Five Types of AI Partner and Which One Fits You

Pick the category first. The shortlist writes itself after that.

Five types of AI development partner exist and they are not competing for the same work, whatever their websites imply. A large consultancy, a boutique specialist, an offshore development studio, an off-the-shelf product vendor, and hiring in-house are five different answers to five different situations. Most bad procurement decisions happen because a buyer compared four suppliers from three categories on price, which is a comparison that cannot be won. Sort by category, then compare inside it. Three inputs decide the category: project size, data sensitivity, and whether you have internal capacity to carry anything afterwards.

What each type of AI development company is actually for

Costs below are published European market ranges rather than any single firm’s price list, and they move with country and seniority. Use them to sanity-check a quote, not to budget. One caveat before the list: the categories blur at the edges, because a large consultancy will happily sell you a small project and a boutique will take on more than it should when the year is quiet. What you are choosing is a delivery model, and the honest signal is the shape of the engagement rather than the label on the website.

• Large consultancy. Roughly €150,000 upwards, six to eighteen months, with partners selling and mixed-seniority teams delivering. Genuinely best at board-level change programmes, multi-country rollouts, and any situation where governance requires a recognised name on the paper. Should not be chosen for a single workflow automation: the engagement model has a floor that your project sits well below.
• Boutique or specialist. Roughly €3,000 to €50,000, four to twelve weeks, with senior people writing the code because the firm is too small for that to be anyone else. Best where the workflow is specific to your business and speed matters. Should not be chosen when you need twenty people on site, or when a procurement policy requires a supplier above a revenue threshold you will not meet.
• Offshore development studio. Frequently 40 to 60 per cent below Western European day rates, with capacity to scale headcount fast. Best when the specification is genuinely complete and the work is build-to-spec. Should not be chosen when the requirements are still being discovered, because the cost advantage is spent on specification churn across a time-zone gap.
• Off-the-shelf product vendor. Per-seat or per-usage licensing, live in days. Best for solved categories: scheduling, invoice capture, transcription, basic support deflection. Should not be chosen when the process is your competitive advantage, because you will be configuring toward a product roadmap you do not control.
• Hiring in-house. Roughly €60,000 to €110,000 a year per engineer in most of Western Europe, plus three to six months to hire and ramp. Best when AI is continuous rather than a project. Should not be chosen for your first system: you will be hiring against requirements you cannot yet write, and one engineer alone has nobody to review their work.

The seniority question that cuts across all five

One dynamic runs through every category and rarely appears in a proposal. The people who sell the work are frequently not the people who deliver it, and the gap between the two widens as the supplier gets larger. This is structural rather than dishonest: a firm of four hundred cannot put its most experienced person on every engagement, and the commercial model depends on spreading senior time thinly. It is still your risk to price.

The test is the same whoever you are talking to. Ask which named individuals will do the work, what their other commitments are this quarter, and whether the key person can be named in the contract. Large firms will often decline the last one for legitimate reasons and should tell you so plainly. A boutique that declines it has a different problem, because there is nobody else to be doing the work. The twenty-one questions to ask before signing put this at number 17, along with the ownership and failure questions that tend to get asked too late.

A decision path that takes about ten minutes

Start with project size. Under roughly €10,000 of value at stake, buy a product or do nothing; the procurement cost of anything custom will eat the return. Between €10,000 and €150,000, you are in boutique or offshore territory. Above that, with multiple countries or business units involved, a large consultancy becomes a defensible choice even at the higher price.

Then apply data sensitivity. Special-category personal data, trade secrets, or anything that would need to stay in a specific jurisdiction narrows the field sharply, and it usually rules out the cheapest offshore options for reasons that are contractual rather than technical. A partner who cannot answer where processing happens, in which region, has answered the question.

Then check internal capacity. This is the input buyers skip and it decides more than the other two. If nobody in your company can own the system after handover, a build-and-leave engagement produces an asset that decays quietly for a year. Either buy the product instead, or budget for the hire alongside the build, or choose a partner who will run it and accept that you are buying a service rather than an asset. Our case studies are all builds where somebody on the client side owned the result, which is not an accident.

Four axes that separate suppliers of the same type

Once the category is settled, capability stops being a differentiator. Every firm in the boutique bracket lists the same models, the same cloud platforms, and the same frameworks, because those are commodities. What varies is behaviour, and four axes expose it.

• Scope discipline. Does the firm reduce your brief before quoting, or accept it whole? A supplier who quotes your entire wish list without argument has told you they will build whatever is asked and let you discover what was unnecessary.
• Production references. Not a portfolio of demos. A system that has been running for more than six months, with somebody willing to say so on a call. The distinction matters because most AI failure happens after month three, not during the build.
• Ownership terms. Code, prompts, fine-tunes and evaluation sets, all assigned to you on payment, in one clause. Split ownership is the most common way an engagement becomes hard to leave.
• Willingness to be measured. Whether they propose the success metric themselves, and whether they will measure the baseline before building rather than reconstruct it afterwards.

Those four are expanded, with the reasoning behind each, in our guide to comparing custom AI development companies. They apply equally well to a large consultancy or an offshore studio: only the price bracket changes.

Four situations where we are the wrong call

We are a boutique specialist, so treat the following as the honest version of our own limits. Do not hire a firm like ours for a multi-country rollout needing coordinated teams in six markets: we would subcontract, and you would be paying a margin for the privilege. Do not hire us when your board requires a recognised name for governance cover, since that requirement is real even when it is not technical, and a boutique cannot satisfy it.

Do not hire us when an off-the-shelf tool already solves your problem. We say this in a meaningful share of first conversations, and the data readiness assessment frequently surfaces a version of it, where the honest recommendation is a form change rather than a model. And do not hire any external partner when the actual blocker is that your leadership has not agreed what the priority is. A partner will build a competent system pointed at a direction nobody settled, the system will be blamed, and the next AI proposal in your company will be harder to pass. That is what the audit is for, and it is the cheapest of the five options by a wide margin.