The follow-up that gets ignored is not the badly written one. It is the one that could have been sent to anyone. Buyers have been pattern-matched on sales emails for ten years. They can spot a templated message in two seconds, and they can spot an AI-written templated message in one. What gets a reply now is not better writing, it is better context, and the difference between a follow-up that earns a reply and one that gets archived comes down to three pieces of information you put in the prompt.
Salesforce State of Sales found that 64% of reps using AI report saving one to five hours per week on routine work. Most of that saving is in the drafting of follow-up emails. The catch is that reps using AI without context inputs are also reporting a slow drop in reply rates over the same period, because their drafts are starting to look identical to every other rep’s drafts. Here is what changes when you fix that.
The follow-up most reps send
A typical AI-drafted follow-up after a discovery call looks like this:
“Hi James, it was great chatting yesterday about your scheduling challenges at Boutique Fitness. I think our software could really help with the issues you mentioned around billing and class management. Would you be open to a quick 15-minute call this week to walk through how it would work for your team? Best, Marcus.”
That email is fine. It is also invisible. The buyer has read fifty versions of it this quarter. The opening compliment, the name-drop of a generic pain point, the fifteen-minute ask, the polite sign-off. Every word of it is replaceable.
The reason is in the inputs. The rep gave the chatbot one piece of context: “boutique fitness studio, scheduling and billing software.” The model wrote the average follow-up that gets sent to that profile every day.
The three context inputs that fix it
Three additions transform the output. Adding all three takes about ninety seconds.
The first is the specific quote. Not a paraphrase of what the buyer said, the actual sentence. “I keep losing track of which clients are on a class pass versus a monthly membership and my front desk staff are getting stressed during the morning rush.” That sentence carries the buyer’s vocabulary, their stated frustration, and a specific operational detail that nothing in your generic value-prop covers.
The second is the disqualifying counter-objection. What did the buyer flag as a reason this might not work? “We are tied into our current platform until November.” Naming that constraint inside the follow-up signals you were listening, and it gives you a reason to write back later that is not “just checking in.”
The third is what you said you would send. If the call ended with “I will send you the case study from the gym in Hamilton,” then the follow-up email is not asking for a meeting. It is delivering the case study and then asking for the meeting. Most reps skip this because the case study is a hassle to find. The model can find it for you in eight seconds if you have linked it once.
The prompt that uses all three
Open ChatGPT or Claude, and run something like this:
“Draft a follow-up email to [BUYER NAME] at [COMPANY] after our discovery call yesterday.
Specific quote they used: “[PASTE THE EXACT SENTENCE THEY SAID, FROM YOUR NOTES OR YOUR GONG TRANSCRIPT]”
Their main constraint: [WHAT THEY FLAGGED AS A REASON THIS MIGHT NOT FIT, IN THEIR WORDS]
What I committed to send: [THE ASSET, CASE STUDY, COMPARISON, OR LINK YOU PROMISED ON THE CALL]
Email structure: open with a single sentence that quotes one phrase from their language back to them. Second sentence delivers the asset I committed to send. Third sentence acknowledges their stated constraint and proposes one specific calendar slot in [WEEK OR DATE] that gives them time to handle that constraint first. Hard limit 90 words. No ‘just checking in.’ No ‘I wanted to follow up.’ No mention of features they did not ask about. Sign off as [YOUR NAME].”
Run that against the same Boutique Fitness example and you can end up with something like:
“You said the morning rush is when the class pass versus monthly mix gets messy. The Hamilton case study I mentioned, the gym that fixed exactly that with our front-desk view, is here: [link]. I know you are tied into your current platform until November. How about a 20-minute call the second week of October so you have time to read it first and we can talk concretely about what a clean migration looks like? James.”
That email reads as if it was written by someone who took notes. Because it was. The model did not invent the morning rush detail or the November constraint. The rep typed both into the prompt. The model just turned them into a tight, sequenced email faster than the rep could.
A counterintuitive note on length
Reps coming from sales-coach training sometimes assume shorter follow-ups always outperform longer ones. The data, when you actually run an A/B, is more subtle. Short follow-ups outperform long ones when the buyer is high-intent and you are pushing for a meeting on a specific date. Long follow-ups (200 to 250 words) outperform short ones when the buyer is mid-funnel and the email is delivering substance, a useful summary, a case study annotation, a comparison they asked for. The model handles both lengths fine. The mistake is using one length for both situations.
The failure mode that gets reps blocked
The trap is feeding the model a buyer name, a company, and asking for “a personalised follow-up.” The model has no real context, so it invents plausible-sounding personalisation. “I noticed your team is growing fast” when you have no evidence of that. “Companies like yours often struggle with X” when X is your sales script’s pain point and not the buyer’s. Every fabricated detail is a tell. Buyers spot tells. Tells get you marked as spam.
The discipline is one specific quote, one stated constraint, one committed-to deliverable, all from your notes, none invented. If you do not have those three inputs from a conversation, you do not yet have a real follow-up to write. You have a cold outreach, which is a different prompt with a different structure.
Where the per-rep tools fit
Gong and Salesloft handle the first input automatically by pulling exact quotes from call transcripts so you do not have to scroll through your notes. Outreach and Apollo wrap follow-up drafting into the same UI as the sequencer, which removes the copy-paste friction. They are convenient. The underlying prompt logic is the same as running it yourself in a chatbot.
The reason these tools have moved fast in the last eighteen months is that pulling the right quote from a fifty-minute call is the work that used to make context-rich follow-ups expensive to write. Now it costs nothing.
What changes about your morning routine
Drafts come together in two minutes instead of fifteen. You stop sending the cold-template version when you do not have time for a custom one, because the custom version is now faster. Reply rates climb because the buyer is reading a message that proves you listened. Booked-meeting rates climb because the calendar slot you are proposing is calibrated to the constraint they named. A pipeline that moves faster without you working longer is the only honest definition of what AI is supposed to do for a salesperson.