What email needs to learn from AI agents
Open, select, delete. open, select, delete.
I set up a Hermes agent a couple of days ago, and yesterday I gave it access to its own email account and Google drive. The practical problem this solved was that on Telegram, Saki*, my AI agent was unable to send me files. Saki is running on a Raspberry Pi 5 (8GB RAM, with a 1TB SSD for fast input-output), and connected via Telegram.
I had to jump through several hoops to set up Saki’s gmail address: sign up for Google Cloud, create a project. Enable API’s one by one. Create a desktop app. Run a complicated authentication procedure. Google doesn’t make anything easy does it?
Still, I feel the need for giving my AI agent an email address so it can mail on my behalf. On the other hand, the other day I got pitched something by someone’s “Karan Claw” and it just felt odd… almost rude. So, I definitely don’t want an agent emailing as me.
I don’t want to give it access to my email, for obvious security reasons. I don’t want to risk hallucinations.
I’ve also been thinking about giving an AI agent access to my email just to manage my email. Like many others, I find managing email overwhelming. It’s not surprising that daily morning email summaries are a popular AI use-case. Email is largely becoming legacy infrastructure because everyone hates email.
Reasoned vibe coding workshops
Last Saturday, I held a (physical) workshop on vibe coding in Delhi for 14 people, each of whom ended up building a website, an interactive website, and a mobile app. One build a pacman game. Another made a guitar tuning Mac app. A meal planner, a physiotherapy exercise tracker.
They came to the workshop with zero coding knowledge, and walked out with the ability to build more. If you’re in Delhi, the next workshop is on June 5th.
Max 15 people.
Sign up here
(Or if you know someone who might benefit from this, send them the link. Thanks.)
The problem with Email
I hate email. My email address is public information, and I get about 400 emails a day, mostly spam from PR agencies. My email setup is quite unique, after years of figuring out what works for me: I have accumulated about 2000 filters and blocklists for auto deleting emails, especially from PR agencies.
It’s still a mechanism for me to communicate with MediaNama readers, respond to client queries, email clients updates, forward emails to people. People in the team cc me on things, and my (human assistant) manages some things. All these decisions are mine, and it’s situational and a very “learnt” experience. My email inbox is segmented into six parts: Active (starred emails go here), Unread, TED, Asia21 (I’m a fellow at both), and large files (for deletion to save space). I hate email but I can’t do without it. Schrodinger’s email, it is.
The consequence of this is that email has become transactional, and overly optimised for opening. Most communication is short because attention spans are short. Like YouTubers who optimise thumbnails for growth, emailers optimise subject lines for opening - Growth Hacking, they call it. When people are tricked into opening emails because of the subject line, they hate it.
The problem is that these systems keep getting gamed, and we keep changing our personal logic in terms of whether we should check an email or just delete without reading.
How AI can improve email
I think email needs to improve with AI, and here’s how:
First, learn the rules: email currently works on universal signals and a “confidence score”. Platform providers pass signals to each other about whether something is a new account, account activity, a large number of emails sent soon after setup, domain related authentication (DKIM). There are workarounds, of course, like emails get “warmed up” before they scale spamming. An evolved approach to emails would go beyond just aggregate logic, and incorporate a users conditional logic.
Second, learn the behaviour: Every user deals with email differently: what they open, what they spend time on (depending on the length of an email), and understand their priorities the way agents can over time by building user-specific memory. It is possible to seek a little more information from users and ask them to mark an email as useful or spam. You shouldn’t need agents to triage your email for you, especially since it’s not that complicated to build personalised triage within email itself. Once the logic is learned, triage becomes easier.
Third, learn the prioritisation and presentation: Replace agents summarising emails for you by integrating summarisation within email itself, and then create a dashboard of segmented emails, in terms of user chosen sections, alongwith suggested actions that can be performed with one click (like deleting promotional emails).
Fourth, act on it: Ask users whether they would like a separate segment for that type of email (for example, separating out transactional emails). If they get a lot of emails from a particular address, or a particular type of address (for example, for me, those with medianama email addresses), separate those out automatically. Gmail gives helpful tips by resurfacing sent emails that haven’t been responded to. Can that be replaced, within Gmail, with a poke like function, instead of having to mail again?
The handover
Email providers are ceding the user experience to AI because AI handles email better than email itself does. That’s the structural shift sitting underneath every “summarise my inbox” demo.
What’s slowing the handover is trust, in two ways:
First, data: giving an LLM access to your personal email means trusting the contents won’t end up as training data, or surface somewhere you didn’t intend. For some users, that risk alone keeps agents out entirely.
Second, and harder to engineer around: a summary is only useful if you trust what it leaves out. I still check my Spam and Trash folders to make sure nothing important got filtered, and Gmail has had years of my feedback to work with. If I can’t trust filters I’ve trained over years, why would I trust an agent’s summary after a week?
The strange part is that almost none of this needs LLMs or agents. Learning rules, learning behaviour, surfacing priorities, acting on them — email could do all of it. It just hasn’t.
And one of the most-used services on the internet refusing to evolve while agents around it improve is a recipe for redundancy.
*named after the pen name of my favourite author Hector Hugh Munro. It is believed to refer to the farsi word for a person who serves alcohol.