“How much will it cost?” is usually the first question about an AI project, and the quote for building it is usually the only answer people get. That’s a third of the picture. An AI tool also costs money every month it runs, and it needs looking after as your process, your data and the models change.
This post breaks the cost into those three parts: build, run and maintain. It shows our own price ranges for the build, works through an example of the running costs and lists the levers that keep the monthly bill under control.
Three kinds of cost
Where the money goes
| Build | Run | Maintain | |
|---|---|---|---|
| When you pay | Once, before launch | Every month | Every month or per change |
| What drives it | How complex the process is, how many systems to connect | Volume, model choice, size of each request | How often the process, data or models change |
| Typical items | Workshop, integrations, screens, testing | Model usage, hosting, search index, voice minutes | Monitoring, prompt and model updates, support |
| Easy to miss | Cleaning up data and exceptions | Staff time checking the output | Retesting when a model is retired |
| Easy to predict | Yes | Partly | Partly |
Build: what you pay once
A build covers more than “connecting ChatGPT”. For a typical AI process it includes:
- Discovery. Mapping the process with the people who run it, collecting real examples and agreeing what a good result looks like.
- Integrations. Reading from and writing to your systems: email, ERP, CRM, document storage. This is often the largest part of the work, especially with older systems.
- The AI part. Instructions for the model, search over your documents if it needs them, and a set of test cases to check quality against.
- The screens people use. Where someone reviews a draft, fixes a field or approves an exception. A good review screen saves more time than a clever prompt.
- Testing and launch. Running real cases, handling the edge cases and showing the team how to use it.
Here are our own price ranges:
What a build costs at QikBuild
- €5–10k One process: a simple tool or AI feature connected to 1–2 systems
- €10–30k Connected system: a portal or internal tool with 10+ features, connected to existing systems
- €30k+ Larger projects: platforms, replacing old software, many integrations
Which tier would a project fall into? Three illustrative examples:
- An accounting firm wants supplier invoices read from email and entered into its accounting software, with a person checking the exceptions. One process, two systems: the first tier.
- A clinic wants a patient portal where people book appointments, fill in an intake form and get answers about preparing for a visit from the clinic’s own instructions. Several features connected to the booking system: the middle tier.
- A distributor wants to replace a 15-year-old order system, add AI order intake from email and connect the warehouse, accounting and a customer portal. A larger project.
What pushes the price up: more systems to connect, old systems with no API, many exceptions, high accuracy requirements (which mean more testing) and detailed permissions. What brings it down: one clear process, good examples and a person on your side who owns the outcome.
We agree a fixed price after a workshop in the first week, then build in two-week stages with a working demo every week. Our past projects took six weeks to three months to reach a first version.
Run: what you pay every month
Running costs come from four places.
Model usage
Language models are billed by the token, a piece of a word. In English a token is roughly three-quarters of a word on average, and Slovak usually needs more tokens for the same text. You pay for the tokens you send (instructions, the document, the conversation so far) and for the tokens the model writes back. Output tokens usually cost several times more than input tokens. We explain this in more detail in tokens and context windows.
Prices can differ ten times or more between a provider’s smallest and largest models, and they change often, so we don’t quote them here. Check the current price list of the provider you use. If you run your own model instead, usage turns into a fixed server cost; see self-hosted LLMs or API models.
Hosting
The application itself, its database, file storage and logs. If the AI searches your documents, there is also a search index, often a vector database (a database that finds text by meaning rather than by exact words). For a small internal tool this is usually a modest monthly line, but it grows with the number of users and documents.
Voice and other services
Voice agents add speech-to-text, text-to-speech and telephony, usually billed per minute. Text recognition, e-signature and text messages are billed per page, per document or per message.
People’s time
This one never appears on an invoice. If a person checks every AI draft, their time is a running cost, and in most processes it is the largest one. Good design makes that check fast, and over time limits it to the exceptions. Our post on human in the loop shows how.
A worked example
Let’s put numbers on a typical case. Everything here is an assumption, so swap in your own.
A distributor receives 2,000 orders a month by email. An AI workflow reads each email and attachment, matches the items to the catalogue and prepares a draft order in the ERP. A person checks each draft before it is confirmed.
Assumptions:
- Each order sends about 4,000 tokens to the model: 2,500 for the instructions and the relevant part of the catalogue, 1,500 for the email and attachment.
- The model writes about 500 tokens back: the structured order.
- The model costs €3 per million input tokens and €15 per million output tokens. This is a made-up mid-range price for the example, not any provider’s current price.
- Hosting, database and logs cost €150 a month.
- A person spends one minute checking each draft, and their time costs the company €30 an hour.
The maths:
- Input: 2,000 × 4,000 = 8 million tokens × €3 = €24
- Output: 2,000 × 500 = 1 million tokens × €15 = €15
- Model usage: about €39 a month
- Review time: 2,000 minutes, about 33 hours × €30 = about €1,000 a month
Monthly running cost in the example
Two things stand out. First, the model bill is small next to the cost of people checking the output. Cutting the check from one minute to 30 seconds saves about €500 a month. Switching to a model at half the price saves about €20. Second, the comparison that matters is with today. If entering an order by hand takes six minutes, the team currently spends about 200 hours a month on it. How to work that out properly is in measuring the ROI of an AI project.
The same job done by an AI agent costs more. An agent decides its own steps and calls tools in a loop: look up the customer, search the catalogue, check stock, check the price list. Each step sends the growing conversation back to the model. With eight steps, input can easily reach 40,000 tokens per order and output 1,500. Under the same assumptions that is about €285 a month instead of €39, roughly seven times more for the same result. Agents earn their cost when the steps can’t be fixed in advance. For order entry they can. More on that choice in when to use AI agents.
How to estimate your own usage
- Count the requests per month: emails, documents, questions, calls.
- Estimate the input per request: instructions, plus the document or the relevant passages, plus any conversation history. A page of plain text is somewhere around 500 to 800 tokens in English, and more in Slovak.
- Estimate the output: a short label is a few tokens, a drafted reply a few hundred.
- Multiply by the provider’s current input and output prices.
- Add a generous margin, say a third to a half on top, for retries, testing and growth.
Then replace the guesses with facts. During a pilot, log the tokens of every request. After a week you’ll have real numbers.
Maintain: what you pay over time
An AI tool isn’t finished at launch. Expect work in four areas:
- Model changes. Providers retire older model versions on a schedule, and new ones behave slightly differently. Each switch means rerunning your test cases and sometimes adjusting the instructions.
- Process changes. New products, a new invoice layout from a big supplier, a new approval rule. The tool has to keep up.
- Monitoring. Error rates, cost per request and what users flag. Without it, you find problems when a customer does.
- Support. Someone who fixes things when they break and makes small changes.
With us, the client owns the code, repositories and accounts from launch. We fix bugs free for three months after launch, and monthly support after that is optional. Whoever does it, put maintenance in the budget from day one. A tool nobody looks after slowly gets worse as the world around it changes.
How to keep the running costs down
Cost levers
- Use a rule where a rule works A rule costs nothing per request and never guesses
- Pick the smallest model that passes your tests Sorting and extracting rarely need the largest model
- Send less Search for the relevant passages instead of pasting whole documents
- Use prompt caching Providers discount input that repeats, such as long fixed instructions
- Batch the work that can wait Processing in bulk, for example overnight, is often cheaper
- Cap agent loops A maximum number of steps and a spending limit per run
- Set budgets and alerts Per month and per feature, so a bug can’t run up a bill
- Track cost per item Cost per order or per answer, not just the monthly total
Mistakes to avoid
- Comparing build quotes only. Ask every supplier for a running-cost estimate at your volume, and for what maintenance will involve.
- Using the biggest model everywhere. Test a smaller one on each step. Many steps don’t need more.
- Forgetting review time. If nobody designs the review step, people either skip it or it eats the time you meant to save.
- Trusting pilot costs. A pilot with 50 tidy examples tells you little about 2,000 messy ones a month. Log real usage in the first weeks of production.
What to do next
Write down three numbers for the process you have in mind: how many items a month, how long each takes today and how much checking the output will need. That is enough for a first estimate of both the build and the running costs.
If you want help with it, book a 30-minute call. You’ll leave with what to build first and a rough price.
