Most AI projects in SMEs don’t fail because of the AI itself. They fail because of the agency — and the decision on which one to choose is made during the initial consultation, long before the first model is up and running.
This is the uncomfortable side of the current boom. Ever since anyone has been able to access a language model via an API, every other marketing or software outfit has been calling itself “AI agency”. The barrier to entry has dropped to zero, whilst the variation in quality is greater than in any other field we deal with. You’re no longer booking a clearly defined service. You’re booking a promise. And promises are cheap in the 2026 AI market.
You can spot the difference between an agency that delivers and one that only delivers slides within the first twenty minutes of an initial meeting. If you know what to listen out for.
Three paths behind the same label: who ultimately delivers becomes clear during the initial consultation.
What an AI agency actually does – and what it doesn’t
Before you make your choice, it’s worth sorting through the options, because the label ‘AI agency’ covers three very different business models.
The tool reseller essentially sells you access: a ready-made chatbot product, a white-label dashboard, or a licence with a set-up fee. This can be useful if your use case is standard. But you’re paying for a configuration, not for a solution.
The consultant provides strategy: potential analysis, roadmap, use-case prioritisation. Valuable at the outset, but risky as an end point. A roadmap that nobody implements is just an expensive PDF.
The implementation partner does both and doesn’t stop at the concept. They analyse the processes, build the automation, integrate it into your systems, train your team and measure whether the project is paying off. In our view, this is exactly what an AI agency is: not just a slide-production service, but someone who stays involved right through to the running workflow in day-to-day business. If you’re unsure what’s actually feasible in your case, it’s worth taking a look at how a AI consultancy with implementation is structured. The spectrum ranges from process automation through to chatbots and voice agents.
In short: don’t just ask “do you do AI?”, but “will you support me right up to the point where the system is running smoothly without you?”. An honest answer to that question weeds out half the market.
The scope of each business model: from the initial idea to measurable results.
How to spot a good AI agency
There is no single key metric. But there are a few indicators that reliably distinguish substance from window dressing in practice.
Concrete references rather than a wall of logos. Every agency will show you client logos. That means nothing. Ask about a project for a medium-sized business from the last twelve months: what was the problem, what was the scope, what was the outcome, and what went wrong? Anyone who answers this in concrete terms and with a touch of self-criticism has actually delivered results. Anyone who gets vague or reflexively deflects the question by mentioning a corporate name may not yet have delivered anything at your level.
A pragmatic starting point. A good AI agency won’t try to sell you a twelve-month transformation programme for 2026, but will start small. A proof of concept that demonstrates, within two to four weeks, whether the AI actually solves your specific problem beats any glossy strategy document hands down. This reduces both your risk and theirs. And it separates those who can deliver from those who merely plan.
Skills transfer is built in, not optional. The crucial difference between a partnership and a dependency: does the knowledge leave the premises along with the invoice, or does it stay with your team? Reputable providers document, hand over and train your team, for example through accompanying AI workshops for your team. An agency that deliberately keeps you in the dark to secure follow-up contracts is optimising its turnover, not your bottom line.
Compliance is factored in from the outset, not added as an afterthought. AI without data protection and a legal framework is not an option for German SMEs. The EU AI Act is being phased in gradually: on 2 August 2026, the transparency obligations under Article 50 will come into force — chatbots and voicebots must then disclose the AI interaction. The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), on the other hand, has officially postponed the deadlines for high-risk applications: to the end of 2027 or 2028, depending on the system class. The GDPR already applies to AI data flows. An agency that only mentions these issues when asked is exposing you to a risk that will later prove more costly than the entire project. Ask early on how they handle personal data, model hosting and documentation requirements.
Measurability before marketing. Three factors determine whether AI is worth it for you: the time to the first productive result, the hard ROI in hours or euros saved, and whether your team can continue working independently afterwards. An agency that refuses to specify these factors before the project begins is selling you a feeling, not an investment.
Red flags: When you’re better off hanging up during the initial meeting
A few patterns are so reliable that they deserve their own warning.
Anyone who starts with “AI is revolutionising everything” and, after twenty minutes, still hasn’t mentioned a single concrete use case for your business is selling hype.
Anyone who refuses to quote a price until you’ve booked a “discovery workshop” lasting several weeks is setting a cost trap.
Anyone who promises “proprietary AI” but is essentially just passing on a standard language model with their own logo isn’t being honest about the value they’re adding.
And anyone who gives you guarantees of success – “95 per cent accuracy, guaranteed” – either doesn’t know your data or is counting on you never verifying the figure.
None of these points on its own is a deal-breaker. But if two of them come up in the same conversation, you’re talking to a salesperson, not someone who will actually build something in the end.
AI SEO Agency: the special case of AI visibility
One special case deserves its own attention because it’s currently being touted as the hottest sub-segment: the AI SEO agency. What’s meant here is Generative Engine Optimisation (GEO), the discipline of ensuring that your brand actually appears in the responses from ChatGPT, Perplexity, Google AI Overviews and the like.
This is not traditional SEO with a new label. The levers are different: structured data, citable content, technical crawlability for AI bots, and clean server-side rendering. An agency that sells you GEO but only talks about keyword density and backlinks has not grasped the concept. Ask specifically how they measure AI visibility . If the answer remains vague, it’s false advertising.
If you want to delve deeper into this, you’ll find our article on the subject, whether AI can find your website at all, the five pillars that determine this. With the free AI Visibility Check, you can see the current status of your domain in just a few minutes. Market rates for ongoing GEO will be noticeably higher than for traditional SEO by 2026. For medium-sized clients, monthly fees in the four- to low five-figure range are standard. This is no cause for panic, but it is a reason to take a close look at what is actually being done for this money.
What an AI agency costs and how to interpret the pricing model
A reputable price range is broad, and that’s fine as long as it’s transparent. A one-day strategy workshop typically costs in the low four-figure range, in line with industry standards. A proof of concept is in the region of around 5,000 to 15,000 euros. A complete initial project comprising a potential analysis plus a productive use case typically falls within the mid five-figure range, whilst large-scale integrated projects are significantly higher.
More important than the absolute figure is the structure behind it. Hidden costs for data preparation, dashboards or ‘model maintenance’ that do not appear in the quote are the real red flag. A fair model makes it clear what constitutes a one-off cost and what involves ongoing expenditure, and ideally links part of the remuneration to measurable results. If an agency is unwilling to calculate the ROI for you, this is rarely a sign of discretion. More often than not, there is no substance behind it.
Start small, then scale up: typical market sizes at each stage.
The questions you ask in the initial meeting
If you take just one thing away from this article, let it be this concise list. Ask them during the initial meeting, and the wheat will be separated from the chaff before a contract is on the table:
- What comparable SME project have you implemented in the last twelve months, and what went wrong with it?
- What is your smallest possible first step, and how much does it cost?
- How do you ensure that my team can continue working independently after the project?
- How do you deal with the GDPR and the EU AI Act in this specific case?
- Which three key performance indicators will we use to jointly assess whether the project was worthwhile?
An agency that answers these five questions without evasion is worth at least a second meeting. One that flounders on three of them has just saved you a lot of money and stress.
What matters in the end
The best AI agency isn’t the one with the largest model vocabulary or the flashiest demo. It’s the one that takes your use case more seriously than the technology behind it: the one that starts small, delivers measurable results, gets your team on board and, in the end, leaves you with a solution that keeps running even without them. Anything else is simply a well-packaged risk in today’s market.
If you’re currently facing this decision and want an honest initial assessment of your specific use case, rather than a pitch deck: that’s exactly what an initial consultation is for. Bring along these five questions. We’ll be happy to answer them and will also let you know if AI isn’t the right solution for your problem at the moment.