Learning has been rebuilt from the ground up. You can now upload your own writing samples, see exactly what the AI extracted from them, turn on a writing profile built from that review, and track how each version of that profile performs over time. Learning is on by default for every workspace, and its analytics now count only results that were actually confirmed, not estimates.
What changed
Previously, teaching the AI your voice was mostly a black box. You could feed it samples, but you had little visibility into what it actually pulled out of them, and no easy way to tell whether the profile it built was helping or hurting your reply rates. This rebuild changes that at every step, from the moment you upload a sample to the moment you decide whether a new version of your profile is actually working better than the last one.
You start by uploading writing samples, the kind of emails you would actually send. The system extracts a writing profile from them: tone, sentence rhythm, how you open and close a message, the kind of phrasing you use. Instead of trusting that extraction blindly, you now get to review it. You can see what the AI took from your samples before you decide to activate it, which means you catch a wrong read on your voice before it ever reaches a prospect's inbox, not after.
Once you activate a profile, it becomes part of how your campaigns draft. And because writing style naturally drifts as you add more samples or adjust instructions, the system tracks each version of your profile separately, so you can see whether a newer version is actually performing better than the one before it, rather than assuming an update helped just because it felt like an improvement.
Two other changes matter here. First, Learning is now on by default for every workspace. You do not have to find a setting and turn it on; new and existing workspaces are already benefiting from it, drafting against whatever profile currently exists, even a minimal one, from day one. Second, the analytics tied to Learning have been tightened. Earlier numbers could include drafts or sends that had not actually been confirmed as successful. Now, the performance figures you see only count confirmed results, so the numbers reflect what actually happened, not an optimistic estimate that could overstate how well a profile is doing.
How to use it
Open any campaign and go to the AI Instructions tab. The guided flow has three steps: Describe campaign, Teach the AI, and Practice your style. Writing samples live in step two. In the Learned writing profile section, paste past emails directly, separating each one with a line of three dashes, or upload files in .eml, .txt, or .csv format. You can paste up to 50 emails and upload up to 20 files, with a 1 MiB size cap. Pick emails that represent how you actually write, ideally ones that got real replies, since those are the samples most likely to teach the AI something useful about what works for you specifically.
Once your samples are in, click Analyze writing samples. This is where you check the AI's read on your voice. If something looks off, such as a tone note that does not sound like you, this is the place to catch it before it shapes every draft the campaign sends. Take this step seriously; a profile built on a misread of your voice will quietly push every draft in the wrong direction until you notice and correct it.
The same step also holds a testimonials section, where you approve which customer quotes the AI is allowed to reference, and a field for explicit instructions, where you spell out rules the AI must follow no matter what the samples suggest. As you fill these in, the panel on the right updates the effective brief, which is the combined set of guidance every draft in this campaign will follow, so you can see exactly what the AI is working from at any moment.
The writing profile applies per campaign, which means you can teach one campaign a formal voice for enterprise prospects and another a casual voice for startup founders, without either bleeding into the other.
It is worth revisiting your samples periodically rather than treating this as a one-time setup step. As you run more campaigns and see which emails actually get replies, feed those specific emails back in as new samples. A profile built once from your first batch will drift out of date as your outreach matures, while one you keep updating with your best-performing writing stays a genuinely current reflection of your voice.
Why it matters
Generic AI-written cold email reads generic, and prospects notice. The whole point of Learning is to close that gap between what the AI drafts and how you would actually write it yourself, in a way you can verify rather than take on faith. Being able to review the extraction before you activate it means you are not handing your voice to a system you cannot inspect. Version tracking means you can actually tell whether teaching the AI more about your style is helping, instead of guessing based on a handful of anecdotal replies.
The shift to confirmed-only analytics matters just as much, even though it is less visible day to day. If your reported performance included unconfirmed results, you could end up trusting a number that was never real, and making decisions, like which profile version to keep, based on it. Now what you see in Learning's analytics is what actually happened with your contacts, which makes it a number you can act on with confidence.
Taken together, these changes turn Learning from a setting you configure once and forget into something you can actually manage, the same way you would manage any other part of a campaign you care about improving. You upload, you review, you activate, you watch the numbers, and you adjust again, with every step visible instead of hidden behind a single toggle, and every claim about performance backed by a result that actually happened.
Availability
Learning is enabled by default across every workspace. Sample upload and extraction review happen in each campaign's AI Instructions tab, under Teach the AI, the second step of the guided flow. Profile activation and version tracking work the same way, per campaign, from that same tab, for both new workspaces and existing ones.