Two kinds of cost, and why pitches blur them
Every automation project has a one-time cost, the build, and a recurring cost, the run. Inflated ROI decks blur the two: they compare a heroic savings estimate against the build alone and let the recurring lines surface later, on your card statement. We would rather you see the whole bill first.
The recurring side has four lines that matter: platform licensing, hosting if you self-host, messaging, and AI model usage. None of them has a fixed price we could honestly print, because vendors change their numbers. What does not change is what drives each line, so that is what we itemize here.
Platform licensing
GoHighLevel is a flat monthly subscription with a small number of tiers, which makes it the most predictable line in the stack. What moves the bill is less the tier than the metered services inside it: SMS, email, and AI features are billed on usage on top of the subscription.
Zapier and Make are usage-based. Zapier meters tasks, Make meters operations, and both grow with every new workflow and every busy month. In the setups we audit, this is the most common source of quiet cost creep.
n8n splits the choice. n8n Cloud is a subscription by plan. Self-hosted n8n is free to run under its sustainable use license, and you pay for the server it runs on instead, so volume stops being a pricing question and becomes an infrastructure question. This is why most of our high-volume designs end up there.
| Line item | Billing model | What moves the number |
|---|---|---|
| GoHighLevel | Flat monthly subscription by tier | Tier, plus metered SMS, email, and AI usage |
| Zapier | Usage-based, per task | How many workflow steps run, and how often |
| Make | Usage-based, per operation | Scenario complexity multiplied by run frequency |
| n8n Cloud | Subscription by plan | Execution volume and plan features |
| n8n self-hosted | Free under its license; you pay for hosting | Server size, and who maintains it |
| SMS (Twilio or GoHighLevel telephony) | Per message segment, plus carrier fees | Volume, message length, destination country |
| AI models | Per token, on input and output | Run volume, context size, model tier |
Hosting, if you self-host n8n
A modest virtual server runs most small business workloads; n8n does not need exotic hardware. The technical drivers are memory, execution volume, and whether you need queue mode for scale. As infrastructure lines go, it is a small one.
The honest cost of self-hosting is not the server, it is the responsibility: updates, backups, monitoring, and someone who notices when a disk fills up. If nobody on your side wants to own a server, n8n Cloud is the honest recommendation, not a failure of ambition.
Messaging, and the A2P 10DLC toll booth
SMS is billed per message segment with carrier fees on top, so the drivers are volume, message length, and destination country. Follow-up sequences multiply volume faster than people expect: one lead can generate a dozen messages across a polite persistence window.
The bigger surprise is not the price, it is the paperwork. US business texting requires A2P 10DLC registration: your brand and your messaging campaign are registered with the carriers before your number can send at volume, and unregistered traffic gets filtered. Approval is measured in days to weeks, not hours.
That is real friction with real lead time, which is why registration sits at the top of our project plans. A system that is built but not yet allowed to text is not finished.
AI model usage
Models are billed per token, roughly per word, on both what you send and what comes back. Three things drive the line: how many runs happen, how much context each run carries, and which model tier does the work.
All three are design choices. Retrieval keeps context small by sending the relevant passages instead of whole documents. Routing sends cheap classification work to small models and reserves stronger models for drafting and reasoning. A well-designed system spends tokens where they earn something.
For most operational workloads, classification, extraction, and qualification, the model line stays small next to the labor it replaces. Long conversational transcripts and heavy document processing are the cases that deserve a real estimate up front, which is exactly the kind of estimate the audit produces.
The one-time build
The build is where quotes vary most, and scope is why. The honest drivers: how many systems are being connected, how many approval checkpoints you want, how much failure handling the process deserves, how messy the existing data is, and how much documentation and handover you expect. With us that last item is not optional.
Beware two quote patterns. A price that seems low usually excludes error handling, documentation, and the unglamorous edge cases, the parts that make a system dependable. A price that arrives before anyone has mapped your stack is not a quote, it is a guess with a signature line.
We build inside your own accounts and you own the workflows outright, so the build is a purchase, not the start of a dependency.
When automation is not worth it yet
Some operations should not buy any of the above yet. We say this in audits regularly, because a system sold into the wrong situation fails in public and poisons the well for the right system later.
None of these disqualifiers is permanent. Volume grows, processes settle, and the honest answer changes. If automation is not the right answer yet, we say so, and you keep the assessment.
Honest disqualifiers
- Too little volume. A handful of inquiries a week does not repay a subscription stack; a disciplined shared inbox beats it.
- The process is not stable yet. If how you qualify, quote, or fulfill still changes month to month, automation bakes the churn in. Stabilize it manually first.
- One person can still do it in an hour a week. Below that threshold, the meetings to specify the system cost more than the work it removes.
- The bottleneck is judgment, not repetition. If every case is an exception, there is no pattern to automate, and pretending otherwise produces a robot that guesses.
The verdict
A number for your system, not a generic quote
Everything above is drivers rather than totals, deliberately. The moment we print a price for an unspecified business, we are doing the thing this guide exists to argue against.
The free audit turns drivers into a number for your operation: which platforms fit, on which licensing model, what your realistic message and model volumes look like, what the build scope contains, and where the 10DLC lead time lands in the plan. If the numbers do not justify the system, the audit says that instead. Either way, you know the honest cost before anyone asks for a commitment.
Straight answers
The questions we hear about this choice.
What does an AI automation system cost per month to run?
It depends on four recurring lines: platform subscriptions, hosting if you self-host, messaging volume, and AI model usage. We do not print totals because vendor prices change, but for a typical service business the recurring stack is modest next to the labor it replaces, provided the volume is there. The audit estimates each line for your specific system.
Is self-hosting n8n cheaper than n8n Cloud?
Usually, once volume is meaningful, because self-hosted n8n has no per-execution pricing and a modest server carries a lot of workload. The tradeoff is responsibility: updates, backups, and monitoring become yours. If nobody on your team wants to own a server, cloud is the better answer even where it costs more.
Why can we not just start texting leads this week?
US business texting requires A2P 10DLC registration of your brand and messaging campaign with the carriers, and approval takes days to weeks. Unregistered traffic gets filtered. We start registration at the top of the project plan so the system is not sitting finished before it is allowed to speak.
Will AI model costs grow out of control as we scale?
Token costs scale with volume, but cost per run is a design choice: retrieval keeps context small, and routing sends cheap work to cheap models. In practice, growing run volume means growing lead or order volume, and the model line stays a small share of what that volume is worth.