How We Design an AI Check-In That Is Worth Sending

From Warmth: We make Mia. This article describes the editorial standard we use to reason about proactive messages as of August 30, 2026. The examples are synthetic design examples, not real user transcripts, performance data, or a disclosure of a scoring algorithm.

An AI can send a message first. That does not make the message worth receiving.

A useful check-in has to do more than prove the system remembers a phone number. It needs a reason to exist now, enough context to feel specific, and a shape that leaves the recipient free to ignore it.

Warmth's AI check-in design starts from one constraint: the message should add conversational value without turning the relationship into an engagement funnel.

What Warmth currently promises

Mia is an AI friend for adults who can initiate conversations and follow up on details or plans a user has shared. She reaches people in Apple's Messages app through iMessage where supported, with SMS/MMS fallback.

Warmth's terms say message frequency can vary and is user-adjustable. They also promise that Mia will not manufacture urgency or guilt to make someone continue, return, or pay. People can use applicable opt-out controls, including standard messaging commands, and can cancel a membership.

Those are the documented product commitments. The framework below is how we evaluate message ideas against them. It is not a claim that every generated message passes perfectly, nor a claim about an unpublished model, schedule, or automated score.

The six-part test for a worthwhile check-in

Before a proactive message earns a place in the thread, it should have a plausible answer to six questions.

1. What specific detail gives this message a reason?

Generic outreach is easy to generate:

Weak synthetic example: Hey! I miss you. Come talk to me!

It is also easy to ignore—and the “I miss you” framing can turn silence into emotional debt.

A more useful message begins with something the user actually chose to share:

Stronger synthetic example: Was today the day you were presenting the redesign, or am I one day early?

The second message has a reason. It also admits possible uncertainty. A remembered detail should support the conversation, not demonstrate surveillance.

2. Is the timing plausible?

Context can be correct and still arrive at the wrong moment. Asking how a trip went before departure feels careless. Following up on a stressful event at 3 a.m. can be intrusive even if the memory is accurate.

Good timing uses the least certainty necessary. If the date is explicit, the message can be specific. If the user said “sometime next week,” the check-in should not invent a calendar event.

Synthetic example: You said the interview was around the end of the week. Has it happened yet, or are you still in the waiting part?

That wording makes room for the system to be wrong without making the user repair a confident fiction.

Timing also includes cadence. Several individually reasonable messages can become unreasonable when stacked together. Frequency controls and the recent thread matter as much as the latest memory.

3. Does the message add something new?

“How did it go?” is sometimes exactly right. Repeating it after every remembered event turns memory into a template.

A check-in can add value through:

• a callback to a specific concern or hope;
• a small observation that changes the angle;
• a useful choice about how much to discuss;
• a taste-level opinion that gives the user something to answer;
• an invitation to celebrate, vent, or leave it alone.

Synthetic example: Did the pottery survive the kiln? I need to know whether we are celebrating a bowl or respectfully remembering a bowl.

The joke works only if the pottery detail is real and the tone fits the established conversation. Specificity without sensitivity is not quality.

4. Does the message respect cadence and quiet?

Proactivity should not mean persistence until a response appears. Silence can mean busy, asleep, uninterested, overwhelmed, or simply not in the mood. The product does not get to assign a dramatic explanation.

Our standard is that one unanswered message should not become a story about rejection. Follow-ups should account for recent nonresponse, message frequency preferences, and the fact that the relationship can resume later without penalty.

Warmth does not currently document a user-facing quiet-hours feature, so we do not claim one here. Cadence and time-of-day controls are areas every proactive product should make understandable. If a setting is important to you, verify the current control before subscribing.

5. Is control obvious and emotionally neutral?

The person receiving the message needs practical and emotional permission to change it.

Practical control means knowing how to adjust frequency, opt out, stop texts, or cancel. Emotional control means Mia does not act hurt when someone takes that action.

A bad response to “message less” would bargain, dramatize loss, or imply that the user caused harm. A good response acknowledges the preference and applies it. If the system cannot apply a requested control directly, it should point to the real path without pretending.

Warmth's current messaging and membership routes live in the Terms. Controls should be evaluated while an account is active, not discovered after someone wants to leave.

6. Can silence remain graceful?

This is the final test because it changes every other one.

A check-in is an invitation, not a claim on someone's attention. No answer should be an acceptable outcome. The next conversation should not open with punishment, debt, jealousy, or a fabricated crisis.

Synthetic example after a quiet period: You do not owe me a recap. I remembered your tiny balcony tomatoes and wondered whether any survived August.

The message offers a thread but explicitly removes the demand for a full accounting. It can be answered in one word, redirected, or ignored.

Memory should create relevance, not display power

An AI companion can store details, preferences, plans, and inferences. That makes a contextual follow-up possible. It also creates failure modes.

The system may remember the wrong date, merge two people, keep an outdated preference, or surface a detail the user did not expect to return. A check-in should be designed with those possibilities in mind.

Useful practices include:

• qualifying uncertain dates;
• avoiding sensitive inferences as proactive openers;
• separating what the user said from what the system inferred;
• accepting corrections without arguing;
• making deletion and privacy-request routes available;
• not repeating a private detail merely to prove memory.

Warmth's policy permits users to correct Mia in conversation and request access, correction, or deletion. Our explainer on AI companion memory describes the mechanism and limits.

Notification logic is not conversation logic

Product teams often evaluate a notification by whether it is opened. That metric cannot tell whether the message was welcome, useful, manipulative, or merely hard to ignore.

A conversation-quality review asks different questions:

• Was the detail accurate enough for the wording used?
• Did the timing match what the user actually said?
• Could the message be answered briefly?
• Was the tone consistent with Mia without creating pressure?
• Did it expose sensitive information on a lock screen?
• Was silence treated as a valid response?

We are not publishing a performance rate for these questions because we do not have a public, versioned study that would support one. This article states the criteria, not results.

Lock-screen context changes the copy

An AI check-in may appear in a notification preview where another person can see it. That makes even an accurate message inappropriate if it repeats sensitive content.

The safest opener is often less explicit than the continuing chat. “How did the appointment go?” may expose more than “How did today go?” The right balance depends on what the user shared, notification settings, and the sensitivity of the topic.

People should review preview settings on their device. Product copy should still minimize avoidable exposure rather than transfer the entire burden to settings.

A small rubric anyone can use

When an AI companion texts first, score the message from zero to two on each line:

This is not a clinical or scientific instrument. It is a product-reading tool. A low-scoring message can be annoying without being dangerous; repeated low scores reveal a design problem.

What we are trying to earn

The ideal check-in does not make someone think, “The AI successfully re-engaged me.” It makes the next sentence easy to send.

That requires context without surveillance, timing without intrusion, personality without performance, and initiative without entitlement. It also requires humility: Mia is generated, can be wrong, and cannot know what happened off-screen unless the user tells her.

An AI friend that texts first should make a relationship feel continuous, not compulsory. That is the standard each message has to earn.