"Do we need an AI agent, or is Zapier enough?" — the question is framed so that it has no right answer. These are different tools, and in most companies of 10–200 employees they work side by side. The practical line runs not even between products, but between two units of account: Zapier counts steps, an agent counts conversations. Below: how each unit works, what it costs as of 8 September 2026, and the sign that shows a rule is no longer carrying the load.
How Zapier charges
A Zapier subscription is built from two multipliers. The plan level — Free, Professional, Team or Enterprise — sets the feature set. The separately chosen task tier sets the monthly allowance of tasks, that is, the working capacity itself.
A task is spent on every successful step in a Zap and on every programmatic call. One tool call through Zapier MCP costs two tasks; MCP is the protocol an external program uses to reach Zapier tools without a visual scenario. Built-in tools — Formatter, Paths, Filters, Delay, Looping, Sub-Zaps, Digests, Storage, Tables, Forms — cost nothing. AI by Zapier steps are billed by model tier: the standard one costs one task, the advanced one three. A Lead Router step costs five. When the allowance runs out, scenarios do not stop: the account automatically switches to pay-per-task until the end of the period.
Public self-service pricing, annual billing, verified 8 September 2026:
Tasks per month | Professional, $/mo | Team, $/mo |
|---|---|---|
100 | free | not available |
750 | 19.99 | not available |
2,000 | 49 | 69 |
10,000 | 129 | 169 |
100,000 | 489 | 599 |
The 750 and 1,500 tiers exist only on Professional; Team starts at 2,000 tasks; Enterprise is quoted separately through the sales team. The price list is kept in dollars: at the ECB rate of 06.07.2026 (1 USD = 0.876 EUR) the 10,000-task tier works out at ≈€113 per month, and the rate is not recalculated automatically.
In this model you pay for the number of runs, not for the quality of the decisions inside them. As long as the steps are simple, there is no cheaper model on the market. The moment a model appears inside a step, the cost of that step is multiplied by its tier, and if you have connected your own provider, your own token bill is added to the Zapier bill.
Zapier Agents and Copilot: what is actually in the product
Zapier has long since stopped being "rules only", and a comparison that stays silent about that is misleading.
Zapier Agents is a separate product with its own unit of account. It is called an activity and lives outside the economics of tasks. The free plan gives 400 activities per month; the Pro plan gives 1,500 activities for $33.33 per month on annual billing; the enterprise tier is marked "coming soon". The agent is connected to company data, works across 9,000+ apps, can browse web pages and is available in a browser extension.
Zapier Copilot is an assistant inside the editor: you describe the task in words and it assembles the scenario or the agent. The free plan has a daily message limit; on paid plans it is lifted.
If your task is limited judgment inside a flow you already own, Zapier Agents may cover it completely, and there is no need to pay anyone for a separate agent. We write the same about the neighbouring platforms: Make has its own AI Agents now too, and in n8n the AI Agent node is a genuine agent runtime. What is deterministic here is the classic scenarios of these platforms, not the platforms as a whole.
The sign that shows the line
The sign: a person currently reads something and then decides.
That reading step is exactly what a rule cannot do. Three situations in which the sign fires most often.
- The input is free text. A customer email, a call transcript, a messenger message, a messy PDF from a supplier. A scenario can take a field from a form, but it cannot understand a paragraph.
- The exceptions never end. You add branch after branch to a Zap and each time you find a new case. Branch sprawl means the decision tree has stopped being finite.
- The path itself changes, not the data in it. Today a request is closed with one reply, tomorrow it needs a clarification and a stock check, the day after that — a handover to a person.
The simplest way to check yourself is to count the branches. The scenario "new request from the website → create a deal → notify the manager" lives for years and never needs an agent. The scenario "incoming messenger message → work out whether this is a price request, a complaint or a clarification about an order already placed → find the order → reply" grows a dozen conditions by the end of the quarter, and every new condition adds a branch that someone has to maintain. The first scenario is work for a rule. The second is a reading step disguised as routing.
The mirror sign is useful too. If a step is described by one sentence, "if X then Y", and that sentence has not changed in six months, an agent is redundant here, and any contractor who says otherwise is selling you something you do not need.
Two budget units on one flow
A large Ukrainian online clothing store with a chain of 30+ offline stores (the name is not disclosed by agreement) puts its stream of enquiries from Instagram, Viber, Telegram and the website chat through an AI agent; the agent writes the results into Bitrix24. Since the end of March 2026 it has handled 6,448 conversations across April–July, and 1,782 in June. The median response is 13 seconds. Model spend is ≈€0.10 (≈$0.11) per conversation. The final order is placed by a manager. A breakdown of this deployment is a separate article.
Compare the units. One conversation is not one step: the customer writes, clarifies, changes their mind, asks for a photo in a different colour, comes back two hours later. To count such a flow in tasks you would first have to decide how many steps one conversation takes, and multiply every step that contains a model by its tier. We deliberately do not quote an "equivalent Zapier plan" for this case: that would be an invented figure, not a measurement.
The honest difference lies elsewhere. With Zapier you budget runs and you answer for the result yourself. With us you budget the conversation with the customer and the KPI it leads to, and we answer for the result. These are different purchases, and confusing their price tags is the fastest way to get the estimate wrong.
What neither side's price list shows
Zapier's pricing does not cover the person who runs the automation. Our price list does not cover your readiness to change the process. Both cost items are real, and they are exactly where first budgets usually miss.
In a scenario an error looks like a stopped Zap: you see a red flag in the list and fix the branch. With an agent an error looks different — it answered confidently and wrongly. That is why an agent needs three layers of protection: rules in the prompt, a separate supervisor model that checks the answer before it is sent, and a person at the edge. The working escalation norm, that is, the share of enquiries the agent deliberately hands to a person, is 10–20% according to our own deployments. If a contractor promises zero, they either did not measure it or are hiding the errors from you. How the three layers of protection work is a separate article with examples.
The second invisible item is observability. In a Zap you look at the run history: the step passed or failed. For an agent that is not enough: you need to see what exactly it said to the customer, how many times it went wrong in the same place, and where quality is heading from week to week. In the case of a network of 500 dealers (the name is not disclosed by agreement) the agent holds 95%+ accuracy on typical questions precisely because it hands the remaining cases to a person instead of guessing.
When Zapier is enough
The most frequent correct answer to the original question is "Zapier". The signs that show it:
- The trigger is structured: a form submitted, a deal stage changed, a row added to a table.
- The rule is stable: the logic has not been rewritten in months.
- The volume is moderate, and every run is cheap.
- You need repeatability and an audit trail: identical inputs must give an identical result every time.
- There is a person inside who owns the automation and maintains it.
If this is your case, take Zapier and do not let anyone sell you an agent. An agent starts paying off from roughly 200 requests per month; at a flow of 50–200 it is cheaper to fix the process first than to add another layer on top of it.
What our side costs
Grow2.ai is the AI division of Auspex. We do not sell a platform subscription, but an agent assembled for your process with a fixed-price pilot.
Starter — €1,800 for a 14-day pilot, then from €49 per month. Practice — €3,600 for 21 days, then from €99. Operator — from €6,000 for 30 days, then from €149. Model spend comes on top; in the case above it is ≈€0.10 per conversation. The pilot is paid upfront; if the agreed KPI has not moved at the day-30 review, we refund it in full. Typical payback is 2–6 months; in the dealer network case it came to 2–3.
Setting up the processes and the CRM the agent will live in is handled by Auspex.
Not sure which case is yours? A free 2-minute AI audit will tell you whether you need an agent at all, and whether a rule would not do. General questions about prices and timelines are collected in the AI agents for business hub.