Every automation conversation starts in the same place: the most visible, most creative, most discussed part of the process. It is also the part with the most judgement in it, the least repetition, and the highest cost of being wrong — which makes it close to the worst possible first project.
The best first project is almost always something nobody enjoys talking about.
The variance test
For any step in the process, ask one question: how much does the output vary given the input?
| VARIANCE | EXAMPLE | TOOL | WHY |
|---|---|---|---|
| None | Renaming, resizing, filing, format conversion | A script | Deterministic. An agent here adds cost and a failure mode. |
| A few known paths | Routing by campaign type, applying one of four templates | A branch | Enumerable. Write the conditions down. |
| Path depends on content | Summarising a brief, drafting variants, classifying inbound | An agent | The work is deciding, which is what the model is for. |
| High, and being wrong is expensive | Final approval, claims, anything legal, the creative decision | A person | Not a technology question. |
The common error is reaching for the third row when the first would do. An agent doing a job a script could do is more expensive, slower, and capable of surprising you, which is three disadvantages and no benefit.
Where the hours actually are
When we audit a content team’s week, the time does not sit where anyone expects. Writing is rarely the largest block. The largest block is almost always coordination: finding the current version, renaming exports, chasing an approval, rebuilding the same status deck, reformatting one asset into six sizes, and looking for a file somebody named untitled-final-v3.
That work has zero variance, high volume, and no creative content whatsoever. It is also invisible in every plan because nobody puts it on a timesheet.
The four places automation should not go
- Final approval. Somebody signs, at any volume. The moment approval is automated, nobody is responsible for what ships.
- Anything making a factual or regulated claim. A drafted claim is fine; a published one without a human check is a liability with a schedule.
- Anything touching a real person’s likeness or voice without a checked consent record. The check is the point and it cannot be inferred.
- Anything where the failure is silent. Automation that quietly does nothing is worse than no automation, because the team stops checking.
The sequence
A workable order, in the sense that each step makes the next one possible:
- Map the process first, one row per step, with an owner and a time estimate against each. Automating an unmapped process automates the mess.
- Fix naming and filing. Everything downstream depends on being able to find things reliably.
- Automate export and reformatting. Zero variance, immediate return, and it removes the most-resented task in the building.
- Add routing. Branches that send work to the right template, the right person, the right queue.
- Then, and only then, add agents to the steps where the path genuinely depends on the content.
- Instrument all of it. If you cannot say what a run cost and what it produced, you have a demo.
What makes it survive to month six
Most creative automation is abandoned within a quarter, and the reasons are always the same three. Nobody owns it, so nothing gets fixed when it breaks. It has no logging, so nobody can tell whether it is working. And it was built around one person’s way of working, so it stops when they move.
The scaffolding that prevents that is boring: a named owner, a run log, a written description of what it does, and a fallback path for when it is down. Building it is about a fifth of the work and it is the fifth that decides whether the other four fifths still exist next year.
One row per step, an owner against each, and the verdict: a person decides it, a model drafts it, or a script handles it alone.
THE WORKFLOW CANVAS →