Vendor selection guide
Rule-based platforms, process suites and AI agents all call themselves automation. They differ in what they can read, who approves, what gets logged and what a run costs. This page gives you the questions, a test method and a cost model, and says plainly where ZeroTwo is not the right pick.
Trigger
What starts a run?
Read
Can it read this document?
Decide
Who or what routes it?
Approve
Does a person see it first?
Record
Is each run logged?
ZeroTwo publishes this page and sells one of the tools in it. We have not run these products against each other, so there is no ranking and no score. What you get is a checklist, a method for a fair test and the few vendor facts we read at the source on 5 October 2026.
The useful split is not AI against no AI. It is how much of the process is fixed rules, how much is interpretation, and how much control you need around the interpretive step.
For example Zapier and Make
Steps are fixed and the data is already fields: a form submission becomes a row, a payment becomes a notification.
The input is a document that has to be interpreted. Both now offer AI steps (sourced below), so the question is how much control and logging surrounds that step.
Larger vendor suites
Many people own parts of one process, there are formal approvals and someone has to prove what happened afterwards.
A small team that needs one working process this week. We have not evaluated any suite for this page, so check the vendor documentation for fit and price.
ZeroTwo is one
The work is reading, summarising or drafting from messy input, volume is moderate and a person reviews the output.
You need guaranteed identical execution at very high volume, or an audit trail you have not confirmed the product provides. Check run history in your own account.
A common claim is that rule-based platforms are limited to one AI provider. Their own documentation says otherwise, which means "has AI" is not a useful way to separate tools. Judge the controls around the step. Read on 5 October 2026 at the pages linked here.
Eight things to find out for each tool before you commit. Where we have not verified an answer, the cell says what to check rather than guessing. Check integration methods and requirements for the integration row.
| Area | Ask of every tool | Zapier and Make | ZeroTwo |
|---|---|---|---|
| Trigger | Does a run start on a schedule, an event or by hand, and which of those do you need? | Check each vendor trigger list. Neither page above documents triggers. | The scheduled tasks page describes recurring runs. Confirm event-based starts in your account before you design around one. |
| Reading the document | Which file types, how large, and what happens with a scanned or photographed page? | Zapier documents PDF text extraction and says it works best on non-scanned PDFs. Make: check its own docs. | Free allows 1 file upload per message; paid plans allow multiple per task. Test your messiest scan. |
| Branching and exceptions | Can an unreadable or low-confidence document go to a person instead of failing? | Check each vendor router, filter and error-path features. | Design the exception path in the instruction and test it with a bad file. |
| Approval before side effects | Is there a pause where a person approves before anything is sent, posted or written? | Check each vendor approval and human-in-the-loop options. | The scheduled tasks page mentions approvals. Confirm what pauses a run in your account. |
| Run history and recovery | Can you see each run, why it failed, and re-run it without duplicating the effect? | Check each vendor run log, replay and retry behaviour. | The scheduled tasks page describes run history. Check what a failed run shows. |
| Integration effort | Are your systems connectable, and how much setup and authentication does each take? | Compare each vendor app directory against your actual tools. | See the connectors page and the integration methods and requirements overview. If a tool is missing, export a file. |
| Cost per accepted run | What is the billing unit, and what does one successful run cost including retries? | Zapier meters AI steps in tasks by tier. For Make and Zapier prices, use their pricing pages. | Credits per run depend on model and task length. Record usage during your test. |
| Data handling | Where do documents go, who can read them and how long are they kept? | Read each vendor security and privacy documentation. | Read the Privacy Policy and Terms before connecting financial or customer data. |
Sources: the two vendor pages above, read 5 October 2026, and ZeroTwo's own pricing, scheduled tasks and connectors pages. For current prices see Zapier pricing and Make pricing.
Method, not a result
We have not run this comparison, so no outcome is shown. It is the method we would use, and the only one that tells you which tool completes your workflow reliably.
For example inbound supplier invoices or customer requests that must reach the right person. Write the routing rule in plain words first.
Around 20 real documents, deliberately including a few messy ones: a scan, a missing field, a duplicate, an odd layout.
Same trigger, same fields, same approval step. Do not tune one tool harder than another.
Record routed correctly, routed wrongly, sent to a person, and failed. Note what a reviewer had to edit.
Everything the tool cost in the period, plus review time, divided by runs a person accepted.
Subscription price alone hides retries, review time and heavier model tiers. Compare cost per accepted run for your own workload, not a vendor average.
(tool cost in the period + review hours x hourly cost) divided by runs a person accepted
Credits per run vary with model and task length. See the full plan comparison.
Our opinion, with the cases where ZeroTwo is the wrong answer stated up front.
Fixed steps between apps, high volume and a need for identical execution each time. If the deciding factor is the longest app directory, compare the vendor directories against your list.
Look elsewhere if the input is free text or documents that need interpretation.
Formal approvals, several departments and a requirement to evidence every step later.
Look elsewhere if you are one small team that wants a working process quickly.
Messy input, moderate volume, a person reviewing output, and a wish to try different models on the same task.
Look elsewhere if you need guaranteed identical execution at high volume or a log you have not verified.
Software that runs a repeated, multi-step business process with less manual handling: moving data between apps, routing documents, requesting approvals and recording what happened. Some tools follow fixed rules, some add AI steps for reading and drafting, and some are AI agents that choose steps themselves.
Yes, according to their own documentation read on 5 October 2026. Zapier's AI step lists OpenAI, Anthropic, Google Gemini, Azure OpenAI and Amazon Bedrock, and Make's AI Agents API lists providers including Anthropic, Gemini, Mistral and others. Make marks those endpoints as open beta, so check the current list before you rely on one.
When the data is already structured, the steps are fixed and you want identical execution every time. Moving a form submission into a spreadsheet does not need a model. Add an AI step only where a document or free text has to be interpreted.
Build the same document-routing flow in each tool with the same approval step, run about 20 real documents through it including some messy ones, and record correct routes, wrong routes, referrals to a person, failures and reviewer edits. Then compute cost per accepted run. This page describes the method and has not run it.
Plus is $14.99 per month and Pro is $29.99 per month, or $26.99 per month billed annually. Model, image, video and deep-research runs draw from a monthly credit pool, so cost per run depends on the model and the length of the task. It has not been measured for a specific process here, so record usage in your own test.
This page does not claim that. It depends on how many systems, approvals and exceptions the process has. A bounded document-routing process with a person reviewing the output is a reasonable thing to try yourself. Anything with strict compliance, many owners or heavy integration deserves a scoped estimate before you decide.
There is no single answer, and this page does not rank models. Test two or three on your own documents, score them the same way and keep the cheapest one whose output reviewers accept. ZeroTwo lets you choose from 60+ models for that comparison.
Not necessarily, but the integration part decides it. You can describe a process in plain English, yet connecting systems, handling authentication and defining exceptions still take care. Check that your tools are connectable before you commit, and start with an exported file if they are not.