Custom AI development costs: real ranges by system type
Ask three agencies about custom AI development cost and you will get three numbers that differ by a factor of ten, all technically honest. The spread exists because "custom AI" covers everything from a two-week prototype to a regulated production system. So here are the ranges we actually see, by type of system, with the caveats that make them real.
One framing first: you are paying for engineering, not the model. Model calls are cheap and getting cheaper. The money goes to connecting the system to your data and tools, making it fail safely, proving it works with evaluations, and keeping it observable in production.
What custom AI costs, by type of system
An assistant on your own data, meaning a RAG system that answers questions from your documents with citations and respects who is allowed to see what: expect 30,000 to 60,000 euros for a serious prototype on real data, and 80,000 to 200,000 euros to reach production, depending on how messy the sources are and how strict the permission model needs to be. The naive version costs a quarter of that and answers confidently from stale or forbidden documents, which is worse than nothing.
Workflow automation, meaning AI that reads incoming documents or messages, extracts and validates structured information, and pushes results into your existing systems: typically 40,000 to 150,000 euros to production. The extraction is the easy half. The validation rules, the human review queue, and the integration with an ERP that predates the euro are the expensive half.
A multi-agent system, where several specialized agents plan, retrieve, compute, and check each other’s work inside one process: 150,000 to 500,000 euros and up. These numbers are for systems that genuinely need the architecture. Many quoted multi-agent projects are single-agent problems wearing a costume, and the fastest way to cut this budget is to notice that early.
Treat all of these as our reading of the market from projects we scope and deliver, not a universal price list. A vendor quoting far below these ranges is usually selling the demo and leaving out production. Far above, and you may be funding their platform ambitions.
What actually drives the number
Four factors move the price more than anything else. First, integration surface: each system the AI must read from or write to adds real work, and legacy systems without APIs add more. Second, data condition, though less than people fear: AI tolerates messy documents better than classic software, but access rights and completely absent digitization still cost money. Third, the stakes of being wrong: a system whose errors reach customers or regulators needs evaluation sets, guardrails, and review workflows that an internal drafting tool does not. Fourth, compliance: GDPR is manageable by design, but sector rules like healthcare hosting change the infrastructure and the timeline.
Notice what is not on that list: the model, the vector database, the framework. Technology choices move the number by percent. Scope moves it by multiples, which is why the highest-return hour of the whole project is spent scoping the AI project down to one workflow and one metric.
The running costs nobody budgets
After launch, three lines appear that the build quote often omits. Model usage: for most mid-size workloads this lands in the hundreds to low thousands of euros per month, and it can be engineered down substantially with routing and caching, which we covered in your AI bill at ten times the volume. Monitoring and evaluation: someone must watch quality and cost, because AI systems degrade silently, not loudly. And maintenance: models get deprecated, source systems change, the workflow itself evolves. As a planning figure, we tell clients to budget roughly 15 to 20 percent of the build cost per year for keeping the system healthy. A vendor who quotes a build with no running-cost conversation is not being generous. They are not planning for the system to survive.
How to keep the number down
The cheapest project is the short one that proves value fast or dies fast. Start with a 2 to 4 week prototype on your real data, with a metric agreed in advance, and make production funding conditional on the metric moving.
Also compare against the alternative honestly: hiring your own senior AI engineers costs more than most builds and takes two quarters just to start, which is why many mid-size companies land on an embedded external team instead.
A real number, as opposed to a range, takes a scoping conversation about your workflow, your systems and your data. That costs you an hour, not a study.