Skip to content
DT Nova

Artificial intelligence cost: what an AI project costs to build and to run

Artificial intelligence cost is set less by the technology than by four choices: which approach you take, how much work your data needs, how often the system is used once it runs, and how much oversight and compliance the use case demands. Two AI projects that sound alike in one sentence can differ many times over in cost. This guide explains each part of the budget, the running costs that ordinary software does not have, the EU rules that add work, and how to keep the total under control.

It does not quote our own prices. It explains how an AI budget is built, so any quote you receive, from us or from anyone else, can be compared on the same basis. For software in general, see our guide to custom software development cost; for what building an AI program involves step by step, see how to make an artificial intelligence program. For two common projects in detail, see chatbot development cost and AI agent development cost.

The short version

Every AI budget has five parts:

  1. The build: discovery, integration with your systems, prompts or model work, the screens around it, and testing.
  2. Data: collecting, cleaning, labelling and giving the system safe access to the data it needs.
  3. Running: usage fees for hosted models, or compute for your own, which grow with every request.
  4. Oversight: evaluation, monitoring, retraining and the people who review uncertain cases.
  5. Compliance: data protection, AI Act duties and the records that prove both.

The approach sets most of the cost

ApproachBuild effortRunning costData you need
Rules and automation, no AILow to moderateLow and flatClear rules
AI feature in software you already useConfiguration and trainingA licence or add-on per userWhat the product already holds
Hosted model through an API, with your prompts and rulesModerate: integration, prompts, testingUsage-based, grows with volumeExamples to test against
Hosted model answering from your documents (retrieval)Moderate to high: search index, access rights, testingUsage plus search and storageCurrent, well-organised documents
Fine-tuned modelHigh: training data, training runs, evaluationUsage or hosting, plus retrainingMany good, checked examples of the task
Your own model trained from your dataHighest: data science, compute, deploymentHosting and compute, plus retrainingA large labelled history

The cheapest working approach is usually the right first one. Many problems presented as AI are solved better by rules: if a person can write the rule down, write it down. Where judgement on text, images or patterns is needed, start with a hosted model and your own content, and move to fine-tuning or your own model only when a measured limit, in quality, speed, data location or usage fees, makes it worth it.

Data: the cost that surprises

Most AI budgets that overrun do so on data. Records are spread across systems, fields mean different things in different places, history is incomplete, and the examples a model needs, such as images of defects or correctly classified documents, often do not exist yet and must be labelled by people who know the work. Access is its own task: an assistant that answers from company documents must respect who may see what, or it will show salary files to the whole company.

Ask early: where does the data live, who owns it, how much history is usable, and who will label examples. A short data check before the build is the cheapest way to avoid paying for a model that cannot be trained.

Running costs: the part ordinary software does not have

A conventional system costs roughly the same to run whether it is used ten times or ten thousand times a day. AI does not. Hosted language models are typically priced per amount of text sent and generated, image and speech models per item or per minute, and your own models by the compute they occupy. The bill therefore follows usage, and it grows quietly when a successful feature is rolled out to more users or when each request sends more text than it needs.

The useful measure is cost per transaction, such as the cost of one answered support question or one processed invoice, next to the value of that transaction. Four techniques keep it low: send the model only the relevant part of a document instead of the whole thing, use a smaller model for simple steps and a larger one only where needed, cache answers to repeated questions, and process non-urgent work in batches. Set monthly limits and alerts from day one.

People's time

As with any software, most of the build is hours, and what an hour costs follows local labour costs. Eurostat's release of 31 March 2026 (EU hourly labour costs ranged from €12 to €57 in 2025, read on 11 November 2026) puts average hourly labour costs across the whole economy at €34.9 in the EU and €38.2 in the euro area. Skills are the larger constraint: among EU enterprises that had considered AI but did not use it in 2025, the most common reason was lack of relevant expertise (70.89%), followed by lack of clarity about the legal consequences (52.52%) and concerns about data protection and privacy (48.83%) (Eurostat, Use of artificial intelligence in enterprises, read on 11 November 2026).

Count your own people too. Someone must define what a good answer is, review test results, decide what happens when the AI is unsure, and own the system after launch.

Oversight: testing and monitoring never stop

An AI system is tested against a set of real examples with known right answers, measured before launch and again after every change of model, prompt or data. After launch, quality can drift: products, customers, documents and the models themselves change. Budget for regular evaluation runs, monitoring of errors and complaints, adjustments, and the time staff spend reviewing the cases the system flags as uncertain. That review time is a real running cost, and also the main safeguard.

Compliance: what the EU rules add

The EU AI Act became generally applicable on 2 August 2026. For most business uses, the obligations are modest: the AI literacy rule in Article 4 has applied since 2 February 2025, and as replaced by the AI Omnibus, in force since 27 July 2026, it requires providers and deployers of AI systems to take measures to support the development of AI literacy of their staff, without requiring any specific level for any individual (Regulation (EU) 2026/1744, read on 12 November 2026), and systems that interact with people must tell them they are dealing with an AI, unless that is obvious (Article 50(1)).

High-risk uses cost more. Annex III lists, among others, AI used for recruitment and selection and AI used to allocate tasks or to monitor and evaluate workers (AI Act Service Desk, Annex III, read on 11 November 2026). Under Article 113 as amended by the Digital Omnibus on AI, the high-risk rules apply from 2 December 2027 for Annex III uses and from 2 August 2028 for AI in products covered by Annex I (AI Act Service Desk, Article 113, consolidated text as at 27 July 2026, read on 11 November 2026). Risk management, technical documentation, logging and human oversight then become design requirements, and they are far cheaper built in than added later.

Getting it wrong has a price of its own. Under Article 99, breaching the prohibited practices can lead to fines of up to EUR 35 000 000 or 7% of worldwide annual turnover, whichever is higher, and breaching most other obligations up to EUR 15 000 000 or 3% (AI Act Service Desk, Article 99, read on 11 November 2026). Personal data adds the GDPR: processing likely to result in a high risk calls for a data protection impact assessment before it starts, under Article 35 of Regulation (EU) 2016/679 (read on 11 November 2026).

Build, buy or use what you have

Check the software you already pay for first. If you decide to build, our guide on how to choose an AI development company covers the partner. Office suites, CRMs, help desks and accounting tools increasingly include AI features, and for standard tasks, such as drafting text or summarising calls, a licence is cheaper than a build. A custom build pays off when the task is specific to your business, needs your own data and systems, runs at volume, or must meet rules a product cannot show it meets. Compare over three to five years: licences per user against build, usage and maintenance, plus the cost of switching later.

How to keep artificial intelligence cost under control

  • One use case, one baseline. Measure today's cost and error rate first, so the AI's value can be shown in money.
  • Prove it small. A pilot on real data, run in parallel with the current process, before any wider roll-out.
  • Smallest model that works. Test cheaper models against the same examples; use the expensive one only where it wins.
  • Limits and alerts. Monthly caps on usage fees, and a report of cost per transaction.
  • No lock-in. Code, prompts, test sets and data in your own repository and account, so the model supplier can be changed.
  • Classify early. Decide whether the use case is high-risk before the budget is set, not after.

Questions to ask before accepting a quote

  • Which approach is proposed, and why not a simpler one?
  • What data is assumed, and who cleans and labels it?
  • What will running it cost per month at today's volume, and at ten times that?
  • How is quality measured before launch, and how often after?
  • Who reviews uncertain cases, and how is that time estimated?
  • How is the use case classified under the AI Act and the GDPR?
  • Where is the data processed and stored, and is it used to train anyone else's model?
  • Are the code, prompts, test sets and accounts in your company's name?

How we price AI work

We start with a short discovery phase that checks the use case, the data and the expected volume, then quote a fixed price for a pilot with a written scope and a measured goal. Running costs are estimated per transaction and monitored from the first day. The code, prompts and test sets are yours, data is processed in an EU data region by default, and models are chosen so they can be replaced. See AI consulting for how an engagement runs, chatbot development for assistants built on your own content, and our guide on manufacturing and artificial intelligence for one industry in detail.

Questions people ask

How much does artificial intelligence cost for a business?

It depends mostly on the approach and the volume of use. Using an AI feature that already exists in your software costs a licence. Building on a hosted model through an API costs the integration work plus a usage fee that grows with every request. Training or fine-tuning your own model adds data preparation, compute and the work of keeping the model current. Any figure given without the use case, the data and the expected volume is a guess.

Why does AI keep costing money after launch?

Because most AI is paid per use and needs looking after. Hosted models charge for each request or for the amount of text processed, so the bill follows usage. Models also drift as products, customers and documents change, so they are tested and adjusted, and people review the cases the AI is unsure about. Budget for running costs per month, not only for the build.

Is it cheaper to train our own AI model?

Rarely, at the start. Training your own model needs a large set of good examples, specialist time and compute, and someone to keep it current. A hosted model with your own documents and rules usually reaches a useful result faster. Your own model can pay off later for a narrow, high-volume task where data stays in-house, response time matters or the usage fees grow large.

Does the EU AI Act make AI projects more expensive?

For most business uses the extra cost is small: telling people when they talk to an AI, AI literacy for staff, and good records. High-risk uses, such as AI in recruitment or worker evaluation, carry real obligations from 2 December 2027, including risk management, documentation, human oversight and logging, which add design and testing work. Classify the use case before the budget is set.

How can we keep the cost of AI under control?

Start with one use case and a measured baseline, use the smallest model that does the job, send the model only the text it needs, cache repeated answers, set usage limits and alerts, and track the cost per transaction next to the value per transaction. Keep the code and prompts in your own repository so you can change model supplier when prices or quality change.

Sources