AI Bureau: What It Means and What to Expect

A category, not a qualification.

An AI bureau is a small firm that scopes, builds, and runs AI systems for other companies, usually with the same people doing all three. The word is Danish, Dutch, and Norwegian for what English calls an agency, and it carries a useful implication the English word lost: a bureau is somewhere you send work, not somewhere you buy a licence. Nothing about the label is regulated. Anyone can print it on a website. So the practical question is never whether a firm calls itself a bureau, it is which of four engagement shapes it actually sells, and whether that shape matches the problem you have.

Why the AI agency label sorts nothing

Three very different businesses use the same word. The first is a rebranded software agency: strong engineers, real delivery record, and AI added to the service list in 2023 because clients started asking. The second is a reseller: a thin layer over one vendor’s platform, paid on licences, technically capable of configuration and not much more. The third builds custom systems and takes responsibility for whether they work in production. All three answer the phone the same way.

You separate them by asking what happens after launch, because that is where the three diverge sharply. The rebranded agency hands over a codebase and moves to the next project. The reseller keeps billing the licence and has limited ability to change model behaviour when your process changes. The third stays close enough to see the system drift and fix it. None of these is dishonest. They are different products, and the word bureau covers all of them, which is why the label sorts nothing on its own.

The engagement models you will be offered

Most European bureaus sell some version of these four. The differences matter more than the day rate.

• Paid discovery, then a decision. A fixed-scope assessment, typically one to three weeks, that ends in a written recommendation and a costed plan. You own the output whether or not you continue. The point of paying for it is that a bureau with no build attached has no reason to recommend a build.
• Fixed-scope build. One defined system, fixed price, fixed deliverables. Works when discovery has already happened and the requirements are genuinely settled. It goes wrong when the scope was guessed, because every change becomes a negotiation instead of a decision.
• Retained capacity. A monthly allocation of engineering time against a roadmap you steer. Suits companies running several systems, or one system that keeps evolving. The risk is drift: without a quarterly review it becomes a subscription nobody audits.
• Build then transfer. The bureau builds, documents, and hands the system to your own team, with a defined support window. The only model that ends with you independent, and the one to ask for if you have or plan to hire technical staff.

Published European market rates for custom AI work sit roughly between €3,000 for a narrow single-workflow build and €50,000 for a multi-role platform, with most mid-market projects landing between €8,000 and €35,000. Those are market ranges rather than any one firm’s price list. What moves a quote inside that band is rarely the model: it is the number of systems you need integrated, the state of your data, and how many people must approve the output. Our breakdown of AI consulting cost goes through the drivers line by line.

Bureau, software agency, or management consultancy

Give the same problem to the three category types and you get three different products at three different prices. Comparing them on four dimensions is more useful than comparing logos.

• Who does the work. At a bureau, the person who scoped it usually writes the code, because the firm is too small for that to be anyone else. At a large consultancy the partner who sold the work and the team who deliver it are different people, often at very different levels of experience. That gap is a category risk, not an accusation: it is structural, and you should ask about it directly.
• What you own at the end. A bureau engagement usually ends with you holding the code and the documentation. A platform reseller ends with you holding a licence. A consultancy frequently ends with you holding a strategy document and a recommendation to procure. All three can be right. They are not interchangeable.
• How the money works. Bureaus bill for time or fixed scope, so their incentive is to finish. Resellers bill recurring licences, so their incentive is renewal. Consultancies bill for people on the ground, so their incentive is duration. Read any proposal with the vendor’s incentive in mind and the odd clauses start making sense.
• What happens when it breaks at 9am on a Tuesday. Ask for the actual escalation path and the actual name. Small firms tend to answer this well because the answer is a person. Large firms answer it with a process, which is more durable and slower.

Company size is the honest tiebreaker. Below roughly 500 people, a bureau usually wins on speed and on senior attention. Above that, or where a board demands a recognised name for governance reasons, the large firm is frequently the correct answer even when it costs more, and a bureau that pretends otherwise is selling you something.

What actually happens in the first quarter

Weeks 1 to 2, discovery. Interviews with the people doing the work, not only the sponsor. A bureau that only talks to the executive who signed the contract will build the process as management imagines it, which is reliably not the process. Output is a written map of where time goes and which steps are candidates.

Weeks 3 to 4, scoping and a hard conversation. The candidate list gets cut, usually hard. This is where a competent bureau tells you which parts of your idea are not worth building, and it is the single most useful hour of the engagement. Roughly a third of the audits we run end with us recommending less than the client came in asking for.

Weeks 5 to 10, build. One workflow, end to end, in production with real users. Not five workflows at demo quality. The WA Center platform we built for an education institution went live this way, one messaging workflow first, with the additional roles and reporting layered on afterwards; the WA Center case study shows how that sequencing worked in practice. Depth through one workflow beats breadth across five, because only the deep one survives contact with actual users.

Weeks 11 to 12, handover and measurement. Documentation, access, and a number compared against the baseline taken in week one. If nobody wrote down the baseline, nobody can prove the system worked, and that omission is common enough to be worth checking for in the proposal. You can see the range of shapes these engagements take across our case studies, and what we build in general on the capabilities page.

Three situations where you want somebody else

Hire a management consultancy instead if the actual problem is that your leadership team disagrees about strategy. A bureau will build you a competent system that serves a direction nobody has agreed on, and the system will be blamed. Hire a general software agency instead if the work is mostly conventional software with a small AI feature at the edge: paying AI-specialist rates for CRUD screens is a waste, and an honest bureau will say so.

Buy an off-the-shelf product instead if your problem is genuinely standard. Scheduling, invoice capture, transcription, and basic support deflection are solved categories with mature tools, and a custom build has to beat a product that thousands of companies already fund. We tell clients to buy the tool in maybe one conversation in five. The cases where a bureau earns its cost are the ones where the workflow is specific to your business, the integration surface is yours alone, or the output has to carry your organisation’s judgement rather than a generic model’s.