You tried to automate your agency. You bought the tool, you sat through the webinars, maybe you hired somebody. Six months later the workflows sat unused and your team was doing it by hand again.
This is the autopsy on that. It is the companion to the three eras, and specifically it is the autopsy on era two, so if you have not read that one this will make more sense after you do.
The first thing to get out of the way: it was almost never the tool. AgencyZoom works. Zapier works. Your AMS works. Blaming the platform is comfortable and it is usually wrong, and worse, it sends you shopping for a replacement that will fail the same way.
What actually killed it
Every automation you have ever built was reasonable on the day you built it. Somebody described a situation, you wrote a rule for that situation, and the rule worked. Then a slightly different situation showed up and you wrote a rule for that. Then an exception to the first rule. Then a rule to catch what the exception missed.
None of those decisions were wrong. That is the part that makes this hard to see coming. There is no single bad rule to point at in the wreckage.
The failure mode is the accumulation. Rules are cheap to build and expensive to own. You pay for them one at a time, in small amounts, at the moment you write them, and you pay for the pile of them every day afterward for as long as it exists.
Two things go wrong once the pile is big enough, and they are both bad.
The first is that nobody will touch it. You know there are rules in there that are not doing anything useful, and you also know you cannot tell which ones without tracing every dependency. So the pile only ever grows. Removing a rule feels risky in a way that adding one does not, which is exactly backwards from the truth.
The second is quieter. Some of those rules stopped firing a year ago, because a field got renamed or a pipeline stage got restructured or somebody changed a form. They did not throw an error. They just stopped. Nobody found out, because the absence of an automation looks identical to the absence of work.
Why era three does not rescue this
Here is the sentence I want you to leave with, because it is the one that saves people money.
There is a real temptation, when you are buried in two hundred rules, to think the answer is something smarter sitting on top. It is not. Pointing a model at a sprawled system does not untangle the sprawl. It gives the sprawl a faster engine, and now the wrong thing happens more often and with more confidence.
| Applied to a clean workflow | Applied to a sprawled one |
|---|---|
| You can tell what it did, because there is one path through the work. | You cannot tell what it did, because there are nine paths and it may have taken any of them. |
| When it gets something wrong, you find the stage and fix the stage. | When it gets something wrong, you find four rules that could have caused it. |
| Speed compounds. More volume through a path you trust. | Speed compounds. More volume through a path nobody has audited in a year. |
| Adding a stage is a small decision. | Adding anything means guessing what else it touches. |
If you are in the sprawled column, the work in front of you is not an AI project. It is deciding which of those rules you actually still want, and turning the rest off. That is unglamorous and nobody sells it, which is probably why it does not get talked about.
The guardrail that makes the difference
There is one rule that separates the era three builds that hold up from the ones that quietly poison a database, and it is small enough to say in a sentence.
Never let it fill a field it does not know.
If it does not have the roof age, it does not put a roof age. It does not estimate one, it does not infer one from the year built, it does not leave a plausible placeholder. It asks, or it stops and stages the record for a human.
This sounds obvious written down. It is not what happens by default. A model asked to complete a record will complete the record, because that is what it was asked to do, and a confidently wrong roof age is far more expensive than an empty one. An empty field gets noticed. A wrong field gets quoted.
Two agencies on the same call, who had never spoken, had both independently built a holding place for exactly this. Work the machine could not finish goes there instead of getting guessed at. One of them calls it staged, the other calls it triage. Neither set out to invent a pattern. They both hit the same wall and built the same door, which is usually a sign that the wall is real.
So what do you do differently
Not much, and none of it is exciting.
Automate by volume, not by complexity. The thing you do four hundred times a month is worth automating even if it is boring. The elaborate referral sequence you run twice a quarter is not, no matter how satisfying it would be to build.
Ship the version that handles most of it. A workflow that covers the common case and goes live in two weeks beats a workflow that covers every case and goes live never. The edge cases will introduce themselves.
Decide who owns it before you build it. Not who builds it, who owns it. If the honest answer is one person who is already busy, you are building the next black box, and you know how that ends because you have one already. That question deserves more than a sentence, so it gets its own piece.
And before any of it, be able to describe how the work happens today. Not how it is supposed to happen. How it actually happens, including the part where Denise checks something on Thursdays for a reason nobody wrote down.