The worst automated email anyone receives is the fourth one in a sequence, sent to someone who replied to the first.
It is a small thing that says a large thing. It tells the reader that nobody is on the other end, that the reply they took ten minutes to write went into a system that did not notice, and that the next email will arrive on Thursday regardless.
Follow-up automation fails here more often than it fails on strategy — in the plumbing. So it is worth separating two jobs that get bought as one.
Two jobs, two failure modes
Follow-up automation decides when to contact someone and what to say. It fails on triggers: the sequence fires on the wrong signal, or fails to stop on the right one.
Lead scoring decides who deserves attention first. It fails on inputs: the score is built from data that doesn’t predict anything, and sales learns to ignore it within a month.
They’re sold together, they live in the same tool, and they break independently. Build them in that order — follow-up first, scoring second — because a sequence with no scoring is merely undifferentiated, while scoring with no sequence is a number nobody acts on.
Part one: follow-up automation
Start from exit conditions, not entry conditions
Everyone designs the entry: someone downloads a thing, they enter the sequence. Fine, and easy.
The exits are what make it feel human, and they’re the part that gets skipped because each one is a separate piece of work. At minimum, someone should leave the sequence when they:
- reply to any email in it;
- book a meeting;
- become an opportunity in the CRM;
- enter a different sequence;
- unsubscribe, obviously;
- go quiet after the full run — they exit rather than loop.
The reply exit is the one that breaks most often, and it deserves its own section.
The reply-detection problem
For your automation to stop on a reply, it has to know a reply happened. That’s less automatic than it sounds.
If your sequence sends from a connected mailbox, replies usually land as a detectable event. If it sends from a marketing platform via a shared sending domain, the reply may go to an address nobody’s system is watching, and your automation will never learn about it.
Test this before launch, with a real mailbox: send yourself the sequence, reply from an outside address, and confirm the record actually exits. Do not trust the settings screen. This is a fifteen-minute test that prevents the single most damaging failure in the whole build.
While you’re there, check the out-of-office case. An auto-responder can register as a reply and pull someone out of a sequence they should have stayed in — the opposite failure, quieter, and it looks like the sequence simply underperformed.
Timing: the honest version
There is a lot of published advice about optimal send times. Most of it is derived from someone else’s list, in another industry, at a different price point.
Two things do hold generally. Gaps should widen as the sequence progresses — day 1, day 3, day 7, day 14 rather than four emails every two days — because the person’s interest decays and the pressure of a fixed cadence reads as pursuit. And the first follow-up should be fast, because it arrives while the person still remembers what they downloaded.
Everything else is worth testing on your own list rather than importing.
Deliverability, which is the constraint nobody plans for
Automating follow-up means sending more email from your domain, and volume is the thing mailbox providers watch.
The failure is gradual and hard to see from inside. Sequences go out, the platform reports them as delivered, open rates drift down over a few months, and the team concludes the copy has gone stale. What has actually happened is that a growing share of the sends are landing in spam, where nobody opens them and no bounce is generated.
Three things protect against it:
Send from a subdomain, not your primary domain. Something like mail.yourcompany.com for automated sequences. If reputation degrades, it degrades on the subdomain and your ordinary business email keeps arriving.
Authenticate properly. SPF, DKIM and DMARC records for whatever domain the sequences send from. Google and Yahoo both tightened bulk-sender requirements in 2024, and the direction of travel since has been consistently toward stricter enforcement rather than looser.
Watch complaint rate, not open rate. Opens have been unreliable since Apple’s Mail Privacy Protection started pre-loading images in 2021 — a high open rate may just mean a lot of Apple Mail users. Spam complaints are the number that predicts trouble, and mailbox providers act on it long before you notice anything.
There is also a direct connection back to the exit conditions above. Continuing to email someone who replied, or someone who has ignored six messages, is exactly the behaviour that produces complaints. Good exit logic is a deliverability feature, not just a courtesy — which is a useful thing to say out loud when someone argues that removing people from sequences reduces reach.
Write the sequence to be readable out of context
Email three will be read by people who never opened emails one and two. Writing “as I mentioned” to a person who has seen nothing is a small tell that the sequence was written as a story rather than as a set of independent messages.
Each email should stand alone: what this is, why they’re getting it, one action. The same discipline applies to the sequences themselves — this is the sequencing half of email and inbound systems, and it is where most of the work actually is.
Part two: lead scoring
The case against scoring
Start here, because it saves some teams a month.
Scoring is a ranking system. Ranking only matters when there are more leads than your team can work. If two salespeople receive thirty leads a week, they can call all thirty, and a score adds ceremony without changing behaviour.
There is also a data floor. A model built on fifty historical conversions is fitting noise — you cannot tell a predictive signal from a coincidence at that sample size, and neither can a spreadsheet.
If you have fewer leads than capacity, or fewer than a few hundred historical conversions, skip scoring. Route by one or two obvious rules — company size, inbound versus outbound — and revisit when volume justifies it. That is not a compromise; it is the correct answer at that stage.
Two kinds of input, doing different jobs
| Firmographic | Behavioural | |
|---|---|---|
| Answers | Should we want them? | Are they in market now? |
| Examples | Company size, industry, role, country | Pricing page viewed, demo requested, email clicked |
| Available | Immediately | Only after they act |
| Over time | Stable | Decays — fast |
| Fails by | Ranking a perfect-fit company that isn’t buying | Ranking a tyre-kicker who reads everything |
Firmographic data is uncorrelated with whether the person is buying anything this quarter. Behavioural data is the opposite: it says nothing about fit, and a pricing-page visit means something this week and very little in April.
A score that mixes them without decay produces the classic failure — a lead who read everything in February ranking above a lead actively evaluating you today. If your scoring tool supports time decay on behavioural points, use it. If it doesn’t, cap behavioural points to a rolling window.
Build the model backwards
Do not start by assigning points to actions you think matter. Start with closed-won deals from the last year and ask what was true about them before they closed.
Frequently, the honest answer is that one or two attributes carry nearly all the signal — company size band and whether they came inbound, say — and the remaining eleven scoring rules are decoration that make the model look sophisticated.
Ship the two-variable version. It’s easier to explain to sales, which is the difference between a score people use and a score people override.
The number needs a meaning
A score of 73 means nothing on its own. Bands do: call today, nurture, ignore for now. Define the bands and what happens at each one, or you have built a column rather than a system.
Wiring the two together
The connection is a routing rule: a score crossing a threshold changes what happens to the record.
A workable minimum:
- Lead enters — ideally after validation has filtered the junk, since scoring a fake record wastes the same effort as scoring a real one — and gets a firmographic score immediately.
- Behavioural points accrue as they act.
- Crossing the threshold stops the nurture sequence and creates a sales task.
- Falling back below it after a decay period returns them to nurture rather than leaving them in limbo.
Step three is the one that gets missed. A hot lead who keeps receiving the educational drip while a salesperson calls them is being talked to by two versions of your company that don’t know about each other.
When the score and the salesperson disagree
They will, and how you handle it decides whether the score survives.
A salesperson who works a lead the model scored low, and closes it, has produced information — either the model is missing a variable or that deal was an exception. The failure mode is treating every such case as an exception, which is how a score stays wrong for two years.
Give people a way to override with one field: a reason. “Referral from existing customer”, “attended our event”, “knows our head of product”. Read them quarterly. When the same reason appears repeatedly, it is a variable your model doesn’t have, and adding it is the highest-value change you can make to the score.
An override log is also the only honest way to answer whether the score is working. If sales overrides it constantly, the model is not being used, whatever the dashboard says.
The maintenance cost nobody budgets for
Automation is presented as work you do once. It isn’t, and the ongoing cost is predictable enough to plan for.
Every offer change breaks a sequence. Rename a product, retire a lead magnet, change a price — some email in some sequence now references something that doesn’t exist.
Scoring drifts. The model was fitted to last year’s deals. The market moved; the score didn’t.
Integrations break quietly. A field renamed in the CRM breaks the sync that fed the score. Nothing errors. The number just stops updating, and it looks like a slow quarter. This is the ongoing half of marketing and sales automation, and it is the part that gets left out of every implementation quote.
Budget an hour a month to walk the sequences and a quarterly look at whether the score still correlates with closed-won. Teams that skip this don’t discover the problem — they conclude that automation doesn’t work for them.
Short version
- Build follow-up before scoring.
- Design exit conditions before content.
- Test reply detection with a real mailbox before launch.
- Skip scoring entirely below a few hundred historical conversions.
- Separate firmographic from behavioural, and decay the behavioural.
- Build the score from closed-won backwards, and ship the two-variable version.
- Make the threshold stop the nurture, not run alongside it.
- Put a recurring hour in the calendar, because it does not stay built.
If the distinction between the two halves is still fuzzy, marketing automation vs sales automation draws the line more carefully.