The math on manual pre-call research has never worked well for high-volume outbound. Fifteen to twenty minutes per account to review the company's recent news, job postings, and LinkedIn activity. If you are working a territory of 300 accounts and have four hours of prep time per week, that gives you realistic research capacity for maybe 12 accounts. For the other 288, you either skip the research or reuse a template that is not grounded in anything specific to that account.
AI drafts do not eliminate research. They compress the time it takes to extract and apply the most relevant signal. The difference is significant: what used to take 15 minutes per account takes under a minute when the signal is already identified and the draft builds from it. The rep's job shifts from finding and synthesizing information to reviewing and adjusting a draft that already has the right context embedded.
What Makes an AI Draft Actually Useful
The gap between a useful AI draft and a useless one is almost entirely determined by the quality of the signal input. AI language models are good at constructing grammatically coherent, plausible-sounding sentences in the tone and format of outbound emails. They are not good at generating the specific, current, account-level context that makes a first touch feel researched.
A draft that starts with "I noticed your company is in the B2B software space and is focused on growth" is not a personalized email. It is a template with a very wide match condition. The rep who sends it is not saving research time. They are just automating a template that was already not working.
A draft that starts with "You posted a VP Revenue Operations role three weeks ago and followed it with two SDR openings. That sequence usually means someone is rebuilding the outbound motion from the top down" is doing something different. It is grounded in a specific, current, verifiable signal. The reader knows you have looked at their hiring activity in the last month. That changes the email from a blast to a note worth responding to.
The Signal-to-Draft Pipeline
Effective AI drafting in outbound works as a pipeline, not a one-step generation. The stages are: signal identification, signal ranking, context extraction, draft generation, and rep review.
Signal identification is the account monitoring step. Which of the accounts in your territory have shown a relevant signal in the last 7 to 14 days? This step is automated in Leadbay: the system monitors job boards, news sources, and web activity continuously and surfaces the accounts with new, relevant signals each morning.
Signal ranking answers the question: for a given account, which signal is most relevant to the outreach context and most likely to resonate with the specific contact you are targeting? An engineering hiring surge is most relevant if you are reaching out to a technical buyer. A VP Sales posting is most relevant if you are reaching out to the CRO or founder. The draft should open with the signal most likely to register with the specific reader.
Context extraction is where the draft gets its specificity. The system reads the signal and constructs an inference: what does this signal imply about the account's current situation and what problem are they probably trying to solve? That inference becomes the opening line of the draft.
The rep review step is not optional. AI drafts built on good signals are usually 70 to 80 percent of the way to a sendable email. The rep adjusts the tone to match their voice, corrects any inference that does not quite fit what they know about the account, and adds any context that the signal monitoring would not have captured (like a personal connection or a recent direct conversation).
What Changes for the Rep's Workflow
With manual research, the rep's morning starts with opening a browser and spending 15 minutes on each target account. With signal-based AI drafts, the rep opens their queue and sees a ranked list of accounts with associated signals and draft first touches already prepared. The first decision the rep makes is not "what do I know about this account" but "does this draft capture the right angle for this account."
This shift has some second-order effects that are not immediately obvious. Reps spend more time on tone adjustment and less time on information gathering. The quality of tone calibration tends to improve when reps are not also fatigued from 20 minutes of research. The decision about whether to send also gets better: when the signal and draft are both in front of the rep at the same time, it is easier to assess whether the timing is actually right rather than defaulting to "I did the research so I should send something."
The risk is the opposite failure mode: reps who treat AI drafts as final outputs and send without review. Drafts built on weak signals produce openings that sound specific but contain inferences the reader will recognize as off. An inference that does not quite land is sometimes worse than no inference at all, because it signals that the sender used automation without judgment rather than not researching at all.
The Tone Matching Problem
One of the more interesting challenges in AI-drafted outbound is matching the rep's voice. Two reps working the same accounts with the same signals will produce very different emails if left to their own styles. AI drafts that sound generic or corporate will be edited heavily by reps who have developed their own voice. Drafts that are too casual will be edited by reps who write in a more formal register.
The approach we take in Leadbay is to train the draft model on each rep's sent email history. After 8 to 10 sent emails, the model has enough patterns to produce drafts that are much closer to the rep's natural style in terms of sentence length, formality, and vocabulary. This reduces the editing work substantially and increases the likelihood that the draft goes out close to how the rep wrote it.
This does not make tone matching a solved problem. Reps whose sent history is mostly internal emails, marketing templates they forwarded, or unusually long messages will generate training data that does not reflect how they actually write cold outreach. The model needs a clean sample of actual rep-written first touches to calibrate well.
When AI Drafts Are the Wrong Tool
AI-generated first touches work well for accounts where the signal is strong and recent, the product fit is clear, and the outreach is genuinely cold. They work less well for accounts that are already in a conversation, have had a previous rejection, or require a highly specific and technical opening that the signal data alone cannot inform.
Reactivation outreach, where you are contacting an account that went cold six months ago, needs a different approach. The reason for reaching out again needs to be credible and specific to why timing has changed, not just the strongest current signal. For those cases, the rep's judgment about what changed in the relationship context is more important than the automated signal ranking.
Using AI drafts for 100 percent of first touches is a mistake. The right usage is for the accounts where a signal-grounded opening will genuinely add value and the rep does not have additional context that would produce a better opening manually. That is probably 60 to 70 percent of first touches for a well-structured territory, not all of them.
