Notifications Could Be Smarter with AI. So Why Aren’t They?

Notifications were supposed to make life easier – keeping us informed without demanding too much attention. Instead, they’ve morphed into digital noise, pinging us at 3 AM about package deliveries or flooding our screens with “personalized” messages that feel anything but personal.

It’s no wonder, then, that only 12 percent of notifications actually arrive at the right moment. The other 88 percent? They’re withdrawals from a trust account most users don’t even realize they’re maintaining.

This tension set the stage for a discussion at the Shift conference in Kuala Lumpur this November, where Tanisha Sharma (AI Developer Relations Engineer, SuprSend) explored how, when done right, notifications can genuinely simplify life for users and become a powerful promotional tool for companies – and why, in practice, they often end up becoming the exact opposite for both sides. 

Don’t spend trust on bad alerts!

“Think of every notification as a transaction – not of data, but of trust,” says Tanisha:

When you send a message at the wrong time, you’re spending that trust. When you respect quiet hours, understand user context, and deliver something meaningful, you’re depositing trust back. Most companies are overdrawing. 

And personalization? That’s another myth.

“Slapping a name on a message isn’t personalization; it’s mail merge circa 1995,” she says.

Real personalization means knowing when a user sleeps, what device they prefer, and whether they’ve enabled Do Not Disturb. It means understanding intent: did they click because they cared, or because they just wanted the red badge gone?

The truth is simple, though a bit painful:

Activity doesn’t equal interest, behavior doesn’t equal concern and curiosity isn’t loyalty. But these flawed assumptions still run notification strategies everywhere – leaving users drowning in alerts and companies chasing phantom engagement.

How do we get notifications right?

Here’s the model:

• It starts with a trigger – something happens, like a user resetting a password or moving a task.

• Next is context: where is the user, what device are they on, which channels do they prefer, and do they have quiet hours or Do Not Disturb active? Notifications should be paused if needed.

• Then comes policy: rules like no notifications during quiet hours or limiting messages to three per day (or whatever suits your product).

• AI can optionally help by tracking preferred channels, fallback options, and user intent.

• A human override is crucial for urgent cases, like password resets, to ensure important notifications reach the user.

• Next, consider channels: email, push, SMS, or other systems – send through the channel the user prefers.

• Timing also matters: who receives it first and when.

• Finally, delivery and feedback: track if the user opened the notification, disabled it, or uninstalled the app. Logs and audit trails ensure transparency and accountability.

AI assists, but rules govern the system.

AI handles the boring stuff, users benefit, developers win

Tanisha then moved on to APIs – specifically, the human API.

Tanisha explains, “Every day, we manage user IDs, usernames, templates, priorities, and fallback options. Each element is captured and processed to ensure notifications reach the user accurately. This system has worked for years, but it’s 2025 – the AI era.”

We no longer have to rely solely on the human API; it’s time to make it AI-friendly.

By explicitly specifying intent – for example, sending a notification about the “order of the day” – the AI agent can handle the rest: choosing the right channel, managing fallbacks, and sending at the correct time. It can even respect quiet hours and adjust delivery accordingly.

AI has the power to manage these tasks efficiently, reducing friction for both developers and users. Users receive fewer irrelevant notifications, and developers spend less time on manual handling.

It’s important to note that the agent-ready API is not meant to replace developers. Instead, it assists them, performing repetitive tasks faster, better, and more efficiently – a true win-win. Both human and AI APIs can accomplish the same objectives, but the AI-enhanced approach makes the system smarter and more scalable without adding extra burden.

Keep notifications safe and smart – with a human backup

“If AI is involved, notifications must be safe and protected by strict safeguards,” says Tanisha.

There should always be human override. For example, if a user has DND from 11 PM – 6 AM but is waiting for an important notification, the human override ensures they still get it. Critical transactional notifications should always bypass DND or other restrictions.

She reminded the audience that AI-generated templates can be really helpful, but they’re not perfect. That’s why someone needs to review and approve them before they go out. Notifications should also respect the user’s context, preferences, and quiet hours so people don’t feel bombarded. Ideally, users should only get a few alerts a week, with one main notification for anything truly important.

Of course, things can go wrong, so having good logs and monitoring is key. Users should be able to see why a notification didn’t arrive, and the system should track delivery, performance, and interactions so nothing slips through the cracks.

Redundant messages are another no-go. Sending duplicates or overly similar alerts frustrates users, so notifications and APIs need to work smoothly for both AI and humans.

Finally, multi-channel support and fallback options are a must. If one channel fails, the notification should automatically go through another, with all attempts logged to keep things reliable and maintain user trust.

And if your users are developers, provide them with the ability to trace notifications, check logs, and debug easily. Integrating notifications into developer workflows improves efficiency.

Key lessons from SuprSend

In the end, building an effective notification system comes down to a few key principles:

• Always consider the context – know why and when a notification should be sent.
• Fallback mechanisms are essential to ensure important messages reach users, and thorough logging and observability help track performance and troubleshoot issues.
• Avoid overwhelming users with too many alerts, but also make sure critical notifications aren’t missed.
• Human override should always be possible, giving a safety net for urgent cases.

Even a simple task management app can demonstrate these ideas: handling preferences, channels, and overrides efficiently allows users to stay informed without frustration. The goal is clear: respect user preferences, maintain safety, keep logs, handle failures gracefully, and provide developer-friendly tools.

As Tanisha puts it: “Every notification is a tiny trust transaction – spend it wisely!”