Ask three vendors what automation costs and you'll get three different numbers. Not because anyone's lying, but because each one is pricing a different tier. Run a single low volume workflow, say a hundred executions a month, through Zapier and two years of it costs roughly six hundred dollars. Build the same job as custom software and the same two years costs roughly ten thousand dollars, before the thing has automated anything at all. Both figures are true. They just answer different questions, and the useful move is working out which one you're actually asking.
There are three tiers, and judgement decides which one fits. Not company size. Not how impressive the tool sounds in a demo. Event driven rules that need no interpretation, a lead landing in a form and needing to reach a CRM, belong on a connector like Zapier or Make: cheap to start, cost climbing with task volume rather than build time. Identical repetitive steps on a stable screen, a fixed PDF that needs turning into a spreadsheet every Monday morning, belong to robotic process automation instead. Setup runs medium to high, and it gets expensive fast the moment the underlying interface changes underneath it. Then there's the tier for anything needing variable judgement with a human still watching whatever leaves the building: triaging an inbox, drafting a reply, flagging the one message that actually needs a person to read it. That's an AI agent's job, and the cost runs from tens of dollars a month in inference for a lean build to low five figures for a proper pilot. Six figures once it spans several departments.
The clearest real number we've found is a documented migration. An agency running fifty workflows moved off a three thousand dollar a month Zapier bill onto a custom stack. The engineering cost of the move worked out to roughly ten months of the old bill. It paid for itself in about eight months, once the new stack's running cost settled near eighty dollars a month, plus a few hundred more for the long tail of workflows nobody bothered migrating. That crossover only shows up past roughly five hundred workflows, or one workflow clearing something like a million runs a month. Below that line, the custom build is the expensive option, not the clever one. A typical development quote of ten to fifty thousand dollars only earns its keep once volume actually clears it.
Plenty of processes aren't worth automating yet. The honest signals are duller than anything in a pitch deck. Can you describe the whole flow in one sentence with no "it depends" hiding inside it? Does nobody need to read an email thread before deciding what happens next? Is the volume too small to repay the build within a year? Any one of those, leave it alone. The dullest tool that solves the problem beats the most impressive one nearly every time. A plain connector or a short script usually outlasts an expensive agent pilot built for work that never needed judgement in the first place.
Working out which of those three tiers a given process actually needs, and whether it clears the volume to be worth building at all, is the part that's easy to get wrong from the outside. That's the call worth bringing someone in for, before any tool gets bought or built. The wrong tier costs more than doing nothing.