FAQ
Frequently asked questions
We answer the questions teams ask before an AI project: cost and time, your data and the law, the technology, and how to work with M-Squad.
What holds companies back on AI? Bitkom asked 603 German companies in 2026 (opens in a new tab). Those not yet using AI named three barriers most: missing know-how (85%), legal uncertainty (66%), and costs that are hard to calculate (63%). The first three blocks answer them.
Getting ready for an AI project
Is this a real AI use case, or would ordinary software do?
Ask that first. Make or buy still applies: if a product on the market does the job, buy it. If fixed rules solve the task, a script is cheaper, faster and easier to test. Messy input, such as language, documents, images or judgment calls, can make AI useful. It does not make AI the right choice by itself. In many good systems, plain code does most of the work and AI handles the exceptions. Sometimes the best AI project has very little AI in it.
What is a good first AI project?
Small and self-contained. One journey, a narrow set of inputs and outputs, and data you already have. It should show a result within weeks and have an expected return you can measure. That keeps the blast radius small and gives you a reviewable first step.
Check your idea against the first-step checklist →
Is our data good enough for AI?
We find that out before you spend on a build. Discovery includes a check of your data: where it lives, how clean it is, and who may use it. Many first projects need less data than you expect, because the model is not trained on it. It only works with it. If the data is not there yet, we say so and name what is missing.
We have little AI know-how in-house. Can we still start?
Yes. We bring the AI know-how: choosing the use case, the architecture, the evals and the delivery. Your team brings what it knows about your product and your customers. We build with your people, so the know-how stays with you after the project.
What do you need from our side?
Three things. People: someone who can decide, and someone who knows your systems and data. Access: the relevant APIs, a test environment and realistic test data. Time: for questions, reviews and feedback, most of it in the first weeks. For AI tooling, we bring our own assistants and licences, or work inside yours if your policy requires it.
Cost, time and what you get
What does an AI project cost?
A small project with a clear scope can start at €2,000 net, at a fixed price. One example: a small tool that analyses an agreed export of app-store reviews or customer feedback. It groups recurring themes, scores sentiment and suggests actions, each linked to the comments behind it. Ongoing data collection, integrations and hosting are agreed separately. Larger work runs in phases, at a fixed price or on time and materials. Before each phase starts, we agree what you get, how we check it, how we bill it and what is out of scope. We also agree what we need from your team. If a phase misses its agreed result, the contract and the billing model decide what follows. After each phase, you decide whether to continue, change direction or stop. A first phase should earn the next one.
What does the AI cost to run once it is live?
Building and running are two bills. To build and maintain something small, budget about €100 a month for an AI subscription. That covers a regular user. A heavy user needs more. For a live product, budget for the whole workflow: model usage, hosting, storage, integrations, monitoring, maintenance and your team’s checking time. A cheap answer that needs correcting can be an expensive result. Automated evals can catch errors before a person has to. In a limited live test, we measure which errors they catch and how much checking time remains. We use smaller, cheaper models, or plain code, where they meet your quality bar. Before go-live, we estimate the monthly cost at your expected volume and at higher volume, including calls that have to be repeated. We can set budget limits and spending alerts, and agree what happens at a limit: stop, fall back, or ask for approval. An alert alone does not stop the bill.
How do we know an AI project will pay off?
You do not know in advance. So we estimate, then test the estimate. First, we measure your current workflow: volume, time, errors and cost. Then we compare it with the new one, including the cost to build, run and check it. Time saved is not automatically money saved. It counts when your team can use the time, cut spending or improve a result that matters. Additional revenue needs evidence too. Before the pilot, we agree with you on the baseline, the target, the payback period and the result that makes us stop. A good demo is not a business case.
How long does an AI project take?
Discovery typically takes about two weeks. A working prototype follows in a few weeks, depending on scope. Production takes longer, because security, testing and a limited live test come first.
Who owns the code?
You get all the source code, prompts and evals, with the right to use, change and extend them, with us or without us. You are not tied to us. We keep the right to reuse the general building blocks and methods we develop along the way. That never includes your data, your business logic or anything confidential. Third-party and open-source components keep their own licences. If you need exclusive rights, we agree that before the project starts.
Your data and the law
Is our data used to train AI models?
That depends on the provider, the product, its settings and the contract. We check them before any of your data goes into a tool. We check three separate things: whether your data may be used for training, how long it is stored, and who can access it. “Not used for training” does not mean “not stored”. The same check covers connected tools and logs. Our default: the model works with your data at run time, without training on it. Fine-tuning a model on your data is a separate decision. It needs a clear benefit, the rights to the data, and agreed controls.
Where is our data processed? Can it stay in the EU?
We clarify that before the project starts, for every system and model in the flow. Where your data has to stay in the EU, we design the flow for it. The options include EU-hosted models, self-hosted models, and on-device AI, which runs on the phone. We validate the setup in a small test before we commit to it. Some widely used AI tools process data outside the EU. If that is not acceptable for your project, we switch tools.
How do you keep our data out of the wrong hands?
The AI gets only what the task needs. Agents work with scoped permissions, not broad access. Sensitive actions need approval, and requests are logged. Before the project starts, we agree which tools may see which data. We sign an NDA if you want one.
Do you sign a data processing agreement (AVV)?
Yes, wherever we handle personal data for you. Real customer data in a back-test is the typical case. We then act as your processor under GDPR, and the agreement lists the tools and sub-processors that may touch the data. Where possible, we work with anonymised or synthetic test data, so less personal data is in play.
How do you handle GDPR and the EU AI Act?
We design for the rules from the start and prepare what your reviewers need. That means data flows, an EU AI Act risk classification, action logs, and the technical input for a data protection impact assessment (DPIA). GDPR requires a DPIA when processing is likely to put people’s rights at high risk. The AI Act’s transparency duties have applied since 2 August 2026, for example telling users that they are talking to an AI. The rules for stand-alone high-risk systems apply from 2 December 2027. Which duties apply depends on your use case and your role, so we check instead of assuming. Where a question needs a lawyer, we bring in legal experts for AI and data protection. The decision stays with your company, advised by its data protection officer and legal team. We are engineers, not lawyers.
What about our works council and our employees?
Involve both early. In Germany, the works council has co-determination rights on technical systems that can monitor staff behaviour or performance. Many AI tools can. We join those talks and explain what the system records, what it does not, and who sees what. What is acceptable is for your company and its works council to agree. Acceptance grows when the people who will use the tool test it early and help shape it.
Technology and quality
Does it work with our existing app and systems?
In most cases, yes. We build on your existing app, APIs and identity systems where possible, so you do not have to rebuild first. We can add an agent to an existing mobile app or backend. For broader access, an Agent Gateway sits in front of the APIs you already have.
How do you choose AI models, and can we switch later?
We choose the model per task, by quality, cost, speed and where your data may be processed. That can be a cloud model, a self-hosted model or AI on the device. The most powerful model is not automatically the best choice. We pin model versions where the provider allows it. Because evals test the same tasks, you can check a newer or cheaper model before you switch. Switching still takes work: prompts, connections to your systems and answers can behave differently. So we recheck quality, permissions, speed, cost and data handling before a replacement goes live.
What happens when the AI gets it wrong?
It will, sometimes. AI is not infallible, and the same question can get slightly different answers. So your process needs some flexibility, and someone on your team should own the cases that need human judgment. We agree with you which mistakes are acceptable and which are not. A draft your team edits can live with small errors. An unauthorised change to a customer account cannot. Before the build, we define evals: repeatable tests based on real tasks and the results you would accept. Every change runs against them, and we agree which failures block a release. A high average score does not excuse a serious failure. In the product, we design the AI to ask for clarification, hand over to a person, or stop. Risky actions need approval from an authorised person, actions are logged, and the first release goes to a small group of users. We measure the remaining errors and checking effort before you decide whether to expand.
What is an Agent Gateway?
A secure action layer between AI agents and your product. It lets customer AI, partner agents and voice assistants use selected app and product capabilities. It works through MCP (the Model Context Protocol) and your existing APIs where possible. Agents do not get direct write access by default. For sensitive actions, the agent proposes, the user approves, the gateway executes, and everything is logged. You can start small, with a Lighthouse Project: one valuable journey and a small set of safe actions.
Working with M-Squad
How does an AI project run, from first call to go-live?
In small steps, each with a result you can check. Steps one to four are discovery, which typically takes about two weeks.
- Map and select. We list candidate projects with you and pick the one with the best expected return.
- Discover. We check your data, then define the architecture, APIs, security and privacy concepts, performance targets, and the test plan with its test data.
- Define evals. We agree the test cases that show what a good result looks like.
- Make, adapt or buy. We check what the market and open source already offer before we build anything new.
- Plan in phases. We split the work into phases and sub-projects that can run in parallel.
- Build and test. We test, measure quality and the cost of model usage, tune, and test again.
- Run a limited live test. Internal testers or a small group of real customers use it first.
- Go live. We monitor, fix and improve, with a roadmap for what comes next.
What if the pilot fails its targets?
Then the pilot, the limited live test, has not earned a wider rollout. We compare the result with the targets agreed before the pilot. We show where it fell short: quality, cost, speed, adoption or business benefit. If a specific change is worth testing, we agree its scope and billing before doing more work. If the evidence says the use case does not work, we recommend stopping. You decide. Either way, you have committed to one agreed phase, not a rollout. A pilot can produce a useful “no”. Finding that out while the scope is small beats spending more to defend the original idea.
Will you sign an NDA?
Yes. We sign one before the first detailed conversation if you want.
How do we get started?
With a short call. Tell us what you want to change and what you already have. If a small first project makes sense, we propose a scope and a fixed price. Call Daniel +49 (179) 691 3629 or use the contact page at m-squad.com/contact.
Who maintains it after launch?
Usually your own team. We build with your people, so by launch they know the code, the evals and the release steps. You get a defined handover. After that we are there when you need us, for a model update, a fix or the next phase. If your lead at M-Squad is unavailable, a colleague steps in. AI providers change and retire models. Under a maintenance contract, we re-run the evals and handle the switch. Without one, we do it on request.
About M-Squad
What does M-Squad do?
M-Squad is an AI-first mobile studio in Munich. We help product and engineering teams shape, prototype, and deliver mobile apps, AI agents, Agent Gateways, and digital product systems, using AI as a force multiplier for your whole team, across Germany, DACH, and Central Europe.
What does "AI-first" mean at M-Squad?
AI accelerates the well-understood parts of the work, boilerplate, glue code, tests, adapters, translations, and documentation, so senior engineers can focus on architecture, edge cases, security and product. It is the opposite of "vibe coding": people decide the architecture before code is written, every change goes through review gates set at a production bar, and specifications and the type system are the source of truth.
Where is M-Squad based?
In Munich, Germany (Schellingstr. 141 Rgb., 80798 München). We focus on Central Europe / the DACH region (Germany, Austria, Switzerland).
Who runs M-Squad?
M-Squad GmbH was founded in 2001 and is led by Co-Founder and Managing Partner Daniel Melter, who co-founded and chaired MobileMonday Germany (2006–2013) and co-founded GenAI Wednesday in Munich.
Who does M-Squad work with?
Mobile operators and MVNOs, established consumer brands (media, entertainment, travel, retail, automotive, financial services), mobile businesses digitising their workforce, and emerging mobile firms entering or expanding in the German market.
How fast can M-Squad deliver?
In one internal sprint, one person rebuilt the core of a mobile operator self-care app in four days. The working prototype included an AI assistant, voice and several languages. It showed how quickly we can make an idea testable. It was not a production launch. Your schedule depends on scope, access to your systems and data, and the decisions your team needs to make. Four days is evidence from one sprint, not a promise for every project.
Learn more about our AI-first approach, why we build with Flutter, or get in touch.