Why almost nobody publishes AI pricing
Search for what an AI system costs and you will find the same non-answer everywhere: it depends. There is a reason for that. Scope genuinely varies enormously, and vendors are reluctant to anchor a number before they understand your problem. But the effect is that buyers walk into conversations with no reference point at all, which helps nobody. So let us be useful instead: the ranges below reflect what custom AI development typically costs in the European and Gulf markets in 2026, based on the shape of projects we see. Treat them as a map, not a quote — your actual figure depends on the factors in the next sections.
The three tiers you will actually be quoted
Custom AI work tends to cluster into three bands. A focused single-workflow system — document extraction, a support assistant grounded in your own content, a report generator — typically lands in the €15,000 to €40,000 range and ships in 6 to 12 weeks. A departmental system with multiple workflows, real integrations into your CRM or ERP, user roles and an admin panel usually runs €40,000 to €120,000 over 3 to 6 months. An enterprise platform serving multiple departments, with strict compliance requirements, audit trails and high availability, starts around €120,000 and climbs from there. If a quote sits far below the relevant band, something is being left out — usually integration, security, or the ongoing running cost.
What actually moves the number
Five things drive most of the variance. First, integration surface: connecting to one clean API is cheap; connecting to a legacy ERP with no documentation can cost more than the AI itself. Second, data readiness — if your documents are scanned PDFs of inconsistent quality, budget for a data pipeline before any model work. Third, accuracy requirements: getting to 85% is quick, 95% takes real evaluation infrastructure, and 99% may be scientifically unrealistic for your task. Fourth, compliance: EU AI Act obligations, data residency, or audit logging add genuine engineering, not paperwork. Fifth, the number of human languages — Arabic in particular needs dialect handling and right-to-left interface work that generic vendors routinely underestimate.
The cost everyone forgets: running it
The build price is not the total cost. A production AI system carries recurring costs that need to be in the business case from day one: model inference (per token or per request, which scales with usage), hosting and vector storage, monitoring, and — most importantly — periodic re-evaluation as your data and the underlying models change. As a planning rule, budget roughly 15% to 25% of the build cost per year to keep a system healthy. A vendor who never mentions running cost is either inexperienced or is leaving you to discover it later.
Where budgets get wasted
Three patterns account for most of the waste we see. The first is boiling the ocean: trying to automate an entire department at once instead of the single most painful workflow. These projects tend to stall before anything reaches production. The second is paying custom-build prices for something an off-the-shelf tool already does well — if a standard product covers 80% of your need, buy it and spend the budget on the 20% that is genuinely specific to you. The third is skipping evaluation. Without a test set and a measurable accuracy target agreed before development starts, you cannot tell whether the system works, and you end up paying for rework driven by opinion rather than evidence.
How to sanity-check the investment
The arithmetic is usually simpler than people expect. Take the process you want to automate, multiply the hours it consumes per month by the loaded cost of the people doing it, and compare that to the build cost plus a year of running cost. In document-heavy processes — tender responses, claims handling, compliance checks, contract review — the time savings are typically large enough that a well-scoped system pays for itself in months rather than years. If you cannot construct that arithmetic for a proposed project, that is a strong signal the scope is wrong, not that the price is wrong.
What to ask before you accept any quote
Four questions separate a real proposal from a hopeful one. What is the measurable success criterion, and how will we test it? What exactly is integrated, and who is responsible if the third-party system changes? What is the monthly running cost at our expected volume? And what happens after launch — who maintains it, and under what terms? A vendor who answers all four crisply is quoting a system. A vendor who deflects is quoting a prototype. If you want help scoping your own case, we are happy to walk through the numbers with you before anyone talks about a contract.