The CEO of Zapier — the company whose entire business is no-code automation, and which has spent two years bolting AI onto it — is now going around saying: no code is code.
Wade Foster's point, backed by his own AutomationBench testing data: frontier models still stumble on realistic multi-step tasks, and the winning agentic pattern isn't "AI everything." It's a pipeline where a handful of steps use model reasoning and everything else runs on dumb, deterministic code. Filters, lookups, formatters, plain IF statements.
That's not a vendor being humble. That's the person with the best seat in the house telling you where the failure modes actually live. And it maps perfectly onto how a small business should be building automations right now.
Why AI Inside a Predictable Step Is a Liability
A deterministic step — a filter, a lookup table, a formatter — has a beautiful property: the same input produces the same output, every time, forever. It fails loudly and obviously. You can debug it in one look.
Put a chat model in that same step and you inherit three new problems:
- Variance. The model that formatted the invoice correctly on Tuesday may add a flourish on Thursday. Same input, different output.
- Cost. You're renting tokens to answer a question a lookup table already knew for free.
- Silent failure. A broken IF statement throws an error you can read. A model that's 90% right fails politely, in prose, and you find out from a customer.
If you can write the IF statement, don't rent a neuron.
The Three-Tier Decision Rule
Route every step of any workflow to exactly one tier. No step gets to skip the routing.
Tier 1 — Deterministic. No model at all.
The correct output is fully determined by the input: invoice generation from line items, reminder emails on a schedule, routing a form submission to the right owner, formatting, receipt sending, calendar moves, status notifications. Use your automation tool's native steps, filters, formatters, or a ten-line script. Cost: pennies. Failure modes: nearly none.
Tier 2 — Cheap decision models.
The step has to read unstructured input, but the correct answer is short: a label, a category, a yes/no. Inbox triage (urgent or not), ticket bucketing (billing / scheduling / complaint), review sentiment, lead qualified or tire-kicking. A small model returning structured output does this at a fraction of chatbot cost. The step needs to read, but not to think out loud.
Tier 3 — Expensive model judgment, behind a human gate.
The step genuinely requires taste, nuance, or synthesis: customer-facing prose, ambiguous exceptions, analysis that informs a decision, anything with negotiation in it. This is where frontier tokens earn their price — and where a human still approves before anything goes out the door.
The Two Questions That Do the Routing
For any step in any workflow, ask:
"If the same input arrived twice, should the output be identical?"
If yes — Tier 1. Full stop. The step is deterministic and doesn't need a model, however smart the demo looked."Does the step need to read unstructured input or pick between options?"
If it reads and answers short — Tier 2. If it needs to produce prose, exercise taste, or weigh tradeoffs — Tier 3.
That's the entire framework. Two questions.
Worked Example: The New-Lead Workflow
A typical small-business lead intake, routed correctly:
| Step | What happens | Tier |
|---|---|---|
| 1 | Form submission lands in the CRM | 1 — plain zap step |
| 2 | Qualified, needs-info, or not-a-fit? | 2 — decision model, one label out |
| 3 | Route to owner by service type + area | 1 — lookup table, not a model |
| 4 | Draft the first reply | 3 — frontier model, human approves before send |
| 5 | Create the callback slot | 1 — calendar API |
The common mistake is wiring one chat model across all five steps. That's five places to be wrong differently every Tuesday, at 50x the token cost, for exactly two steps that needed thinking.
Note what Tier 1 is doing in step 3. Routing leads by service type and zip code is a business rule, not a judgment call. The moment you find yourself prompting a model to reproduce a decision you could write as a table — you've built variance into a step that never needed any.
The Pricing Shift That Makes This a P&L Question
Seat-based software pricing is quietly giving way to usage-based pricing as agents do more of the work per seat. The practical consequence for a small business: your cost per automated step is becoming a line item you actually control.
That turns the tier decision into a pricing decision, not just an engineering one. Every step you downgrade from Tier 3 to Tier 1 stops billing you forever — a subscription you cancel by making one lookup table. Most 10-person businesses have at least a few steps paying chat prices to answer questions a filter already knew.
What Not to Do
- Don't add AI to a step because the demo looked smart. Demos are optimized for the moment of watching, not the thousandth unattended run.
- Don't put a model where a filter would do. Deterministic steps fail loudly; models fail politely. Loud is what you want.
- Don't rebuild a working automation wholesale into an "agent." Downgrade or upgrade one step at a time and watch it run for a week.
The Quick Win
Open your automation tool tonight. Find the workflow with the most AI steps in it and run the two questions over each step. Most stacks contain at least one step paying frontier prices to answer a question a lookup table already knew. Downgrade that one step. That's a permanent cost cut and a reliability upgrade in the same move — and it's the pattern the no-code industry itself is quietly converging on.
Want ready-made automations built on this exact pattern — deterministic where it counts, AI only where judgment lives? The Boring Automation Pack includes workflows with this routing logic you can adapt to your own stack.
Top comments (0)