Most automation quotes answer one question: what does it cost to build. That number is easy to compare across vendors, so it gets all the attention. The number that decides whether the automation is still running in two years is different — it is the monthly bill after go-live, and almost nobody puts it in the business case.
We see the same pattern often enough to name it. A workflow gets built for somewhere between $8,000 and $30,000, works well, and then quietly accumulates $300 to $900 a month in licences, tokens, hosting and human babysitting. Nobody budgeted for that, so it lands in an operating line with no owner, and the first time someone reviews software spend the automation looks like an unexplained cost rather than a saving.
This article breaks the run cost into its actual lines, gives current price ranges, and shows which architecture decisions move the number most.
The five lines on the monthly bill
Every production automation we have deployed has costs in these five buckets. The proportions vary wildly, and the one people expect to dominate usually does not.
1. Platform licences
This is the orchestration layer — the thing that runs the workflow. For n8n, cloud plans start at roughly AUD $35 to $40 a month at current list pricing, with the mid tiers landing somewhere in the AUD $100 to $200 range depending on execution volume and number of active workflows. Self-hosting the community edition removes the licence but not the cost: a small VPS in Sydney runs about AUD $20 to $60 a month, plus a managed Postgres instance if you want backups you can rely on, plus your own patching time. Check n8n's pricing page before you commit — the tiering has changed more than once.
Retool prices per user, and this is where budgets get surprised. Someone who logs in twice a month to approve a credit note is still a user. A ten-person internal tool typically lands between AUD $150 and $500 a month at list prices, and the cost scales with headcount rather than with value delivered. See Retool's pricing for the current tiers, and check specifically how your plan counts read-only or occasional users.
The practical implication: platform cost for n8n is roughly flat as volume grows, while Retool cost is roughly linear in the number of humans who touch the tool. Design accordingly.
2. Model and API spend
This is the line clients expect to be frightening, and for most document and text workflows it is the smallest one.
Take invoice intake at 2,000 documents a month. A typical extraction call sends about 3,000 tokens of input and returns 500 tokens. That is 6 million input and 1 million output tokens a month. At the sort of mid-tier pricing available in 2026 — around US$1 per million input tokens and US$5 per million output for a capable non-frontier model — you are looking at roughly US$11 a month. Even tripling the volume and doubling the model tier keeps you under AUD $150.
Agentic workflows behave completely differently. An agent that loops through 20 tool calls, carrying a 30,000-token context each turn, burns around 600,000 tokens per run. Five hundred runs a month is 300 million tokens, and now you are in the hundreds of dollars — for the same business volume. The cost driver is architecture, not throughput.
Two mitigations are worth knowing about, because both arrived or matured recently. Prompt caching means repeated context (a long system prompt, a policy document, a schema) is charged at a heavily reduced rate on the major providers. Batch endpoints trade latency for a substantial discount on work that does not need to be real time — overnight reconciliation, bulk classification, backfills. Current rates are on the OpenAI and Anthropic pricing pages; both have moved downward repeatedly, so any figure in an old business case is probably too high.
3. Infrastructure and storage
Self-hosted anything needs compute, a database, object storage for documents, and monitoring. For a single-workflow deployment this is genuinely small — AUD $50 to $150 a month all in. It grows in one direction, though: if you are storing every processed PDF and every execution log indefinitely, storage and log retention creep up over years. Decide a retention period on day one.
4. Third-party services per transaction
Easy to forget because they are metered. OCR on scanned documents, address validation, SMS notifications, e-signature envelopes, enrichment lookups, bank feed connectors. Individually a few cents. At volume they can quietly become the largest line on the bill. Every metered dependency should appear in your cost model with an assumed monthly volume next to it.
5. Human time
The biggest line, and the one that never appears on an invoice.
Exception handling. If a workflow handles 95% of 2,000 monthly items cleanly, that leaves 100 exceptions. At four minutes each that is nearly seven hours a month of someone's attention. Push the clean rate to 88% and you have doubled it. Exception rate matters more to run cost than any licence decision.
Maintenance. Expect 2 to 6 hours a month per meaningful workflow in the first year, less after that if nothing upstream changes. This is not speculative work: API versions get deprecated, a supplier changes their invoice layout, someone rotates a credential without telling anyone, a field gets renamed in your CRM, a model you depend on gets retired with 6 to 12 months notice and the prompts need retesting. Model deprecation on a roughly 12 to 18 month cadence is now a standard maintenance event, not an exception.
Oversight. Someone has to read the weekly failure summary and care about it. Ten minutes a week is a real cost if nobody is assigned it, because then it does not happen.
A worked example
Invoice intake and coding for a 25-person business, 2,000 invoices a month, self-hosted n8n, a Retool approval screen used by four people, and a mid-tier model for extraction.
| Line | Monthly (AUD, approx) |
|---|---|
| Hosting, database, storage | $110 |
| OCR on scanned documents (~600/mo) | $90 |
| Model spend, including caching | $40 |
| Retool, 4 users | $180 |
| Exception handling, ~7 hrs at $55 loaded | $385 |
| Maintenance, 3 hrs at $150 | $450 |
| Total | ~$1,255 |
About $15,000 a year against a build cost of maybe $22,000. If that offends you, check what it replaced before you react: two people spending a combined 20 hours a week on manual entry and chasing is roughly $57,000 a year loaded. The automation is still comfortably worth it. But a business case that showed $22,000 once and nothing after was wrong by $15,000 a year, and that gap is where automations get killed in the second budget cycle.
Note that software is under a third of the bill. Cutting the model to the cheapest available tier saves $25 a month and might cost you three points of accuracy, which adds an hour of exception handling. The maths runs backwards on that trade almost every time.
Three decisions that actually move the number
Use models only where judgement is required. Regex, lookups, database joins and conditional logic cost effectively nothing and never hallucinate. If a step can be deterministic, make it deterministic and reserve model calls for extraction, classification and drafting. This is the single largest lever on both cost and reliability.
Pick the model per step, not per project. Routing, triage and classification run fine on small cheap models. Only the steps where a mistake is expensive need a frontier tier. Standardising the whole workflow on your best model is the most common source of an inflated token bill.
Cap agent loops. Set a maximum number of turns, and make the workflow escalate to a human when it hits the cap. Without a ceiling, one badly formed input can loop dozens of times, and you find out at the end of the billing period. A turn cap is a cost control and a blast-radius control at the same time.
Budgeting it up front
Our planning heuristic — a rule of thumb from our own delivery work, not an industry statistic — is to budget 15% to 25% of build cost per year for platform, hosting, maintenance and oversight, then add metered spend (models, OCR, SMS) on top based on assumed volume. For most SME workflows that puts total three-year cost at roughly 1.5 to 2 times the build quote.
Three things to insist on before go-live:
- A named owner for the monthly bill, with the licences and API keys in your accounts rather than your vendor's.
- Spend alerts on every metered service, set at about 150% of expected monthly volume, going to a human who will act.
- A measured exception rate after the first full month, so you know whether the human time assumption in the business case held.
Run cost is also the honest test for whether to automate at all. If a process runs 40 times a month, has high variance, and is likely to change when you migrate systems next year, the run and oversight cost will eat most of the saving. When projected run cost plus oversight exceeds roughly two thirds of the current manual cost, the automation is marginal and you should fix the process first.
Pick your strongest candidate process, count last month's volume and how many items needed a human decision, and run those two numbers through the ROI calculator before you brief anyone. If the exception rate is the number you cannot estimate, that is the thing to measure this month.