The draft that waits for you

Everyone asks the same question about marketing automation: how much can it do without me? It is the wrong question, and it produces a wrong answer in both directions — people either hand over more than they should or, more often, hand over nothing at all and keep doing everything by hand. The better question is almost the opposite. How much would you let a machine attempt if you knew, with certainty, that nothing it made could reach a single human being without you saying yes first?

The answer, for most people, is a great deal more. And that gap is the whole point. An automation nobody checks has to be timid, because every mistake it makes is a mistake real people receive. It gets one narrow job with clear rules, and anything ambitious stays off the table forever. An automation that cannot ship on its own is a completely different animal: it is allowed to be wrong, frequently, at no cost beyond the seconds it takes you to say no. Restraint at the exit buys you ambition everywhere before it.

The arithmetic is what makes this concrete, and it is worth doing with your own rough figures. Writing a marketing email from an empty page is maybe forty minutes for most people — the staring counts. Reading one that is already written, deciding whether it is true, whether it sounds like you, and whether it should go at all, is closer to three. Say twelve pieces of work in a month: emails, a page rewrite, ad copy, the outline of a video. From scratch that is eight hours you do not have. Reviewed, it is a bit over half an hour of actual decisions, spread across a month, at half an hour a day.

Look at what that unlocks and it is not mainly speed. It is the ability to reject. If a piece of work costs you forty minutes to produce, you will talk yourself into sending a mediocre one, because throwing it away means throwing away the afternoon. If it costs three minutes to read, you can kill eight of twelve and still be far ahead of where you started. Being able to say no cheaply is what raises the standard of what actually goes out — and it is a luxury that only exists when you were not the one who wrote the thing.

a machine that cannot publish is a machine that is allowed to be wrong. that is what makes it useful.

So what does reviewing actually consist of, given three minutes? Not proofreading — the typo is the least of your problems, and a model rarely makes one. Three questions, in order. Is anything in here asserted as fact that nobody measured: a percentage, a customer count, a comparison, a superlative that quietly claims something about the world. Does it sound like your product rather than like marketing in general — your customers' words, the promise you actually make. And the one people skip: should this go out at all, or is it competent filler produced because something was scheduled? If you cannot say in a plain sentence what the reader gets from it, you do not have an email, you have a slot, and skipping it is invisible to everybody but you.

The failure mode of all this is not dramatic and it is extremely common: the gate that always says yes. Approving in a batch without opening anything, at eleven at night, because it has been sitting there since Tuesday. A gate that never rejects is not a gate — it is a delay you have added to your own automation, with a story attached about being in control. The honest check is a count. Over the last month, how many drafts did you send back or bin? If the answer is zero, either the machine is astonishing or you stopped reading somewhere around week two, and you know which one it is.

There is a second failure, in the other direction, and it is the one people arrive at after a scare: the gate that nothing ever gets through. Drafts stack up unread, nothing ships for three weeks, and the automation fails quietly by doing nothing. The fix is not more discipline, it is fewer gates. Reversible, private things — a plan, a suggestion, numbers pulled from your connected tools, a draft rewritten twice — should not need anyone's permission, because their worst case is a wasted minute. The gate belongs on the actions that touch real people and cannot be undone: something sent, something published, and above all something spent. Money earns the strictest treatment there is, which is why a campaign worth building automatically is still a campaign that should arrive paused and spend nothing until a person starts it.

What makes the reviewing survivable is mostly logistics, and it is worth designing on purpose. One place you already pass by, not a separate app you would have to remember. One batch at a time you are actually awake, rather than eleven interruptions across a day. And each waiting item carrying the context you need to judge it — who it would go to, what page it points at, which number it was meant to move — because a draft you have to go investigate is a draft you will approve blind. An approval nobody has time to read is not consent. It is a send with better manners.

This is roughly the shape we build toward in SiteOps, and it is worth being precise about where it stops. The system drafts the email, the post, the ad copy and the plan for the week, from the metrics your connected tools actually report rather than invented ones, and it holds all of it for you instead of shipping on its own confidence. Campaigns arrive paused. Scheduled sends can decline to fire and tell you plainly why — a broken page, an empty audience — rather than going out into a bad moment. And when something is not connected yet, the panel says so instead of drawing a zero you might read as a result. What it does not do is decide, because which promise your product makes was never a calculation.

So stop measuring your automation by how much it can do unsupervised. Measure it by how much more you are willing to let it try because it cannot get out on its own — and then protect the only thing that makes any of it work, which is that your yes still means you read it.