How to choose an AI development partner: 9 questions

Why isn’t a general vendor checklist enough for AI?

Because an AI vendor ends up holding things a web agency never touches: the prompts, the evaluation set, the log of model calls and access to your data. If those stay on their side, the system is rented, not yours. The nine questions below poke at exactly those points.

In the same forecast Gartner estimated that out of the thousands of companies calling themselves agentic AI vendors, roughly 130 actually are. That number does not say avoid AI. It says selection is the riskiest phase of the project.

The general criteria still apply: references, team size, contract framework, communication rhythm. We wrote those up in twelve points in our piece on choosing a web development agency. This article covers only what shows up with AI and nowhere else.

1. How does the price break down, line by line?

A quote becomes comparable when discovery, data preparation, integration, building the evaluation set, go-live and monthly operation each sit on their own line. A single total is a negotiating position, not a quote. Ask for the running cost separately, because it is not inside the development price.

The most detailed public phase breakdown comes from the development agency Musketeers Tech, published on 2 April 2026. Worth knowing: the international “AI agent cost guide” genre is almost entirely content marketing written by development agencies, not independent market research. Good as data points, not as statistics.

Most quotes stay silent on the running cost, even though it is easy to calculate. A RAG chatbot handling 500 conversations a month burns between USD 1.50 (Gemini 2.5 Flash-Lite) and USD 31.50 (Claude Sonnet 5) in raw model cost, based on the providers’ official price lists as of 14 August 2026. That is roughly EUR 1.30 to EUR 27. Behind a managed chatbot subscription priced around HUF 24,900 a month, about EUR 69 at the MNB mid-rate of 363.28 HUF/EUR on 14 August 2026, most of the fee is service rather than tokens.

Agentic operation is where the number explodes, and the driver is the number of loops. The same 1,000 documents cost USD 8.70 with a single extraction call and USD 232 in an eight-step agent loop on the same model family. Twenty-seven times the price. So the one pricing question that genuinely matters is: how many model calls make up one transaction in your design? Detailed price bands sit in our AI development cost guide.

If you are buying from another EU country, the hourly rate is only half the picture. German agency guides put senior rates at EUR 120 to 160 an hour, and the Freelancer-Kompass 2026 panel published on 2 July 2026 puts the DACH IT freelancer median at EUR 95, against HUF 16,000 to 24,000 (roughly EUR 44 to 66) on the Hungarian market. A cheaper rate still loses if the architecture spends ten model calls where one would do. More on that trade-off on our nearshore development page.

It is a good sign when the vendor brings a running-cost estimate with the assumptions attached, and states the maintenance figure. It is a bad sign when maintenance is “something we can discuss later”. On the Hungarian market it typically lands at 15 to 25 percent of the implementation fee per year.

2. Who owns the source code and the IP?

Whoever the contract names. With an AI system, though, source code on its own is thin. Six things need to be listed explicitly, otherwise the code becomes yours while the working system does not. Most quotes skip this part, and that is not an accident.

• The repository with full commit history, not a zip file.
• The version history of the prompts, because the behaviour of the system lives there.
• The evaluation set together with its expected answers.
• The vector index and the indexed documents in processed form.
• Fine-tuned model weights, if any fine-tuning happened.
• The infrastructure definition and the environment configuration.

Plenty of vendors advertise full IP transfer without describing what it covers. Ask specifically, and ask for those six items to appear by name in the contract. In all three of our AI packages the complete codebase belongs to the client, with no monthly licence fee.

3. Where does the data live, and who can reach it?

Three things to establish: which provider runs the model call in which region, whether that provider trains on submitted data, and how long the logs are retained. The third gets asked least and causes the most trouble later, because compliance work and audits both depend on it.

For a German or Austrian buyer this is the top objection in the market. In Bitkom’s survey of 604 companies with 20 or more employees (11 March 2026), data protection requirements were the number one external barrier to AI adoption at 77 percent. The Bitkom Cloud Report of 17 June 2026 found 98 percent say the origin of the provider matters, and 68 percent would prefer an EU provider against 45 percent who use one today.

A supplier inside the EU means no third-country transfer, so no standard contractual clauses and no transfer impact assessment. What you still need is the Article 28 GDPR data processing agreement, the Auftragsverarbeitungsvertrag in German practice, with all ten mandatory elements. A missing or defective one carries up to EUR 10 million or 2 percent of global annual turnover.

Training rules differ by provider and by tier. Google’s own pricing documentation states that on the free Gemini API tier submitted content is used for product improvement, so client data cannot go there. Anthropic’s documentation states that uploaded images are not used for model training, and zero data retention is available for document processing. Those details decide whether the system survives a data protection impact assessment.

A bad answer is “EU-based, GDPR compliant”. That carries no information. A good sign is when the vendor volunteers the sub-processor list and tells you which data crosses which border.

4. How do they measure whether the system works?

With an evaluation set. That means 100 to 300 real questions drawn from your own data, each with a recorded expected answer, re-run after every prompt change and every model change. Without it, “it works well” is an impression collected during a demo. A vendor who cannot answer this has no measurement.

Language matters more here than most buyers expect. Published retrieval benchmarks are overwhelmingly English. As of August 2026 there is no public, maintained Hungarian retrieval benchmark, and the most recent independent test measured substantial overlap between the similarity distributions of related and unrelated Hungarian sentence pairs. If your knowledge base is not in English, international accuracy figures do not transfer to it.

One concrete metric shows whether there is engineering behind the system. In Anthropic’s own measurement, hybrid search plus reranking cut the retrieval error rate from 5.7 percent to 1.9 percent. If a vendor can quote a number of that kind about their own system, they measured. If the answer is “very accurate”, they did not.

Ask about continuous monitoring too. Langfuse is free and open source when self-hosted, so the real monitoring cost on an SME project is running the container, not USD 199 a month. We compared the two widespread tools in our LangFuse and LangSmith review.

5. What happens when the model changes or is retired?

This is the question of month six, and quotes almost never address it. The model shifts underneath you: a new version arrives, pricing changes, the old version gets deprecated. Settle who re-runs the evaluation, who retunes the prompts, and whose invoice carries the difference.

Two concrete 2026 examples that show this in money. The new tokenizer introduced from Claude 4.7 produces roughly 30 percent more tokens for the same text, so operating cost rises even at an apparently unchanged unit price. And the promotional pricing on Gemini 3.6 and 3.7 Flash runs until 31 December 2026, after which input goes from USD 0.75 to USD 1.50 and output from USD 3.75 to USD 7.50 per million tokens. Anyone modelling 2027 operating cost should use those figures.

Ask about cost control as well. A monthly budget ceiling is not protection, because it triggers after the runaway run has finished. What protects you: a per-run token cap, a step limit, a per-step timeout, and model routing so routine steps run on a cheap model. According to TechCrunch on 5 June 2026, Uber had consumed its entire 2026 AI coding budget by April.

6. How does it connect to your existing ERP and CRM?

This is where you find out fastest whether you are talking to an engineering team. The weak answer lists system names. The good answer tells you which API, which authentication, what happens when the other system does not respond, and how duplicate postings are avoided on a retry.

Five questions that need concrete answers. Is there an API at all, or are you left with database reads and file exchange. Who owns idempotency, meaning what guarantees that a message sent twice does not post twice. What rate limit applies on the other side. Does the agent get write access to a production system, and under what approval. Finally, what is the plan for master data quality, because on dirty ERP data the best retrieval still returns the wrong answer.

MCP is worth asking about too. The protocol moved to the Agentic AI Foundation under the Linux Foundation in December 2025, and the specification released on 28 July 2026 added enterprise managed authorization. In Anthropic’s own measurement, a task copying from Drive into Salesforce fell from 150,000 tokens to 2,000 once intermediate data stopped passing through the model. The practical architectures are on our system integration page. Add whichever national e-invoicing regime you file under, such as NAV Online Számla 3.0 in Hungary, and ask about it by name.

7. Who is responsible for EU AI Act and GDPR compliance?

Usually you, not the developer. Article 3 of the AI Act defines the provider as whoever places the system on the market or puts it into service under their own name or trademark. If the chatbot runs on your website under your brand, the provider obligations are yours and the developer is a supplier. Fix the role allocation in the contract.

What is live now: the Article 50 transparency obligations have applied since 2 August 2026, so a customer has to be told they are talking to an AI. For systems already on the market before that date, the grace period for machine-readable marking ends on 2 December 2026. The deadline for high-risk Annex III systems, on the other hand, moved to 2 December 2027 under Regulation (EU) 2026/1744, the Digital Omnibus, which entered into force on 27 July 2026. A lot of published guidance still shows the old date, so check the date on anything your vendor cites.

The fine under the second tier of Article 99 is EUR 15 million or 3 percent of worldwide annual turnover, whichever is higher; for SMEs and startups the cap is the lower of the two. One trap is worth knowing: under Article 25, if you change the intended purpose of a system so that it becomes high-risk, you become the provider even though you did not build it. On the deployer side, Article 26 requires logs to be kept for at least six months, from 2 December 2027. We break down classification and documentation on our EU AI Act compliance page.

One more thread for German buyers. The German NIS2 implementation act was published on 5 December 2025 and covers around 30,000 companies. If you are in scope, supply chain security requirements land on your AI vendor too, so ask what evidence they can produce for your NIS2 documentation.

8. What is the exit plan if you want to switch?

Ask for a written exit clause before you sign. It should fix three things: exactly what gets handed over, in what format, and within how many working days. If the question is awkward at contract signature, it will be far more awkward at termination, when you no longer have a negotiating position.

The exit package is the same six items as the IP list, plus the runtime logs, which you need anyway once the deployer retention rule applies. One technical requirement matters a lot here: the model call should sit behind an abstraction layer, so that changing provider stays a configuration question instead of a rewrite.

It is a bad sign when prompts live in a web admin interface with no version control, because then only the current state can be handed over, not the history of the system’s behaviour. It is a good sign when the vendor offers a handover rehearsal: an outside developer gets two days and stands the system up from the documentation.

9. Will they show you a project that failed?

This is the cheapest filter of the nine and the most revealing. Anyone who has been building AI systems for two years has a pilot that got stopped. If they have no such story, they have either done few projects or they are not being straight with you. A good answer is specific: what the use case was, where it broke, what they would do differently today.

In practice this field is empty. We went through twelve AI vendor sites in one European market in August 2026, and all twelve communicate success only. None of them writes about when not to build AI. Of the seven Hungarian and DACH nearshore providers we reviewed the same week, none published anything on EU AI Act or NIS2 compliance either.

Two common failure patterns worth probing. The first is agent washing: the task would be a deterministic workflow, but it gets sold as an agent. Anthropic’s own guidance says a coded workflow beats an autonomous agent for most tasks. The second is an AI layer bolted onto an old process. In McKinsey’s 2026 survey, EBIT impact correlated most strongly with redesigning the process from the ground up, not with piecemeal rollout.

When is AppForge the wrong choice?

Four situations where you are better off elsewhere, and it is cheaper to clear them up on the first call than in month three.

• When an off-the-shelf chatbot is enough. Self-service widgets start at around HUF 1,400 a month, roughly EUR 4, and a managed SaaS chatbot runs around HUF 24,900. For a static FAQ and a few scripts, do not pay for custom development.
• When the requirement is a 24/7 multi-shift operations SLA. We are a team of fewer than ten, and we cannot commit to that honestly.
• When you need a parallel programme with dozens of developers. Large software houses are the right address for that.
• When the process itself is broken, or the data is not digitised. Then AI just produces the same wrong result faster. Fix the process first, add the model second.

What is worth knowing about us: we have two references we can name. We built the AI-powered search for the Hungarian Popular Culture Value Repository for kultura.hu, and we spoke at AI Summit Budapest 2025. The rest of our work sits under NDA, and for those we show anonymised architecture instead. Ask every vendor to draw that same line, because an “80+ projects” figure cannot be verified. Our public work is on the AI portfolio page.

Summary and frequently asked questions

Run these nine questions across three quotes and the prices become comparable, and it usually turns out the three vendors were not offering the same thing. If you want to try them on us, request a quote and we will answer all nine in writing.