Almost every conversation about AI automation focuses on the build. What it costs to make, how long it takes, which platform to use. Very few focus on the part that actually determines whether the system survives its first year, which is what it costs to keep running every single month after launch.
This is where a lot of small businesses get caught out. A workflow that cost $1,500 to build quietly turns into a $340 monthly line item that nobody forecast, and six months later somebody in finance asks why an "automation" is more expensive than the part-time assistant it replaced. Usually the answer is not that AI is expensive. It is that one specific thing in the workflow is expensive, and nobody has looked closely enough to find it.
So here is the honest breakdown: the four line items in an AI automation bill, a worked example with real 2026 prices, the five things that make bills spike, and seven changes that reliably bring spend down without making the system worse.
Whatever you have built, whatever platform it runs on, your monthly cost sits in four buckets. If you cannot say roughly what each one costs, you do not really know what your automation costs.
This is the thing that runs the workflow. n8n, Make, Zapier, or your own code on a server. Pricing models differ in ways that matter a lot at scale. n8n bills per workflow execution, so a 12-step workflow counts as one run. Zapier historically bills per task, so that same workflow counts as 12. At low volume the difference is trivial. At 5,000 events a month it is the difference between a $50 bill and a $500 one.
As of 2026, n8n Cloud Starter sits around 20 euros per month on annual billing for 2,500 executions, and Pro around 50 euros per month for 10,000. In USD that is roughly $22 and $55. Self-hosting removes the execution cap entirely, which I will come back to.
Every call to an LLM costs input tokens plus output tokens. Current list prices per million tokens run roughly $1 in and $5 out for a small fast model like Claude Haiku 4.5, $2 in and $10 out for a mid-tier model like Sonnet 5, and $5 in and $25 out for a frontier model like Opus 5. OpenAI's GPT-5 family sits in a comparable band depending on tier.
The important thing here is the asymmetry. Output is four to five times more expensive than input, but input is where volume hides. Nobody writes a 40,000 token answer by accident. Plenty of people send 40,000 tokens of context by accident, on every single call.
WhatsApp Business conversations, SMS, and voice minutes are billed by the provider, not by your automation platform, and they scale directly with customer volume. SMS in the US runs roughly $0.008 to $0.02 per segment before carrier fees. Voice agents are the heaviest item on this list by a wide margin, which is why I built a separate voice agent cost calculator rather than trying to squeeze it in here.
Vector database hosting, lead enrichment APIs, document OCR, transcription, email verification. Individually small, collectively the bucket that most people forget entirely. A $0.10 enrichment call on 800 leads a month is $80, which is more than your automation platform costs.
Take a marketing agency in Austin or Manchester getting 500 inbound leads a month. The automation captures the form, enriches the company, classifies intent, drafts a personalised reply for approval, and runs a two-step follow-up if there is no response.
That is about four workflow executions per lead, so 2,000 executions a month. On n8n Pro that is roughly $55.
Each lead triggers three model calls: classification, enrichment summary, and reply drafting. Say an average of 4,000 input tokens and 600 output tokens per call. Across 500 leads that is 6 million input tokens and 0.9 million output tokens. On a mid-tier model at $2 in and $10 out, that is $21 a month.
Add enrichment at $0.08 per lead, so $40, and email sending at $20. Total: about $136 per month, or roughly 125 euros, or 105 pounds.
Notice what that says. At small business scale, the model is rarely the expensive part. Tokens are 15 percent of that bill. This is the single most useful thing to internalise, because it tells you where not to waste your optimisation effort.
Costs in these systems do not creep. They step. Something changes, and the bill doubles the following month. Five causes account for most of it.
Retry loops. A downstream API starts returning errors. Your workflow retries. The retry fails, the error handler re-triggers the workflow, and now you are running the same expensive path 400 times a night. This is the number one cause of shock invoices and it is entirely preventable with a retry cap and a dead letter queue.
Context creep. Someone improves answer quality by passing the full CRM record and the last 30 messages into every prompt. Quality goes up. Input tokens go from 4,000 to 80,000. At 1,500 calls a month on a mid-tier model that is $240 instead of $12. The feature worked; the cost was invisible at review time.
Duplicate triggers. A webhook fires twice, or a polling trigger overlaps with a webhook you added later, and every event runs the workflow two or three times. You pay two or three times and, worse, your customer may get two emails.
Model upgrades applied everywhere. A new frontier model launches, someone swaps it in globally, and now a classification step that a small model handled fine at $1 per million tokens is running at $5 per million.
Volume you asked for. The happiest reason. Marketing worked, leads tripled, the bill tripled. This one is fine, provided you noticed and provided the unit economics still make sense.
In roughly the order I would apply them to a live system.
1. Put a hard cap and an alert on every spend source. Before optimising anything, set a monthly limit and a 60 percent alert threshold on your model provider, your enrichment API, and your messaging provider. Most cost disasters are not expensive systems, they are cheap systems running out of control for three weeks unnoticed.
2. Do the deterministic work deterministically. If a step can be handled by an if-statement, a regex, or a database lookup, do not send it to a model. Checking whether an email domain is a free provider does not need an LLM. This routinely removes a third of model calls.
3. Route by difficulty, not by habit. Use a small cheap model for classification, extraction, and routing, and reserve the expensive model for the one step where output quality is customer visible. In the example above, dropping classification and enrichment summary to a small model cuts token spend by roughly half with no visible change.
4. Trim the context you send. Pass the five fields the prompt actually uses, not the whole record. Summarise long histories once and reuse the summary instead of re-sending raw transcripts on every call. This is usually the largest single saving in a mature workflow.
5. Turn on prompt caching. If your prompts share a long stable prefix, system instructions, brand guidelines, product catalogue, caching that prefix cuts the input cost of repeat calls substantially. It is a configuration change, not a rewrite.
6. Make workflows idempotent. Give every event a unique key and check it before processing. This kills duplicate charges and duplicate customer messages in one move.
7. Batch anything that is not urgent. Weekly reports, list cleaning, and re-enrichment do not need to run per record in real time. Batching cuts execution counts, which is what your platform actually bills you for.
Gideon Wafula builds custom AI automation systems, n8n, WhatsApp, Voice AI, and more.
See Services →Self-hosting n8n on a VPS costs roughly $12 to $40 a month and removes execution limits entirely. Compared to a 50 euro Pro plan that looks like an obvious win, and for high-volume workflows it is.
But the infrastructure bill is not the real cost. The real cost is version upgrades, database backups, restore testing, monitoring, and the Saturday morning where the queue is stuck and it is your problem. Price that at even two hours a month of a developer's time and cloud hosting is cheaper below roughly 10,000 executions a month.
Self-hosting genuinely wins in three situations: sustained high execution volume, data residency requirements that make a hosted US service awkward for EU or UK clients under GDPR, and agencies running many client workflows on shared infrastructure where the per-client marginal cost approaches zero.
If you are a freelancer or agency building these for clients, the mistake I see most often is quoting a build fee and staying silent about running costs. Then the first invoice arrives from OpenAI or Twilio and the client feels ambushed.
Two clean options. Either the client holds the accounts and pays providers directly, so you quote the build plus a support retainer and running cost is transparently theirs. Or you bundle everything into a managed monthly fee with a stated volume band, for example up to 1,000 leads a month, with overage priced explicitly. Both work. What does not work is leaving it undefined.
Either way, put the estimated monthly running cost in the proposal in writing, with the assumptions it depends on. It is a trust signal, and it protects you when volume grows.
Once a month, take five minutes and check these five numbers. Nearly everything in this article gets caught by this list.
The businesses that stay happy with their automations two years in are not the ones that built the cleverest systems. They are the ones who treated running cost as a number they own and watch, the same way they watch ad spend. I write more about the operational side of keeping these systems healthy on my blog, and you can read more about my background as Gideon Wafula, AI Automation Engineer, or see how I structure AI automation services for businesses in the US, Europe, and beyond.