The question on every pricing call is what an automation costs to build. A better question is what it costs to run in month six, when the launch excitement is gone and someone has to own it. Most budgets cover the AI model and stop there. Here is the full list, with numbers from a real pipeline.
The example is the emailed-quote workflow we built for Lumen Logistics, a 220-person freight brokerage. It reads every quote request that lands in a shared inbox, extracts the route, checks history, applies the rate card, and files a draft for a dispatcher to approve. It handles about 2,600 requests a month. The figures below are from its fourth month in production, rounded.
The five line items
- AI model usage
- Tool and platform seats
- Hosting and data
- Monitoring
- Maintenance
Only the first one shows up in most estimates.
1. AI model usage
One run of the pipeline makes one AI model call: extract the structured request from an email and any attached PDF. The prompt is about 1,500 tokens (the small chunks of text models charge by) of instructions and format, which we cache. The email plus attachment averages 3,200 tokens. Output is around 400 tokens of structured data. Call it 5,000 tokens in and 400 out.
At the mid-tier model rates we were paying that month, a run cost just under three cents. Times 2,600 runs is about $75. Add retries on badly formed output (about 4% of runs), a weekly test pass against 150 held-back emails, and occasional reprocessing when a customer forwards a corrected request. The realistic AI usage bill lands near $105 a month.
That is the number people budget. It is about a tenth of the total.
2. Tool and platform seats
The workflow runs on n8n's cloud plan, a little over $60 a month at the tier with enough executions for the volume. Gmail was already paid for. The transport system (TMS) offers API access at no extra charge on Lumen's plan, which we confirmed before scoping because some vendors charge for it. A small hosted Postgres database holds route history and run logs for about $30. Total: roughly $90.
Two things to check before you sign a scope: whether the tools you already pay for charge extra for API access, and what execution-based pricing does when your volume doubles.
3. Hosting and data
The pieces that do not live inside a subscription tool: a small server for reading PDFs, file storage for attachments (kept 90 days, then deleted), and backups. About $40 a month. This is the smallest line, and the one that grows quietly if nobody sets a time limit for keeping data. Storing every attachment forever tripled one client's storage bill inside a year before anyone noticed.
4. Monitoring
Every AI model call is logged: input, output, response time, cost, and whether the dispatcher approved, edited, or rejected the result. We use Langfuse for this. Self-hosted, it costs a few dollars of compute. On the hosted plan it is about $60 for this volume. Either way, budget it. When a run goes wrong, the log is the difference between a five-minute fix and a two-day investigation.
5. Maintenance
This is the line nobody budgets, and it is the largest. In four months on this pipeline we logged:
- Two email format changes from large shippers, each of which dropped extraction accuracy for that sender until we updated the instructions and the evaluation set. About 3 hours each.
- One TMS update that renamed a field. 2 hours, including the test run.
- One notice that the AI model was being retired, which meant re-running the evaluation set on the replacement and adjusting two instructions. 4 hours.
- A new equipment type the format did not cover. 1.5 hours.
- A monthly review of rejected and edited quotes with the dispatch lead, to catch problems before they become a pattern. 1 hour a month.
That is about 20 hours across four months, or five hours a month, arriving in lumps. At a blended engineering rate of $140 an hour, that is about $700 a month. On our retainer it is covered. In-house, it is a real slice of someone's calendar and should be written down as one.
An automation you stop maintaining does not stop running. It keeps running, slightly wrong, until a customer notices.
The whole picture
| Line item | Month 1 | Steady state (month 4) |
|---|---|---|
| AI model usage | $140 | $105 |
| Tool and platform seats | $90 | $90 |
| Hosting and data | $35 | $40 |
| Monitoring | $60 | $60 |
| Maintenance | 12 hrs ($1,680) | 5 hrs ($700) |
| Total | about $2,000 | about $1,000 |
Month one is heavier because the evaluation set gets run more often and the first format surprises arrive early. For our clients those first 30 days of fixes are included in the build, so the month-one maintenance line shows what it would have cost. It was not invoiced.
Around $1,000 a month in steady state, of which the model is a tenth. Against that: Lumen's dispatchers each got 14 hours a week back. Across the eight dispatchers who handle quoting, that is roughly 480 hours a month. At a loaded cost of $38 an hour, the pipeline returns about $18,000 a month of dispatcher time, before counting the 31% increase in quotes sent per day. The run cost is about 6% of the value. It would still be a good deal at three times the price. It would be a bad deal at any price if nobody owned the maintenance.
How to keep the number down
- Cache the fixed part of the prompt. Instructions and format are identical every run. Caching cut our input cost by about 40%.
- Route by difficulty. A small, cheap model decides whether an email is a quote request at all. The larger model only sees the ones that are. That filtered out 30% of runs.
- Check before you act. A format check after extraction catches badly formed output before it reaches the TMS, so retries are cheap and failures are loud.
- Set time limits on everything. Attachments, logs, run records. Ninety days covers almost every dispute.
- Put a spend alert and a kill switch on the model account. One runaway loop is a month's budget in an afternoon.
- Name an owner. A person, with maintenance hours on their calendar and a 30-minute monthly review of the numbers above.
The budgeting rule
Two rules of thumb have held across our client base. Total run cost lands between four and ten times the AI usage line once maintenance is counted. If your estimate is under four, something is missing. Maintenance for a live workflow with one AI step runs two to six hours a month in steady state. A full agent with several tools and gates is closer to eight to fifteen. Budget the hours before the tokens. The tokens take care of themselves.
If you want the same breakdown for a process you are considering, send us a paragraph about it. We will come back with a fixed build price and a monthly run estimate with all five lines filled in.