AI automation is usually the wrong choice when the rules are already clear, when a mistake is expensive or hard to reverse, or when nobody on the team can define what a "correct" output actually looks like well enough to check the AI's work. In each of those situations, a plain rules-based automation — a webhook, a script, a Zapier or Make.com flow with no model in the loop — is more reliable, cheaper to run, and far easier to debug than anything built on an LLM. This isn't a hedge. Several of the automation systems we've actually built and shipped use no AI at all, on purpose, because the workflow didn't call for it.
Why "just add AI" became the default advice
A lot of automation content right now reads like AI is a feature you sprinkle on top of any workflow to make it better. Some of that is genuine enthusiasm. Some of it is that AI integrations are easier to market, easier to demo, and easier to bill for than "we connected two systems with a webhook" — even when the webhook is the better engineering decision.
The result is a lot of businesses paying for AI-in-the-loop where a deterministic flow would have been faster to build, cheaper to run every month, and far less likely to produce a confidently wrong answer nobody catches.
Four modes, not one default
Every automation decision we make sorts into one of four modes. None of them is the "advanced" one — they're just different fits for different kinds of work.
Deterministic automation
Fixed rules, fixed inputs, no interpretation required. A Stripe subscription event maps to an Airtable field update. A form submission maps to a CRM record. There's exactly one correct output for a given input, so there's nothing for a model to add except cost, latency, and a new way to be wrong.
AI-assisted workflow
The input is messy — free text, a scanned document, a support message written in someone's own words — or the task genuinely requires interpretation, like summarizing or classifying. A model earns its place here because the alternative isn't "write more rules," it's "have a person do it manually," and a wrong AI output is cheap to notice and correct.
Human-in-the-loop
The rules might even be fairly clear, but the outcome matters enough — money moving, an account being closed, a message going out under the company's name — that a person should confirm before the action is final. AI or a rules engine can prepare the decision; a human still makes it.
Fully manual, not yet
The honest fourth option: some workflows aren't ready to automate at all, in any mode, because the underlying process itself isn't settled. Automating an unclear process just makes the confusion move faster. This is the mode that gets skipped in most automation pitches, because "don't automate this yet" isn't what an agency wants to open with.
None of the four modes is the "advanced" one. They're just different fits for different kinds of work — and "not yet" is as legitimate an answer as any of them.
The two questions that actually decide the mode
Underneath all four is the same pair of questions: how clear are the rules, and how expensive is a mistake. Clear rules with a cheap mistake is deterministic automation, no argument needed. Fuzzy rules with a cheap mistake is where AI genuinely helps. Either one paired with an expensive, hard-to-reverse mistake means a human checkpoint belongs in the loop — and if the rules aren't even settled yet, the honest move is to wait.
What this looks like in automation we've actually built
Worth being direct about: most of the automation systems in our own portfolio use no AI at all. The Stripe-to-Airtable subscription sync and the Shopify-to-Notion P&L automation are both webhook-driven, rules-based flows — the mapping from a Stripe event or a Shopify order to the destination record is exact, so a deterministic pipeline is simply the correct tool. The Twilio auto dialer and Twilio conference call system are the same story — call state and sequencing logic that has one right answer per event, not a judgment call.
Our own internal operations stack is a mix: mostly deterministic syncing and reminders across Airtable, Notion, and Google Sheets, with AI reserved for the pieces that are actually language-shaped — drafting, summarizing, first-pass classification. That split isn't an accident. It's the same two questions above, applied to our own operations instead of a client's.
None of this means AI doesn't belong in automation — it means it belongs in specific places. Our guide to where AI automation helps most covers the workflows where it's the right call: messy inputs, support triage, drafting, and reporting that needs judgment, not just aggregation.
A short check before adding AI to a workflow
- Can the correct output be described as a fixed rule, not a judgment call? If yes, you probably don't need a model.
- If the AI gets this wrong, how expensive and how visible is the mistake? Cheap and visible is fine. Expensive or silent needs a human checkpoint, at minimum.
- Can someone on the team actually check the AI's output against a clear definition of "correct"? If nobody can articulate what correct looks like, the process isn't ready to automate in any mode yet.
- Would a plain webhook, script, or no-code flow solve this without a model in the loop? If yes, that's usually the cheaper and more maintainable answer.
None of this is an argument against AI automation. It's an argument against treating it as the default layer instead of one tool among four. The right question was never "should we use AI" — it's which of these four modes the workflow actually calls for.