Dossier · future scope
Nine things, ordered by how much each changes what the tool is worth — not by how hard it is to build.
Right now sending is the end of the story. If the reply came back in, the agent would have the whole conversation rather than one opening line — it could draft the follow-up in context, notice when a prospect answered a question it had asked, and stop treating every message as a cold first touch.
It also closes the only feedback loop that actually matters. Today the persona learns from what you edit. With replies, it could learn from what works — which hooks got answered, which openings got ignored — and feed that into hook ranking rather than relying on an intent weighting I chose by hand.
Choose the moment, not just the message. Recipient timezone and working hours are the floor; beyond that, seniority, region and role all shift when a message is likely to be read rather than buried. A founder in Bengaluru and a VP of Sales in New York should not be emailed at the same instant just because a rep clicked at the same instant.
With reply data from the item above, this stops being a heuristic and becomes measurable — you can learn the send window that actually earns answers for a given segment.
Two things hold the persona back today. It needs three consistent edits before it writes a rule down, so the first handful of drafts teach it nothing — and only one persona is active at a time, so a rep selling two products switches between them by hand and the learning from one does not help the other.
Both are fixable without weakening the rule that made it work. Several personas could be live at once and chosen per lead by product or segment, rather than by a switch someone has to remember. And the cold start could be shortened by seeding from messages you have already written, so it starts from your actual voice rather than from a brief describing it.
The threshold itself could be evidence-weighted rather than a flat count: deleting the same closing line three times is a much stronger signal than three loose rewordings, and treating those identically is why it sometimes takes longer to learn than it should. With reply data from the first item, the same machinery could weight rules by whether messages carrying them actually got answered.
Read the account list out, write the activity back. Reading matters more than writing: the CRM already knows who is a customer, who is in an open opportunity, and who a colleague emailed last week. The app enforces those rules today from a column in a spreadsheet, which is the same logic pointed at a much worse source of truth.
OAuth per user, each rep sending from their own address, leads owned by whoever is working them, and duplicate protection across the team. This is what turns a single-operator tool into something a GTM team could actually share, and everything in the access section above collapses into it.
A second and third touch, each researched fresh rather than a templated bump. The interesting version is not “send three emails” — it is that the second message can open on something that happened since the first one, which is exactly the kind of signal the research path already finds.
Today you bring a list, or one company. The natural extension is describing an ICP and having it go and find matching companies and people — the outbound route already does a narrow version of this by walking from one company to its competitors.
Every fact in the research path has to survive a check against the page it came from before it can be used. The competitor list does not get that treatment — it is model output that goes straight into finding contacts.
Closing that gap would make the outbound route as trustworthy as the leads route, and it is the same mechanism already built and working one layer down.
Warm-up, per-domain sending limits, bounce handling, and a real unsubscribe path. Any of these matters the moment volume goes past a handful of messages a day, and none of them exists.
Replies. Everything else on this list is an improvement to a system that currently cannot tell whether it worked. Reply data makes hook ranking, send timing and persona learning measurable instead of assumed — and three of the items above get better for free once it exists.