Email Promise Extractor

Anything that scans for the word "will" finds the easy half. The commitments that actually go missing are the ones that never announce themselves as commitments — this tool labels each row by kind so you can see which sort you're dealing with:

The split between what you owe and what you're owed matters more than it sounds. Both land in the same place — a single Follow up view rather than two lists — because in practice you check one list or none. If you want the reasoning behind collapsing them, our write-up on labels versus folders covers why five to seven action-shaped categories beat a topic tree.

Alongside the commitments, the scan reports the phrases in your thread that sound like commitments and aren't. This list is published in full rather than hidden behind the tool, because the list is the useful part — once you can recognise these, you catch them while you're writing rather than three weeks later:

The fix in every case is the same two-part edit: name the deliverable, name the day. "I'll get back to you" becomes "I'll send the pricing on Thursday." It reads as more of a commitment because it is one.

The model does exactly one job here: find the sentences that carry an obligation and resolve any stated deadline to a calendar day. Everything after that is ordinary arithmetic done on our side, because a model asked to both read a long thread and do date maths will get the maths wrong somewhere, and a commitment list with subtly wrong dates is worse than no list at all.

• Dated commitments are followed up the day after they were due. If that day has already passed, the follow-up is today — a reminder set in the past is not a reminder.
• Undated commitments get today plus your chosen window, three days by default.
• Overdue counts are whole days between the due date and the today you set, so the number stays right whether you scan the thread now or next month.
• Ordering is overdue first (most overdue at the top), then upcoming by date, then the undated ones. Undated items sit last but are flagged loudest — they have no position on a timeline to sort into, which is precisely why they're the highest-risk group on the list.
• Dates the thread never supported are shown as inferred, never as stated. If the model resolves "end of week" to a Friday, the row says so.

One more check runs before anything reaches the page: every quoted sentence is matched back against the text you pasted, ignoring punctuation and whitespace, and any row whose quote can't be found is thrown away and counted. A tool that invents a plausible-sounding commitment — right names, right tone, never actually said — is worse than one that misses a real commitment, because nothing in the output tips you off. If the footnote says two rows were discarded, that check is what discarded them.

• It doesn't connect to your inbox. No Gmail or Outlook access, no OAuth, no permissions. It reads the text you paste and nothing else — which also means it can't see attachments, calendar invites, or the message you forgot to paste.
• It doesn't store your thread. The text goes to the AI model once to be scanned and isn't kept on our servers or used to train anything. Treat it as you would any cloud tool: don't paste passwords or card numbers.
• It can't read a screenshot. Plain text only. Copy the message body, not a picture of it.
• It caps at 14,000 characters and 30 commitments. A year-long thread needs splitting; paste the portion that carries the open items.
• It's a snapshot, not a tracker. Nothing watches these dates for you afterwards. Copy the checklist somewhere that does, or use something that does the watching.
• Low-confidence rows are judgement calls. A row marked Medium or Low confidence means the obligation or its owner was read from context. Check the quote before you act on it.

Split it by who owes what, because the two halves need different actions. For what you owe, the list is a to-do list — the useful move is doing the two-minute ones now and putting a date on the rest. For what you're owed, each row already has a nudge written for you: it quotes the commitment back verbatim, which is the only part of a chasing email that does any real work.

That built-in nudge is deliberately plain and is written locally, not by the AI — it costs nothing and appears instantly. When the nudge needs an argument rather than a quote, the follow-up email generator writes the next message in a thread that's gone quiet, and the gentle reminder generator handles the overdue-invoice end of the spectrum. If the row is something you owe and the reply is the whole job, the AI email response generator drafts it from the message you received.

Timing matters as much as wording. Following up the next day is the most common mistake — across 156,000 sequences, waiting three days instead lifted replies by 31% (Optifai, 2026), and follow-ups account for 58.6% of all replies in Belkins' analysis of 7.5 million emails. The default three-day window in this tool is set to that number rather than to instinct. Our guide to handling email follow-ups has the full cadence.

If the same commitments keep surfacing from the same senders, the problem is upstream of any list. Route those threads out of the general inbox with the Gmail filter generator, and read Inbox Zero at 100+ emails a day for the triage pass that keeps them findable. Client work has its own version of this failure — 66% of clients leave over poor communication, and a dropped promise is what poor communication looks like from the outside.