Jul 10

Email AI

I’m in the middle of a couple books on complexity — and they’ve affected me.    In the email hell that is my worklife, I started to think about an emergent/adaptive inbox.  Here’s what I’m thinking:

Decaying Relevancy:  Imagine all incoming email has a “freshness” date.  It came in X time ago.  That email is either read or unread.  There’s quite a bit of intelligence in these two dimensions.  Email I care about will be read while fresh.  Email I drudge through — or that is complex — will be read and stale.  Email that is unread and stale is useless.  That leaves fresh unread email.  That’s what I really care about making sense of.

Sender Benchmarking:  Now is when it gets interesting.  Imagine that I know, for each sender, a distribution on read/unread and freshness.   That distribution is going to tell a lot — some shapes will be fat-headed, where 80% of the emails I read right away.  Some will be fat-tailed — where I read very few of them right away.  Based on these distributions, I can start to categorize and prioritize my email (or rather, it can be done adaptively based on my normal interaction with email).

Behavioral Conclusion:  The last area that rounds out my adaptive email system is what happens with the read email.  My biggest problem with work email is actually read mail that is sitting in my inbox.  I either need to reply, archive, or take other action.  The first two conclusions are part of my normal email workflow.  Adding a follow-up action usually puts an email in purgatory.  This is where the freshness comes in.

Bringing It All Together

So my problems fall out as follows:

  1. Unread important email
  2. Read, non-concluded, important email
  3. Read, non-concluded, unimportant email
  4. Unread, unimportant email

The key unknown is importance.  Using sender distributions — I can determine importance as a function of freshness.  Adding behavioral conclusion to the freshness metrics, we now can calculate % of items read and % concluded per sender.  We can also weight these by the time it takes to do both of those.

By creating benchmarks/norms, each email can be given an importance rating.  High importance items, with longer times in the unread and/or non-concluded buckets, receive the highest priority.   Just by organizing my email inbox into unread-important and a read-non-concluded views, prioritized by freshness, I know I’d be quite a bit more productive.

Other interesting by-products could be a way to score how productive you are by day/daypart.  You could also create a feedback mechanism to people you interact with via email — how important their emails are to you.  Finally, you could break apart a senders importance distribution to allow them to explicitly rank a message on how important they think it is.  Their rating could then be matched to how your emergent system rates it — and productive feedback loops could ensue.