Three Ways to Automate a Process, and When Each Wins

There are three ways to take a manual process off someone’s desk, and they are not a ladder. Connector tools wire existing apps together and charge by volume. Custom software encodes your exact process once and then runs for the cost of hosting. AI handles the steps where the input is messy and judgement is required. Most processes need the first or the second. The expensive mistake is reaching for the third because it is the one in the news.

The choice is not about sophistication. It is about three questions: how often the process runs, how many exceptions it throws, and how bad a silent error would be. Answer those honestly and the approach usually picks itself. Get them wrong and you either pay a subscription that grows faster than the business or commission a build that a $20 subscription would have covered.

Checked on 19 September 2026. Vendor pricing quoted below comes from the vendors’ own pages on that date and changes regularly.

How do the three approaches compare?

They differ on four things that matter to a business owner: what each is actually good at, how the cost behaves as usage grows, how it fails when something goes wrong, and who can change it afterwards. That last column is the one people forget at the start and resent later, because it decides whether you can adjust the thing yourself or need to book someone.

	Connector tools	Custom software	AI in the loop
Best at	Joining apps that already have integrations	Your exact process, at volume	Messy input, genuine judgement
Cost shape	Per task or execution, monthly, grows with volume	Upfront build, then hosting	Per call, plus the checking time
Typical failure	Runs out of quota, or breaks on an exception	Needs a developer to change	Produces a confident wrong answer
Set-up	Hours to days	Weeks	Weeks, plus ongoing supervision
Who changes it later	Anyone, in a visual editor	A developer	A developer, and someone watching output

The rest of this post is the detail behind that table, and the test at the end is the part worth keeping.

When is a connector tool the right answer?

When the apps involved already have integrations, the volume is modest, and the process barely varies. This is the fastest route from problem to working automation, often the same afternoon, and for a lot of small businesses it is the whole answer. Starting here is almost always correct, because it costs little to find out you were right.

The thing to understand before you commit is how these tools bill, because it decides whether they stay cheap. Zapier charges by task, and its own pricing page is explicit that “Every step in a Zap and every external connector call uses tasks” (Zapier), with “Pay-per-task billing” and paid plans “Starting from $19.99” per month. A five-step workflow therefore costs five times what a one-step workflow costs, every time it runs. n8n prices differently: “Pricing based on monthly workflow executions, regardless of complexity” (n8n), so you “Pay for full executions, not for each step”.

That difference is invisible at ten runs a day and decisive at a thousand. It is the single most useful thing to check before picking a tool, and it is why we default to execution-based pricing on the workflow automation we build. n8n also publishes a self-hosted route: “A standard, self-hosted version of n8n is available on GitHub”, which moves the cost from a subscription to a server and keeps the data on your own infrastructure.

Where connectors stop: exceptions. A connector follows the path you drew. When a real invoice arrives with two purchase orders on it, or a customer replies to a confirmation email instead of clicking the link, the workflow either stops or does something wrong quietly. The rule of thumb we use is that once you are maintaining more than a handful of branches to cover special cases, you have outgrown the tool and are now programming in a visual editor, badly.

When does custom software beat a subscription?

When the process runs constantly, when it is specific enough that no integration exists, or when the per-task bill has quietly overtaken the cost of building it once. Custom software encodes your actual process, including the exceptions, and then costs almost nothing to run. The trade is upfront money and a developer whenever it changes.

The clearest case we have built is an ecommerce operation where orders arrived, had to be matched to the right wholesaler, transmitted, tracked and invoiced. Ninety per cent of it now runs without anyone touching it, and it removed a six-figure annual operating cost. No connector tool covers that, because the matching logic is specific to that business and nothing off the shelf knows it.

The honest test for this tier is arithmetic, not ambition. Take the hours the process consumes each week, multiply by what the person doing it costs, and compare against a fixed build quote spread over a year. If the payback is not inside twelve months, stay on the subscription. We publish what builds cost so you can run that sum before talking to anyone, and we have written separately on where a custom build beats off-the-shelf software.

Worth doing with your own numbers, because the answer flips faster than people expect. Take a process running two hundred times a day through a six-step workflow. On task-based pricing that is twelve hundred tasks a day and roughly thirty-six thousand a month, which sits in the paid tiers rather than the entry plan. On execution-based pricing the same work is six thousand executions, because the step count stops mattering. And on a custom build it is a server bill that does not move whether the process runs two hundred times a day or two thousand. Those are illustrative figures rather than a quote, but the shape holds: the subscription scales with your success and the build does not.

There is a second cost people miss, which is who can change it. A connector workflow can be adjusted by whoever built it, in an afternoon. Custom software needs a developer, which is fine when the process is stable and expensive when it is not. If your process is still changing every month, that argues for staying on a connector tool longer than the arithmetic alone suggests.

Where custom software stops: anything that requires reading meaning out of unstructured text or making a judgement call. A rule cannot decide whether a complaint is serious. That is the boundary where the third approach starts earning its place.

When does AI actually belong in the workflow?

On the specific steps where the input is unstructured and a human would have to read it and decide. Classifying an incoming enquiry, pulling fields off a supplier invoice that arrives in a different layout each time, drafting a reply that needs the details filled in. Narrow steps, inside a workflow that is otherwise deterministic, with the output checked.

What makes this tier different is not capability, it is the failure mode. A connector that breaks stops. A rule that is wrong produces something visibly wrong. An AI step that goes wrong produces fluent, plausible output that looks like every other day’s output. The team that builds agents at Anthropic is blunt about the trade:

“The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments, along with the appropriate guardrails.”

Erik S. and Barry Zhang, Anthropic (Anthropic)

The measured numbers support the caution. On CRMArena-Pro, a benchmark of realistic sales and service tasks, “leading LLM agents achieve only around 58% single-turn success on CRMArena-Pro, with performance dropping significantly to approximately 35% in multi-turn settings” (arXiv). A step that is right roughly a third of the time across a multi-step job is not a step you leave unattended.

Which does not make it useless, it makes it a component. Used on one classification step with a person clearing an exception queue, AI removes real hours. Used as the whole workflow, it produces work rather than removing it. We go through what an agent can actually do day to day, and the plain test for agent versus rule separately.

Where AI stops: anywhere the output is not checkable. If nobody can tell quickly whether the answer was right, do not automate that step with a model.

Which should you actually start with?

Start with a connector tool unless you already know it will not work. It is the cheapest way to find out whether the process is even worth automating, and a surprising number of processes turn out not to be. Run it for a month and watch two numbers: how much it costs and how often it needs a human to fix something.

Then the decision falls out of those numbers. Rising bill but almost no interventions means volume has outgrown the pricing model, so move to custom software. Manageable bill but constant interventions means the process has more exceptions than the tool can express, so also custom software, this time for correctness rather than cost. Both fine means stay where you are, and spend the money on the next process instead.

Only after that is settled does AI become a sensible question, and the question is never “should we use AI” but “which single step in this workflow needs judgement”. Reporting on MIT’s research into enterprise AI, Fortune notes that more than half of generative AI budgets go to sales and marketing tools, yet “MIT found the biggest ROI in back-office automation” (Fortune). The boring middle of the business pays better than the visible front of it, and that has been true of all three tiers for as long as we have been building them.

One more thing worth saying, because the three-way framing can mislead. Almost every system we build in practice is a mix rather than one tier. A typical order-handling setup has a connector catching the email, custom code doing the matching and the writing into the accounting system, and a single AI step reading the awkward supplier attachments that never arrive in the same format twice. Choosing a tier is really choosing it per step, and the skill is drawing the lines in the right places: deterministic wherever the output must be identical every time, AI only where a person would otherwise have to read something and decide, and a connector wherever an integration already exists and works.

If you are not sure which tier your process sits in, that is a short conversation rather than a project. Tell us what the process is, or start with the first workflows worth automating if you are still choosing where to begin.

Frequently asked questions

Is Zapier or a custom build cheaper?

It depends entirely on volume, because they have different cost shapes. Connector tools charge monthly by task or execution, so the bill grows with usage. A custom build costs money once and then runs for the price of hosting. At low volume the subscription wins easily. At high volume the build wins. Work out the hours the process costs you each week and compare against a fixed quote over twelve months.

Do I need AI to automate my business?

Usually not, and assuming you do is the most common expensive mistake. Most processes worth automating follow rules: move this data there, send that reminder, match these records. Rules are cheaper, faster and predictable. AI earns its place on the narrow steps where input is unstructured and a person would otherwise have to read it and decide, and those steps still need their output checked.

What happens when an automation breaks?

It depends which kind. A connector tool usually stops and emails you, which is annoying but safe. Custom software fails how it was built to fail, which is why error handling and alerting belong in the build rather than being added later. An AI step is the dangerous one, because it can fail silently by producing confident wrong output. That is why AI steps belong inside checked workflows, never alone.

Can I move from a connector tool to custom software later?

Yes, and that is usually the right order. Starting on a connector tool proves the process is worth automating and teaches you where the exceptions are, which is exactly the knowledge a good build needs. The work does not transfer directly, but the understanding does, and a month of real usage is worth more than any amount of upfront specification.