustration signals.

The middle three: flexibility within guardrails — handle adjacent questions, admit limits instead of faking omniscience; proactivity in two flavors — clarifying questions that tailor an answer, and directional nudges toward next steps, offered only after the user's actual goal is handled; and emotional responsiveness, which NN/G defines carefully. Acknowledge the customer's situation. Don't pretend the bot has feelings.

The fifth, transparency, has four parts: identity, capability, rationale, and privacy — explained in the conversation itself, not buried in a policy page. Identity is the one with teeth. NN/G notes AI-identity disclosure becomes legally required in the EU starting August 2026.

Why this matters

Chatbot vendors sell you the clever parts — personality, upsell prompts, product recommendations. NN/G's list says the boring parts carry the trust. A bot that traps a frustrated shopper in a loop with no exit doesn't just lose that conversation; it teaches the customer your support is a wall. NN/G's framing is that every interaction builds trust or erodes it, and on a storefront, erosion has a dollar sign attached.

The EU date turns one of the five into a compliance item rather than a design preference. If you sell into Europe and your assistant isn't clearly labeled as AI, you have weeks, not quarters. Disclosure built as the default costs one sentence in the opening message; disclosure retrofitted under deadline costs a scramble.

The proactivity guidance has a commerce reading worth spelling out. NN/G says directional nudges come only after the user's primary goal is handled — which is the polite way of saying your bot shouldn't cross-sell a shopper who's mid-complaint about a late order. Answer first, then suggest. Most storefront bots ship with that order reversed because the vendor's demo optimizes for upsell, not resolution.

One more NN/G observation worth sitting with: users still don't have clear mental models of what a site-specific chatbot can do. Your bot's first message is setting expectations whether you designed it to or not.

What to do about it

Wire the human handoff before anything clever

Decide where "talk to a human" routes — live chat, a ticket, a callback — and then test the phrase literally. If the bot deflects even once when asked directly, that's your first fix under NN/G's never-gatekeep rule.

Add the AI disclosure now

Put it in the opening message and keep a persistent label on the chat widget. The EU requirement lands in August 2026; shipping the disclosure this month makes the deadline a non-event instead of a fire drill.

Script the frustration escalation

Define the triggers — repeated rephrasing, explicit complaints, a failed answer on the second attempt — and route them to the human path automatically. Waiting for an angry customer to ask is already a loss.

Write the capability statement

One short paragraph: what the bot can do (order status, returns, product questions) and what it can't. Use it in the opener. It shapes expectations up front and cuts the dead-end conversations that erode trust fastest.