The question we get asked is which AI tool is best at reading a support ticket. It is the wrong question, and answering it will not tell you what the thing costs.
Here is why. The model that reads the ticket and says “this is a password reset, route it to tier one” is the cheap part of the job. The expensive part is the automation that sits in front of it: the thing that notices a new ticket, calls the model, reads the answer, updates the ticket and tells someone. That runs on every single ticket. It runs on the duplicates. It runs on the auto-replies. It runs on the one that turned out to be a printer that was switched off.
Your bill is set by that second thing, and by how your automation tool counts.
The meter matters more than the price
Three automation tools in our directory are priced in three different units, and the unit decides what triage costs you.
Zapier bills per successful action. A triage workflow that classifies the ticket, filters on the result, updates the ticket and posts a notification is four actions. Four hundred tickets a month is sixteen hundred actions, whether or not any of them needed a human.
Make bills in credits. Core is $12 a month, and the pricing page states that as the “Price for 10k credits/mo”, which makes it a floor rather than a flat fee. Make moved from operations to credits on 27 August, so older comparisons are counting a unit that no longer exists. Credits also expire at the end of the term rather than rolling over, so a quiet month buys you nothing in a busy one.
n8n meters whole workflow runs rather than steps. That is the structural difference worth understanding: the careful twelve-step version of your triage flow costs the same as the rough four-step one. If you expect the flow to grow more defensive over time, and triage flows always do, this is the meter that punishes you least for it.
We priced the same workflow across all three in what n8n, Zapier and Make actually cost. The short version is that the advertised prices tell you almost nothing.
The free tiers will not do this job
Worth saying plainly, because it is where most people start.
Zapier’s free tier is 100 tasks a month and, more decisively, two-step Zaps only. A triage workflow is not two steps. You cannot classify and then act on the classification in two steps. The free tier is not a small version of triage; it cannot express it at all.
Zapier’s own help centre also notes that “Free and Pro plans are individual plans”. For a team sharing one triage workflow, that pushes you to Team, which is $69 a month billed annually and $103.50 billed monthly.
Make’s free tier is 1,000 credits a month, and Make’s FAQ ties storage to it: “If you are on the Free plan, you have 1000 credits per month, which equals 1 MB of data storage and 1 data store.”
Activepieces gives 100 credits a day on its free cloud tier with every connector included, which is the most usable free tier of the three for trying the shape of a flow before you commit.
There is a second bill, and people miss it
Your automation tool calls a model. That model is billed separately, and a consumer subscription does not cover it.
This catches people out constantly, so here is OpenAI’s own help page on what ChatGPT Plus includes: “Not included: API usage is separate and billed independently.” The same page puts it again: “This is different from API usage billed to an OpenAI API account, which is not included in Plus.”
So the $20 a month you may already pay for ChatGPT buys you a chat window. It does not buy the API call your triage workflow makes on every ticket. That is a second account with its own meter, and it is usage-based, which means it scales with ticket volume in exactly the way your automation bill does.
The same separation applies to Claude, where Pro is $20 a month or $17 billed annually, and to Google Gemini, where the plans run from $4.99 to $199.99.
On Gemini there is a second point for anyone triaging real tickets, which carry real customer data. Google’s generative AI terms tell you to follow “any warnings not to input sensitive or confidential information”. A support queue is full of exactly that. We went through where your text goes on each tool in can you put work data into these AI tools.
Internal IT and an MSP are not the same buyer
This is the split that decides which tools you may use at all, and it has nothing to do with features.
If you are internal IT, automating your own organisation’s tickets, almost everything is open to you.
If you run tickets for clients, read the licence first.
n8n is free to self-host, but its licence is the Sustainable Use License, Version 1.0, and it is not an open-source licence. The grant is narrow: “You may use or modify the software only for your own internal business purposes.” Triaging your own company’s tickets is internal. Running it as the triage engine for your clients is a different activity, and the licence is the thing to read before you build on it.
LM Studio has the same shape. It is free at work now, and the old requirement to request a commercial licence was removed on 8 July 2025. But the grant covers “personal and / or internal business purposes” only.
Node-RED has no such condition. It is Apache-2.0, an OpenJS Foundation project with no paid tier, no feature gating and no licence key. For an MSP, that absence of conditions is worth more than a feature list. It is also the tool in our directory most at home talking to hardware and protocols rather than web apps, which suits a lot of IT estates.
Activepieces is MIT except for two directories, packages/ee/ and the enterprise directory under its server API. The split is clean, which is more than most tools manage.
We compared four of these properly in four self-hosted AI tools called “open source”, and the licence question for client work in running AI tools for clients.
Self-hosting moves the bill, it does not delete it
The appeal of self-hosting a triage stack is obvious: no per-ticket meter. Run n8n or Node-RED on your own server, point it at a model you run yourself with Ollama or LM Studio, and the marginal cost of a ticket is electricity.
That is true, and it is not free. The server, the GPU if you want a model that is any good at classification, the electricity, the backups, the upgrades and your own time are all real. We put numbers on that in what self-hosting an AI tool actually costs you.
Two specifics worth knowing before you plan around the free editions.
n8n’s Community Edition has no published execution cap, so your hardware is the limit, but it leaves out SSO, projects, environments, external secrets, log streaming and Git version control. Several of those are things an IT team specifically wants.
Windmill is the code-first option if your team would rather write a script than drag a box, and its free Community Edition allows unlimited executions. Its paid pricing is a calculator rather than a price list: the headline reads from $120 a month, and the page’s own default settings total about $170. Its licence is also not one licence. The LICENSE file describes the code as “variously licensed”, with parts under Apache-2.0, parts under AGPLv3, and a proprietary licence over some of it.
Open WebUI is the usual way to give a team a chat interface over models you host. It is free with no user cap, but the licence changed: BSD-3-Clause up to v0.6.5, and from v0.6.6 the project’s own licence adds a branding condition that applies above 50 users in any 30-day period.
What we would check before buying
- Count the actions in your flow, not the tools. Write the triage steps down. Multiply by your monthly ticket volume. That number, not the headline price, is what you are buying.
- Check the billing unit and whether it expires. Per action, per credit or per run. Credits that expire at the end of the term change the maths for seasonal workloads.
- Price the model separately. The subscription you already pay for almost certainly does not cover API calls.
- Read the licence if any client work is involved. “Free to self-host” and “free to run for your customers” are different claims.
- Check what the free edition leaves out, not just what it includes. SSO and version control are common omissions and both matter to IT teams.
What we cannot tell you
We have not run a triage workflow across these tools and measured it. Everything above comes from the vendors’ own published pricing pages, help centres and licence files, read directly and dated on each tool’s entry. Most were checked on 13 October 2026.
What that method is good for is catching the gap between what a pricing page implies and what it says. What it cannot tell you is how well any of these models actually classify your tickets, which depends on your queue, your categories and your users. That part you will have to test on your own data, and no comparison article can do it for you.