Turning email into tracked work
Why marking a message unread is not a task list, what separates a real commitment from a mention, and how to judge a tool that claims to extract them.
4 min read
Most of the work a business owes its customers is agreed in email and tracked nowhere. Somebody says they will send revised numbers by Friday, the thread moves on, and the only record is a sentence in the middle of a message nobody will scroll back to.
The usual coping mechanism is marking the message unread. That is not a task list. It has no due date, no order, and no way to tell a promise you made from a newsletter you have not opened. It also disappears the moment something newer arrives.
Why copying it out never happens
Every task tool solves this by asking you to retype what the email already said. Open the message, open the tool, write the task, set the date, link it back if the tool supports that at all.
That step is the whole cost, and it is the step that gets skipped on the days you most need it. A system that only works when you are calm is not a system.
What actually counts as a task
The interesting problem is not extraction. It is judgement. Most sentences that mention work are not commitments.
"I will send the proposal tomorrow" is a task. "Once we have the proposal we can move" is not, it is a condition. "Did you get the proposal?" is a question, which may need a reply but is not work you owe. A keyword search for "proposal" cannot separate those three, because the difference is not in the words. It is in who is obliged and when.
That is why this is a reading problem rather than a filtering one, and why the failure mode is different from what people expect. A tool that over-extracts does not produce a slightly noisy list. It produces a list you stop trusting, and an untrusted list is worse than no list, because you now check it and the inbox.
Dates are where models are weakest
There is a specific technical point worth knowing when evaluating any of these tools.
Language models are unreliable at date arithmetic. Ask one what date "end of next week" refers to and it will answer confidently and often wrongly. A tool that lets its model produce the actual due date will quietly generate wrong deadlines, which is the worst possible failure for a system whose only job is remembering when things are due.
The right shape is to split it: the model extracts the hint as written, and ordinary code resolves the hint into a date. The model does the reading, which it is good at, and arithmetic does the arithmetic.
A related test: ask what happens to a vague hint. "Soon" has no date. A well built system produces a task with no date rather than inventing one, because a confident wrong deadline is more damaging than a missing one.
Acting only when sure
The other thing worth asking any vendor is what their tool does when it is not confident. Most marketing pages argue their AI is accurate. Very few show what happens when it is not.
CevroFlow answers this with a fixed threshold rather than a setting:
The task is created
Clear enough that acting on it saves you a step. It appears on your list with the date already resolved.
You are asked first
It shows up as a suggestion you accept or dismiss. A list full of things you never agreed to is worse than no list.
Note what that figure is and is not. It describes how the software behaves, not how often it is right. Those are different claims, and a vendor showing you a confidence threshold has not thereby shown you an accuracy rate. Ask for the second separately, and be suspicious of a precise answer.
Both directions, or it is half a system
The last thing to check is whether a tool tracks only what you owe, or also what you are owed.
Those are the same mechanism read in two directions, and a system that does one is doing the easier half. What you promised is a task list. What others promised you is a follow-up list, and it is the one that quietly costs money, because a supplier who has gone silent looks identical to a thread that simply moved down.
An inbox sorted by recency actively hides this: the longer something has been waiting, the further from view it sits. Sorting by age instead of arrival is the whole fix, and it requires knowing what was promised in the first place.
In CevroFlow, Tasks from Email covers what you agreed to, and Commitments and Follow-up covers both directions with the age of each promise visible.
