Choosing a partner for an AI implementation follows different rules than choosing a team for a regular application. A polished demo in a sales meeting tells you nothing about whether the system will survive its first month running on real data. Below, we look at what actually matters instead of the slide deck — and which questions are worth asking, because they determine whether a project ends up as an “interesting pilot” or actually becomes part of a company’s daily work.

A demo is not production

Any team can stitch together a demo in a few days where the model answers questions from a curated set of examples. The real test starts the next day, when the system meets real, messy data: low-quality scans, invoice formats nobody anticipated, customers writing in ways no prepared scenario covered. We wrote more about this in our piece on what it actually means for AI to run “in production” — the gap between a demo and a working system is fundamental, and most offers on the market don’t address it at all.

So the first question worth asking when choosing a partner isn’t “can you build a demo” — it’s “what happens a month after launch, when data shows up that nobody planned for.”

Questions worth asking a potential partner

Instead of asking generically “do you know AI,” a few specific questions quickly reveal whether a partner is thinking about production or just about the pitch.

Do you have your own product where AI runs in production — not just client projects? This question matters more than it seems. A company that maintains its own AI-based product knows firsthand the cost of errors, support, and keeping a model running over time. A company that only ever does work-for-hire projects often stops caring the day the project is signed off — and that’s exactly when the real problems start.

What happens when the model gets it wrong? A good partner has a ready answer: what mechanisms catch bad outputs, where automation ends and a human takes over, how logging and the ability to reverse a decision work. If the answer is a shrug, or reassurance that “the model rarely makes mistakes,” that’s a warning sign.

Where exactly does the AI’s suggestion end and the system’s decision begin? In a well-designed implementation, these two roles are separated and named explicitly. If a partner can’t clearly point to that boundary, they probably haven’t designed one — which means the model is making decisions that deterministic code should be making instead.

How is the data that reaches the model secured? This question matters especially when financial data, personal data, or client documents are involved. It’s worth asking directly where data is physically processed, who has access to it, and what happens to it after a query is processed.

Who maintains the system after launch, and what does that maintenance actually look like? Unlike a classic website or a simple integration, an AI implementation needs ongoing attention — models change, data changes, and the system’s behavior changes along with them. A partner with no maintenance plan is, in practice, planning a one-off project, not a long-term partnership.

Red flags

A few signals that should raise a warning flag in conversations with a potential partner:

  • “AI will do it all by itself” — with no mention of oversight, escalation to a human, or control mechanisms. A responsible implementation always has a defined boundary for automation.
  • No concrete answer about error handling — vague reassurance instead of a description of the actual mechanism.
  • Pricing that only covers “building” — with no mention of maintenance, monitoring, or updating the model over time.
  • No firsthand experience running AI in production — a partner who has never maintained a product built on language models is learning at your project’s expense.
  • All talk about technology, none about process — no mention of process audits, pilots, or success criteria.

Why this matters to us

We hold ourselves to the same criteria we describe above, because we implement AI not only for clients but in our own products, which run in production every day. Qkwit, our AI-powered accounting product for sole proprietors, uses Claude to read accounting documents, while the actual tax and social security calculations are handled by a deterministic engine — not the model. Brokik, our rental management platform running in 25 countries, uses AI to adapt documents and pricing, always grounded in the local rental law of each market. These aren’t case studies written for a pitch — they’re systems we maintain and are accountable for ourselves, every day.

That’s why, when we advise clients on how to design an AI implementation, we’re not working from theory — we’re pointing to solutions we’ve tested at our own risk first.

What to expect from a good partner

A well-run AI implementation tends to follow a similar shape regardless of industry: it starts with an audit of a specific process (not a general conversation about “AI for the business”), moves through a pilot on real, if limited, data, and only after the system proves stable does it move to a full rollout — with a clear maintenance plan attached. A partner who proposes jumping straight to full automation of an entire process, skipping the pilot stage, usually hasn’t planned for what happens when something goes wrong.

FAQ

Can a smaller team implement AI well, or do we need a large firm? Company size isn’t the main criterion here. What matters more is whether the team has real experience maintaining AI systems in production — regardless of whether that’s a large organization or a small, specialized team.

How long should the process audit stage take before implementation? It depends on how complex the process is, but a good audit always ends with a concrete conclusion: what can be automated safely, and what needs human oversight — not a vague “yes, AI can do this.”

Should we expect a partner to guarantee the model will never make a mistake? No — that kind of guarantee is a warning sign, not a selling point. A responsible partner will tell you plainly that models make mistakes, and show you how the system is designed to handle that.

Does a partner need its own AI-based product to implement AI well for us? It’s not a strict requirement, but it’s a strong positive signal — it shows the partner understands the costs and pitfalls of maintaining such systems from firsthand experience, not just from projects done for other people.

Where should we start the conversation with a potential partner? With a specific process you want to improve, not with the general phrase “AI implementation.” The more concrete the starting point, the easier it is to tell whether a partner actually understands the problem or is just selling technology.

If you’re wondering what that kind of audit looks like in practice, and where to start a conversation about implementing AI in your company, get in touch — we’re happy to walk you through it from the perspective of a team that maintains its own AI products in production.