When AI features became easy to add, my product backlog grew very quickly.
Automatically translate every incoming message. Generate replies before the user asks. Detect customer intent. Choose a saved script. Send a follow-up when a conversation becomes quiet. Summarize every chat. Score every lead.
Each idea sounded reasonable in isolation. Together, they would have turned a small WhatsApp Web utility into a system constantly reading, predicting, and interrupting.
I removed or narrowed several of those ideas. The experience changed how I evaluate “smart” product features.
Automatic translation sounded better than on-demand translation
The first idea was simple: whenever a new message appears, translate it immediately.
That removes a click, which is normally a good thing. It also means processing every message, including messages the user already understands, one-word acknowledgements, emojis, addresses, product codes, and conversations that are not currently relevant.
Automatic translation creates cost, visual noise, and a larger data-processing surface. It can also make the original message harder to scan because every bubble suddenly contains twice as much text.
The narrower design worked better: place a translation action beside the message and remember the preferred language for the conversation. The user asks when translation is useful; the interface stays quiet when it is not.
One click remained. A large amount of unnecessary processing disappeared.
Pre-generated replies created false urgency
Another idea was to generate a suggested reply as soon as a customer message arrived.
This looked efficient in a prototype. In practice, it subtly changed the relationship between the user and the tool. A waiting suggestion made the user feel that the next step was to approve or edit the AI's answer, even when the better next step was to ask a colleague, check inventory, or leave the conversation unanswered for a moment.
The AI was not just saving writing time. It was setting the agenda.
I changed the flow so reply generation begins only after the user requests it. That small decision preserves a useful boundary: the person decides that a reply is appropriate, then the model helps draft it.
Intent detection was less useful than better organization
Automatically labeling a message as “pricing,” “shipping,” or “support” feels intelligent. It can also become another classification system that needs correction and maintenance.
For a small tool, the more immediate problem was often simpler: users could not find the reliable answer they had already written.
A searchable script library solved more of that problem than automatic intent detection. Users could save approved replies, organize them, and insert them when relevant. AI could then help adapt the wording without becoming the source of truth for prices or policies.
This feature looked less advanced. It was easier to trust.
Fully automatic sending failed the reversibility test
The most obvious “smart” feature was also the one I rejected most strongly: let AI read the conversation, decide on an answer, and send it without review.
The problem is not only model accuracy. It is reversibility.
An awkward draft can be edited. A wrong translation can be discarded. A message sent to a real customer cannot be unsent in any reliable business sense. Even if the interface offers a delete option, the recipient may already have read it or acted on it.
The higher the cost of reversal, the more important human confirmation becomes.
MSG.AI therefore produces editable drafts. The user remains responsible for the final message and the final click.
AI features have hidden interface costs
Developers often estimate an AI feature by its API call and prompt. The user pays additional costs that are harder to measure:
- Another button to understand
- Another loading state to wait for
- Another result to verify
- Another setting to configure
- Another failure mode to interpret
- Another category of data processing to explain
A feature can save thirty seconds of typing and add forty seconds of checking.
This is especially true in multilingual business communication. A fluent-looking answer feels finished even when it contains a subtle factual mistake. The better the writing quality becomes, the more deliberate the verification process needs to be.
I now use three tests before adding AI
I evaluate new ideas with three questions.
1. Is the model reducing creation work or making a business decision?
Drafting, rewriting, translating, and summarizing are creation tasks. Setting a price, promising a date, choosing a recipient, or deciding whether consent exists are business decisions.
AI can assist with both, but the second category requires much stronger controls and often should remain outside the feature entirely.
2. Can the user inspect the input and output?
A useful AI feature should make its context understandable. If the user cannot tell what conversation, saved knowledge, or language setting influenced the result, correcting a mistake becomes guesswork.
Visible context and editable output are more valuable than a mysterious confidence score.
3. What happens when the feature is wrong?
If the result can be discarded with no consequence, experimentation is inexpensive. If it changes a customer record or sends a message, the design needs confirmation, logging, and a safe failure state.
The question is not whether the model will ever be wrong. It will. The question is whether the product contains that error before it becomes an external action.
Removing intelligence can make a product feel smarter
After simplifying these features, MSG.AI did fewer things automatically.
It also became easier to understand. Translation appeared when requested. AI replies began with an explicit action and ended as drafts. Saved scripts held stable business information. Bulk tasks required recipient review.
The product felt calmer because it stopped competing with the user's judgment.
That is the paradox I keep returning to: the smartest product is not necessarily the one that predicts and automates the most. It may be the one that knows when to wait.
AI is most useful when it removes effort without removing agency.
Project page: MSG.AI.
