Which agency processes should never be automated

Everyone selling in this category is selling more automation. So a list of what not to automate reads, at first, like a trick — the humble brag before the pitch. It is not. It is the most commercially important thinking we did, because every serious failure we have seen or caused came from automating something on this list, and every one of them cost more trust than the successful automations had earned.

The line is not where most people draw it. It is not “creative stays human, admin gets automated” — plenty of admin should stay human and some creative work automates beautifully. The real line runs somewhere else, and it moved twice while we were finding it.

The test that actually works

A process belongs on the never list when it has one of two properties.

Someone must be accountable for the outcome. Not involved — accountable, in the sense that a client can look at them and ask why. Accountability cannot be delegated to a system, because “the system decided” is not an answer a client will accept, and an agency that starts giving it has begun quietly transferring its risk onto the people paying it.

Being wrong is expensive and slow to surface. A wrong number in a draft report gets caught the same day by the person reviewing it. A wrong judgement in a client relationship surfaces three months later as a churn conversation, and by then nobody can trace it back. The slower the error surfaces, the less automatable the process, whatever the hours saved.

Everything on the list below fails one of those tests, and most fail both.

One: anything a client hears as your judgement

Recommendations. Escalations. The answer to “what should we do about this”. The moment where somebody has to say the honest thing about a campaign that is not working.

A brain can assemble every input to those conversations — and should, because the assembly is most of the labour. Traffic fell eighteen per cent, it is isolated to two templates, it started the week of the deployment. That is retrieval, and machines are better at it than people. What follows — whether to escalate to the client’s CTO, absorb the fix, or use the moment to have the bigger conversation about their technical debt — is a decision with commercial and relational consequences, made by someone who will answer for it. We wrote about that boundary in what an agency brain can do, and it is the one limit that does not soften as the models improve.

Two: the relationship itself

Check-ins that exist to maintain trust rather than transfer information. The renewal conversation. The apology.

This is the first line that moved on us. We started with meeting follow-ups on the automatable side — they are mechanical, after all: who committed to what, by when. The extraction is. But we found a category of message that looks like a follow-up and is actually relationship work: the note after a tense call that smooths the thing that nearly went wrong. Drafted by a system, those read like they were drafted by a system, and the client can tell, and the telling is worse than no note at all. Follow-ups moved to “drafted, always rewritten” and the purely mechanical extraction stayed automated. The line moved inward.

Three: anything sent under a person’s name without their eyes on it

This one is absolute in our shop, and it is the reason every publishing action in our own setup requires a named approval. Not because the drafts are bad — mostly they are good, which is precisely the danger. A good draft invites the reviewer to stop reading. The rule “nothing leaves the building without a person” survives only if it has no exceptions, because every exception becomes the new default within a month.

The corollary surprised us: it is also the answer to disclosure. When a client asks what is automated, “everything you receive was approved by a named person, and here is what they checked” is a stronger answer than any percentage.

Four: the decision to stop

Systems do not notice their own decay. An automation that produced good work in March and mediocre work in September will keep producing mediocre work indefinitely, at scale, confidently. Somebody with standing has to own the review that catches drift, and somebody has to be able to switch a thing off over the objection of everyone who has stopped checking it. That role cannot be automated for the same reason an auditor cannot audit themselves.

This is the second place our line moved, in the other direction. We assumed quality review had to stay entirely manual. It does not: the checkable layer — the contradiction between the summary figure and the table, the client’s name from the wrong template — automates well, and automating it makes the human review better, because the reviewer’s attention lands on whether the work is any good. The judgement stayed human. The floor under it did not need to.

What this looks like from the client’s chair

Clients do not experience your automation decisions as architecture. They experience them as whether the report arrived unchased and whether the person on the call knows their business. The observable half of that story is written up from the client’s side in where we draw that line, and why it moved — including the three questions a client can ask an agency to find out whether it has actually held these boundaries or just claims to.

The short version: an agency that automates the never list becomes efficiently forgettable. The one that automates everything else and spends the reclaimed hours on the never list becomes very hard to replace. The technology is the same in both cases. The judgement about where to stop is the whole difference.

How to hold your own line

Write the list down before the project starts, not during it. Ours is four items; yours may differ, but it should exist on paper, because a line that lives in someone’s head moves under deadline pressure and nobody notices it moving.

Expect it to move twice: once inward, when something you automated turns out to have a human core, and once outward, when something you protected turns out to have a mechanical shell. Both moves are the process working. What is not the process working is the line moving silently — which is why the review in item four is the item that protects all the others.