AI email triage means letting a model read your unread mail and rank it before you do, so what needs you is already at the top. The method is old. Reading all of it is new. The real question is whether that judgment is any good, and what happens the first time it’s wrong.
- A model is reliable on the judgment you’d make in a second. It’s shakier on the one that needs context from outside your inbox.
- Gmail’s Priority Inbox reached roughly 80% accuracy, and most of that came from learning one person at a time rather than from a cleverer model.
- People drop an imperfect algorithm faster than they drop an equally imperfect person. Being able to correct it is what makes it usable.
- Test it against your own inbox for two weeks. The check is cheap, and the answer is specific to you.
What is AI email triage?
Triage is the pass that sorts arriving mail by what it’ll cost you to ignore it, and how email triage works covers the method in full. AI email triage is that same pass, run by a model, over everything that arrived, before you sit down.
Search the term and you’ll mostly find developer tutorials: wire this webhook to that classifier, add a queue, ship it. That’s useful if you’re building one. If you run a firm and want to know whether to let a model near your client mail, you want a different answer. This guide is that answer, and who Point is for takes it trade by trade.
What can a model actually judge?
There’s a good public answer here, because Google published theirs. The Learning Behind Gmail Priority Inbox is a 2010 research note by Douglas Aberdeen, Ondrej Pacovsky and Andrew Slater. It describes a system that “ranks mail by the probability that the user will perform an action on that mail”. Accuracy ran “approximately 80 ± 5% on a control group”.
Two things there are worth sitting with.
The first is the definition. They predicted whether you’d act on a message, and used that as the stand-in for importance. Mostly it’s a good proxy, since the mail you act on is usually the mail that matters. Where behavior and interest come apart, the ranking follows the behavior.
The second is where the accuracy came from. One model for everybody left a 45% error rate. A model per user brought that to 38%. Making each user’s threshold personal too brought it to 31%. As the paper puts it, “importance ranking is harder as users disagree on what is important, requiring a high degree of personalization”. A good ranking is learned from your own firm, one person at a time.
The payoff was measured on Google’s own staff, and it’s real rather than dramatic. Across roughly 2,000 employees running it, they “spent 6% less time reading mail overall, and 13% less time reading unimportant mail”, and came away “more confident to bulk archive or delete email”. The confidence is the bigger win. Time saved is pleasant. Being able to stop checking is what changes a day.
What can it not know?
A model reads the mail, and the mail is all it has. What you agreed on the phone yesterday stays on the phone. The client who’s been quietly unhappy since March reads like any other client. An invoice for this amount is routine for one client and alarming for another, and the message looks the same either way. The partner who usually handles this is away, and that fact lives in your head. You carry those commitments somewhere other than your inbox.
Here’s the reassuring part. A model is good at the judgment you’d have made in a second from the sender, the subject and the first two lines. It makes that judgment on all of your mail, rather than on the top of the pile. It gets shaky exactly where you’d have stopped to think. Those are the messages you were always going to open anyway.
There’s a thinner spot inside the inbox too. Judgment built on how you’ve dealt with a sender starts from scratch the first time that sender writes. A first approach from someone new is the message you can least afford to miss. It’s also the one a model knows least about.
Why the first mistake matters more than the accuracy rate
Algorithm Aversion is a 2015 paper in Journal of Experimental Psychology: General by Berkeley Dietvorst, Joseph Simmons and Cade Massey. They found that people “more quickly lose confidence in algorithmic than human forecasters after seeing them make the same mistake”. That’s five studies, in the Wharton School’s lab and on Mechanical Turk, forecasting rather than sorting mail. Participants dropped the algorithm even after watching it beat the human. One visible error was enough.
This is worth knowing about yourself before you start. Your assistant will misfile something in the first week, and your instinct will be to call the whole thing unreliable. That instinct is normal. A person doing the same job would misfile more.
The same researchers found the way out. Their 2018 follow-up in Management Science put an algorithm’s forecasts against the participant’s own. People were “considerably more likely to choose to use an imperfect algorithm when they could modify its forecasts”. That preference “held even when participants were severely restricted in the modifications they could make”. A little control was enough.
That gives you the test to apply to anything you’re considering. Ask what happens when it’s wrong. Ask whether you can fix it in one move.
How much do you hand over, and how do you check it?
You hand it over a piece at a time, and each kind of work gets its own amount. Point puts a separate setting on each kind of task. The settings run from suggest-only, through review, to handling the work outright. Review is what you get out of the box. At the top of a setting, Point does that work on its own, which is why the settings move one at a time.
Sorting and filing is usually the first thing people hand over. Anything that goes to a client tends to stay on review far longer, and the reason is worth stating plainly. Filing is reversible: one click on the record puts the thread back in your feed. A sent reply is not, because it’s already sitting in somebody else’s mailbox. Undo covers what Point did to your inbox, and stops there.
Point writes down everything done for you, in order and in plain language. Corrections go in the same way, rather than through a settings screen. Tell Point once that a summary buried the deadlines, and that becomes a standing rule. Star a contact and everything they send scores higher from then on. That’s your judgment entered directly, rather than a guess the model has to make. You can audit the whole thing whenever you like: one tap switches the feed from ranked to plain time order, everything in the sequence it arrived. A ranking you can check is one you trust on evidence rather than faith, and an inbox deserves that much.
The trust question starts before the ranking, of course, particularly in a practice whose mail is full of other people’s financial affairs. Point holds suspicious mail at the door, unread and unprocessed, in a quarantine only you can release. Individual sensitive messages can be locked end to end, so only you and the recipient can read them. The security questions worth settling first go through the whole checklist, and the benefits page covers what you get in return.
How do you test it before you trust it?
Two weeks is enough, and the whole test runs with your hands still on everything. Choosing between products is a different exercise, covered in choosing email triage software. This one asks a narrower thing: you have the tool, and you want to know whether to believe it.
- Keep the ranking on and the acting off. Let it sort, and leave everything else on review. Judgment is the thing you’re testing.
- Read the top of the list, then read the rest anyway. For the first week, do both. You’re hunting for a message it put low that you’d have put high.
- Count the misses. A wrong ordering you shrugged off is noise. One important message ranked low is the number that matters, and it should be zero after week one.
- Correct it rather than tolerate it. Star the people who always matter, and say when a summary is off. A system you correct is the one you end up trusting.
- Read the log once a week. If the record is thin or hard to follow, that’s your answer about the tool.
If the important mail is reliably at the top and the log reads the way you expect, you have your evidence. Otherwise you’ve spent two weeks reading your inbox twice.
Common questions
Is AI email triage accurate enough to rely on?
The best public figure is around 80%, from Google’s Priority Inbox, and personalization is what moved it. That’s easily accurate enough to stop ranking your mail yourself. Keep reading your inbox as well. The measure that matters is whether anything important was ranked low, and you can check that in two weeks.
Will it learn how my firm works?
That’s where the accuracy comes from. Google’s result was that a model per user beat a general one by a wide margin, and personal thresholds beat that again. Expect roughly right in week one and noticeably better by week three. Expect to have corrected it a few times in between.
Can I still see my inbox normally?
At any moment. Point keeps a plain chronological view one tap away. Connecting Gmail puts the same three groupings on your actual mailbox as labels, so the judgment travels with your mail, into the app you already open.
What if I disagree with how it ranked something?
Say so, and it should hold. A correction should stick the first time, and one you repeat is a bug. Being able to modify an imperfect system is what makes people willing to use it at all.
The short version
A model can rank your mail about as well as you would at a glance, across all of it, before you arrive. The parts of your job that happen away from email stay with you, and it’ll get something wrong in the first week. What decides whether it’s worth having is whether you can see what it did, disagree in one move, and have that stick. Run two weeks with the sorting on and the acting off, and count the important messages it put low. That number is your answer.