What AI email triage can and cannot do

There is a lot of noise about AI and email, most of it promising an assistant that reads everything and answers on your behalf. Having built an extraction product, we think that promise is both further away and less desirable than it sounds. Here is the honest version.

What models are genuinely good at

Recognising that an obligation exists

Distinguishing "here is the report you asked for" from "please review and confirm by Thursday" is a classification problem with abundant signal. Models handle it well, across languages, including the indirect phrasing common in professional correspondence where the request is buried in courtesy.

Compressing a thread to its current state

A fourteen-message thread usually resolves to one or two sentences of actual state: what was agreed, what is outstanding, who owes what. Producing that summary is the single highest-value thing a model does with email, because it eliminates the re-reading cost entirely.

Extracting structured detail

Dates, amounts, reference numbers, named parties — pulling these out of prose is reliable and easy to verify at a glance, which is exactly the property you want in an automated step.

Where it falls down

Priority without context

A model can see that a message is urgent in tone. It cannot know that this particular client is in a renewal negotiation, that the sender always writes urgently, or that the deadline moved in a conversation that happened over the phone. Priority is a judgement about your situation, not a property of the text.

Anything irreversible

Sending, deleting, and committing on your behalf all share the same problem: the failure mode is unrecoverable and lands on your reputation. A 95% accurate triage assistant is genuinely useful. A 95% accurate autonomous sender is a liability, because you remember the one in twenty.

Missing context that was never written down

A great deal of what makes an email important happened in a meeting, a WhatsApp message, or a hallway. The model sees the text and nothing else, and it cannot know what it is missing.

The useful question is not how much a model can do, but which mistakes you can afford to have made automatically.

Why we require human sign-off

Taskify does not send email on your behalf. It surfaces what needs your attention, drafts nothing that leaves without your approval, and requires you to act. This is a deliberate constraint, and it comes from a simple observation about error costs.

When triage is wrong, you see an item you did not need to see and dismiss it in a second. When autonomous action is wrong, a client receives something you did not write and would not have sent. The two error modes are not comparable, and building as if they were is how AI products lose trust.

How to evaluate a tool in this category

1. 1Ask what happens when it is wrong. If the answer involves an apology to a client, the automation is too aggressive.
2. 2Check whether it shows its source. Every extracted task should link back to the message it came from, so verification takes one click.
3. 3Test it on your worst thread — long, mixed-language, multiple forks. Demos use clean examples; your inbox does not.
4. 4Confirm what leaves your device before you evaluate accuracy at all. A tool that is accurate and indiscreet is not a trade worth making.

The realistic promise is not an inbox that runs itself. It is an inbox where you stop spending two hours a day deciding what matters, because that decision has already been drafted for you and takes a glance to confirm.