What a Custom AI Build Actually Costs in 2026

€3,000 to €50,000 for most of it, and a tenfold spread on the same brief.

A custom AI build for a European SME costs between €3,000 and €50,000 in 2026 for the great majority of first projects, with the figures below breaking that band down by what you are actually building. The number that surprises buyers is not the size of the band, it is the spread inside it: send one brief to five suppliers and quotes of €12,000 and €120,000 come back, and both can be honest. That spread is not negotiating position. It means the five suppliers have made five different assumptions about your integration surface, your data, and how much compliance work sits underneath. They are disagreeing about your business, not about their rates. Which is why the fastest way to compress a quote range is to remove the guesswork, not to haggle.

Custom AI development cost by project type

Custom AI development cost splits by what you are building rather than by how sophisticated it sounds, so the list below is organised that way. These are published European market ranges rather than a price list, and they assume a supplier doing the work at Western European rates with senior people involved. Offshore delivery sits materially below them and a large consultancy materially above. Use them to sanity-check a quote you have received.

• Chatbot or single-purpose assistant: €3,000 to €8,000. One or two workflows end to end. Intake, FAQ deflection, booking, internal lookup. Three to five weeks. The cost is dominated by the conversation design and one or two integrations, not by the model.
• Document processing and routing: €8,000 to €18,000. Extraction from invoices, contracts, forms or reports, plus validation and routing into the system that consumes the output. Five to eight weeks. Each additional source format adds days, and scanned documents of poor quality add a phase.
• Internal assistant over your own content: €9,000 to €25,000. Retrieval over your policies, manuals, tickets or matter files, with permissions that respect who is allowed to see what. The permission model is what makes this cost more than a chatbot, because an assistant that can quote a document the reader should not have seen is a security incident.
• Agent system that acts on other systems: €20,000 to €60,000. Multi-step work with tool use: reading a case, deciding, writing to a CRM, escalating. The build cost climbs because every action needs a guard, an audit line and a reversal path. The run cost climbs more than the build cost does, which is the part buyers miss.
• Trained custom model: €25,000 to €80,000 and upwards. A model trained or fine-tuned on your data, with an evaluation suite and a retraining plan. Above roughly €50,000 the honest comparison stops being to buying software and starts being to hiring, and most companies that ask for this need retrieval instead.

The four-tier version of the same band, cut by complexity rather than by project type, is in our note on what AI consulting costs in 2026. The two cuts agree because they describe the same market from different angles.

Four things that move an AI project cost, none of them the model

Four inputs move an AI project cost far enough to explain a tenfold gap between two honest suppliers. A quote produced without asking about all four has guessed at least one of them, and the guess is always in the supplier’s favour or yours, never neutral.

Integration surface. The single largest multiplier. One system with a documented API and a sandbox is a few days. Four systems, one of which is an in-house tool with no API and a retired author, is a month, and it carries risk that nobody can price accurately in advance. When two quotes differ by 5x, count the systems each one assumed.

Data condition. Digital, accessible and reasonably consistent means you pay list. Scanned PDFs, free-text fields where the meaning depends on who typed them, or a decade of records with no owner means a preparation phase that is a project in its own right. A supplier who quotes without looking at real records has priced an imaginary dataset, and the correction arrives in week six.

Compliance load. A supplier assistant that fails is an annoyance. A system on a regulated path needs a data-protection assessment, documented processing, retention rules, logging that survives an audit, and a human decision point. That work can add half again to an otherwise identical build, and it is the second most common reason two quotes for one brief look like quotes for two different projects.

Who does the work. Day rates across Europe vary by roughly a factor of three, and the same brief delivered by a senior engineer in eight days and by a junior in twenty-five days can land at a similar price with a very different result. Ask who is writing the code and what else they are committed to this quarter, because the answer changes the number and the outcome at the same time.

Where the money goes across the five phases

Across our own engagements the distribution is stable enough to be useful as a check on somebody else’s quote. Discovery and scoping take 15 to 25 per cent of a project. The build itself, meaning the AI behaviour and the interface, takes 35 to 45 per cent. Integration takes 20 to 30 per cent. Rollout, training and documentation take the last 10 to 15 per cent.

Two readings follow from that. If a quote puts 70 per cent of the budget in the build line and almost nothing in integration, the supplier has not looked at your systems yet and the integration cost will arrive later as a change request. If a quote has no rollout or documentation line at all, you are buying software that will be delivered rather than adopted. The week-by-week shape of an implementation maps onto the same five phases, and reading the two together makes an underweighted line easy to spot.

Running costs, and the real cost of a custom AI system over three years

Most quotes stop at go-live. The real cost of a custom AI system is the build plus every month it then runs, and in our experience a fair planning figure for that second part is 15 to 25 per cent of the build cost per year. Over three years it turns a €20,000 build into something closer to €30,000. A supplier who leaves the line blank has not made it free. They have made it yours to discover.

• Model usage. Usually the smallest line, and the one buyers worry about most. At SME volumes it is frequently tens of euros a month. Agent systems break this rule, because a single task can make twenty model calls instead of one, and volume growth hits you multiplicatively.
• Hosting and infrastructure. Servers, database, storage, a queue, backups. Modest and predictable, and higher if the deployment has to stay in a specific region for data-protection reasons.
• Monitoring and evaluation. Somebody has to notice when quality drops. Document formats change, an upstream field gets reused for a new purpose, a model version is retired by its provider with three months notice. Re-running the evaluation set quarterly is cheap; finding out from a customer is not.
• Maintenance. The APIs you integrated with will change. Budget for a few engineer days a quarter even if nothing goes wrong, because something will.

The Biscoito.ai veterinary assistant is a good illustration of why the run line matters more than the build line in the end. It handles 60 to 70 per cent of owner enquiries in under five seconds, around the clock, and it has been doing so continuously. The build was a few weeks. The value came from the months afterwards, and months afterwards is exactly what a zero-euro maintenance line fails to buy. Our other case studies show the same shape, and the capabilities page lists what running one of these actually involves.

Five signs a quote is too low to deliver

• A fixed price arrived before anyone looked at your data. That number is a risk premium wearing a discount. It is either padded by 30 to 50 per cent, or it is a real number for a smaller project than the one you described, and you will find out which in week five.
• The integration line says API integration. No system named, no count, no mention of credentials or a test environment. The supplier is quoting the part they can imagine.
• There is no evaluation set and no measurement plan. Without a baseline nobody can later argue about whether the system works, which suits a supplier and not you.
• Ownership is described in the meeting but not in a clause. Code, prompts, fine-tunes and evaluation data, assigned to you on payment. If that is not in the contract, the low price is a subscription with a deposit.
• Nobody is named. A quote with no named engineer is a quote for whoever is free in March.

Three cases where none of these numbers apply to you

An existing product already does it. Scheduling, invoice capture, transcription and basic support deflection are solved categories with mature tools at per-seat prices. Paying €15,000 to rebuild one of them because it would fit slightly better is a decision you will explain to a CFO in eighteen months.

The annual value at stake is under about €10,000. The procurement cost, your own time in scoping and the run cost will consume the return. This is the most common finding we deliver that clients did not want, and it is the cheapest one to receive.

The data does not exist yet. If nobody has been recording the outcome you want the system to predict, no budget buys the model. What you need first is a change to how the process is recorded, running for a few months. That is a form change, not an AI project, and any supplier who takes your money without saying so has sold you an expensive way to learn it.