Hiring an AI Agency: 12 Questions for the First Meeting
This article was created with AI assistance
The text and images in this article were generated with the help of AI systems. Labelled in accordance with Art. 50(4) of the EU AI Act. Responsible for publication: ArkeonTech.
Choosing an AI agency is a decision without a benchmark for many mid-sized companies: you buy something you cannot fully assess technically, in a market that is growing fast and is hard to survey. The good news is that you do not need to be a developer to separate substance from sales talk. You just need to ask the right questions. These twelve matter most.
Key takeaway: Four areas decide the outcome: integration depth (does the system write into your software or stay an island?), data protection (EU hosting, data processing agreement, EU AI Act), cost structure (fixed price or open bill, minimum term, who owns the result) and operation (who maintains the agent after go-live?). Get every answer in writing before you sign.
Why the right questions decide project success
Because most AI projects fail not on technology but on what was never clarified beforehand. Research from MIT Project NANDA (2025) concludes that around 95 percent of the generative AI pilot projects studied deliver no measurable return. The core problem it identifies is not the quality of the models but systems that neither adapt to nor learn from the company's workflows.
That is exactly what a first meeting can reveal. Asking the following questions usually tells you in the first meeting whether a provider sells a tool or solves a problem.
Technology and integration: questions 1 to 3
1. Does the AI agent write into our existing systems, or does it only produce text? This is the single most important question. An agent that answers an enquiry but does not write the lead into the CRM and does not book an appointment creates follow-up work instead of saving it. Ask which systems will be connected: CRM, ERP, calendar, inventory, phone system.
2. What happens when the agent cannot answer a request? A good system recognises its limits and hands over to a human, including the conversation history. A poor one guesses. Ask about the escalation path and how often it triggers in comparable projects.
3. Where does the agent get its knowledge, and who keeps it current? Products, prices and opening hours change. Clarify whether the knowledge base is maintained manually, who does it and whether it is included in the price. An overview of system types is in our article Which AI agents exist.
Data protection and law: questions 4 to 6
4. Where is data processed, and is there a data processing agreement? If the provider processes personal data on your behalf, you need a data processing agreement under Art. 28 GDPR, and the same obligations must be passed on to every sub-processor. If processing takes place outside the EU, a legal basis for the transfer is also required, such as an adequacy decision or standard contractual clauses. Ask to see the contract and the list of sub-processors before commissioning, not afterwards.
5. How does the solution meet the EU AI Act transparency obligations? Since August 2026 the transparency obligations of Article 50 of Regulation (EU) 2024/1689 apply: users must be able to recognise that they are interacting with an AI system. A provider who cannot answer this confidently has not engaged with the topic. Details in our EU AI Act guide for SMBs.
6. Will our data be used to train third-party models? The only acceptable answer is no, contractually guaranteed. In sensitive sectors, professional confidentiality applies on top, for example under Section 203 of the German Criminal Code in healthcare.
Cost and contract: questions 7 to 9
7. Is this a fixed price, and what exactly is included? Ask explicitly about concept, integration, testing, training, operation and model costs. Channels such as WhatsApp add usage-based third-party fees that should be calculated upfront. How this breaks down is shown in our articles on AI chatbot costs and the WhatsApp Business API.
8. Is there a minimum contract term? A long commitment is a risk for a tool you can only evaluate in operation. Clarify the term, the notice periods and what happens to data and configuration when you cancel.
9. Who owns the result, and what happens if we switch providers? Clarify usage rights, data export and documentation. If switching means starting from scratch, that is a cost risk not stated in the quote.
Collaboration and operation: questions 10 to 12
10. Who will actually work on our project? The people you speak to in sales are not always the ones who implement later. Ask who is technically responsible for the project and who your contact will be after the contract is signed.
11. What is the first measurable milestone, and when does it arrive? Good projects start with a narrowly defined use case and a testable result within a few weeks, not with a complete overhaul. Ask when you will first see something real.
12. Which metric should change, and how do we measure it? Without a defined target, success cannot be assessed afterwards. Sensible metrics include enquiries answered outside business hours, processing time per case, or the rate of missed calls.
Which answers should make you cautious?
Four patterns that experience shows point to problems, regardless of provider:
| Answer in the meeting | Why it is a warning sign |
|---|---|
| "We will sort that out later" on data protection | Retrofitting GDPR compliance is expensive or impossible |
| No concrete naming of systems to connect | Suggests an island solution without process integration |
| No traceable cost structure | Makes comparison and budgeting difficult |
| Success promises without a metric | Without a metric there is no yardstick later |
A transparent provider answers all of this in the first meeting without hesitation, and in writing in the quote.
Regional agency or national provider?
Distance matters less often than a search for an agency in your own city suggests. It counts mainly at the start: when an agent reaches into established workflows, such as order entry or appointment booking in a medical practice, half a day at the workplace uncovers exceptions that no process description contains. Build, testing and operation then run over shared screens and ticket systems in most projects anyway.
Three questions help to weigh it up:
- Does the provider come on site for the project start, and is that included in the quote? Travel costs that only appear on the invoice distort every comparison.
- Are there references from your industry? A comparable project from another region tells you more than an unrelated project nearby.
- How quickly does the provider respond during operation? What counts is the agreed response time for incidents, not the travel time.
Once these three points are settled in writing, regional and national offers can be compared using the same twelve questions.
Frequently asked questions about choosing an AI agency
How do I recognise a reputable AI agency? By three things: it names concretely which of your systems will be connected, answers data protection questions immediately and in writing, and proposes a defined first use case with a measurable target instead of a complete project.
What does hiring an AI agency cost? It depends on the use case and the integration effort. Reputable providers list the one-off costs for implementation and the running costs for operation, model usage and third-party fees such as WhatsApp separately. A quote without that split cannot be compared.
Should I choose a local agency or a national provider? It depends on complexity. For deep process integrations, being able to sit down on site and review the workflows together helps. For clearly defined standard cases, distance matters less.
How long does it take from commissioning to go-live? It depends on the scope. A clearly defined first use case should be testable within a few weeks; a deep integration with an ERP system or phone system takes longer. Ask for the date of the first testable milestone in the quote.
What must an AI agency quote contain? The scope with all cost components, the systems to be connected, the data protection framework including the processing agreement, the rules on usage rights and data export, the term, and the agreed success metric.
Do I need my own technical staff for operation? Not necessarily. If the provider handles hosting, monitoring and adjustments, your team mainly works with the results, such as enquiries in the CRM or booked appointments in the calendar. Anyone who wants to run the system themselves needs their own capacity for it. Specify in the quote who takes on which task.
Sources
- MIT Project NANDA (2025), reported by Fortune: MIT report: 95% of generative AI pilots at companies are failing
- European Union (2024): Regulation (EU) 2024/1689 (EU AI Act), Article 50
AI chatbot for sales & support
Answers customer enquiries in seconds, qualifies leads and hands over to your team - live in 2-4 weeks.
More Articles
Which AI Agency Fits a Company Without Its Own IT Department?
Four provider types compete, but only one reliably works without in-house IT staff: the agency with contractual operation. What consulting, handover and managed operation each mean, what the models cost in 2026 and which five points reveal in a first call whether the provider actually runs the system.
AI Agents Alone Don't Drive Growth: Strategy, Brand & AI
A chatbot or voice agent rarely lifts revenue on its own. Why real growth only emerges when strategy, brand and AI work together as a system.