Most AI consulting engagements end the same way: months of discovery workshops, a polished 60-page strategy document, a roadmap deck — and not a single line of software you can actually use. You paid for a plan. What you needed was proof. At Crux Digits B.V., a Utrecht-based AI consultancy and software studio, we flipped that model. By the second call, you are not reading another report; you are clicking a working AI MVP built on your own data. This article explains how that is possible, what it honestly is (and is not), and why a working AI proof of concept changes everything about risk, buy-in and time to value.
The promise is simple to state and harder to deliver: a working AI MVP instead of a strategy document. Not a mock-up, not a Figma flow, not a slide that says "imagine if." Real software, running against your real data, that you can put in front of a real user. That is the entire point.
The problem with the 60-page document
The traditional discovery-and-strategy model made sense in an era when software was slow and expensive to build, so you de-risked with planning first. AI broke that logic. The expensive, uncertain part of an AI project is no longer drawing the diagram — it is finding out whether the model is actually accurate and useful on your data. A 60-page document cannot tell you that. It can describe an architecture, list use cases, and estimate a business case, but it quietly assumes the hard part will work. Frequently, it does not work the way the slide promised, and you discover that months and a large budget later.
There is a second cost that rarely gets named: momentum. A long discovery phase drains energy from a project before anything tangible exists to rally around. Stakeholders nod at the deck, the document goes in a drawer, and the initiative stalls in the gap between "interesting idea" and "thing that works." We wrote about closing that gap in AI-native software delivery — the shift from documenting software to shipping it.
What Crux does instead: a working AI MVP, fast
Our answer is to compress the distance between conversation and working software. The first call is genuinely a conversation: we listen to the problem, identify the single riskiest assumption, and agree on the smallest thing that would prove or disprove it. Between calls, we build. By the second call you are looking at software you can click — a focused AI proof of concept running on a slice of your real data, not a hypothetical.
Concretely, that might be a retrieval-augmented assistant answering questions from your actual document set, a classifier scoring your real records, or a computer-vision model running on your own images. The interface may be rough and the edges unfinished — but the core works, and you can see it work. A live demo on your data settles arguments that a deck only opens. This is what we mean by an AI consultancy that ships working software, and it runs through every service we offer, from AI implementation to application development.
How that's even possible: tight scope and reusable building blocks
The fair question is how fast can you build an AI proof of concept without cutting corners — and the answer is that speed like this is not magic, and it is not a heroic all-nighter. It comes from three disciplines. First, ruthless scope: we deliberately build the one thing that proves the core, and consciously leave everything else out. The temptation in AI projects is to boil the ocean; the skill is to pick the single slice that carries the most risk and answer it first.
Second, reusable building blocks. We do not start every engagement from a blank repository. Data connectors, retrieval pipelines, evaluation harnesses, authentication, deployment scaffolding — the plumbing that makes up most of an AI application — are components we have built before and assemble quickly. That is exactly the production AI stack doing its job: the commoditised layers are ready, so our time goes into the part that is specific to you.
Third, the audit-to-proof-of-concept path. A short, structured AI audit and strategy step maps your data, constraints and the highest-value use case, so that when we build, we build the right small thing rather than guessing. The audit is not a 60-page deliverable for its own sake; it exists to aim the proof of concept. This design-first discipline — clarity before code — is something we unpack in design-first AI.
What "by the second call" honestly means
Here is the candid part, because credibility matters more than hype. "By the second call" describes how we work, not a contractual guarantee, and the thing you see is a focused, working prototype — not a finished, fully-hardened production system. It proves the riskiest assumption on real data. It does not yet have enterprise authentication wired through every edge case, exhaustive monitoring, load testing, full security review, or polished UX across every screen. That hardening is real work, and it is exactly what the production phase is for.
The timeline depends on honest inputs, too. If your data is accessible and the scope is genuinely tight, a working AI MVP in a couple of weeks is realistic. If data access requires legal sign-off or the data itself needs significant cleaning, that adds time — and we will tell you so up front rather than promise a date we cannot meet. An AI proof of concept in weeks, not months, is the norm we aim for; it is a working approach, not a number we put in a contract. Honesty about scope is what keeps the second-call MVP credible instead of a sales trick.
The impact: why a working MVP beats a document
This is the heart of it. A working AI MVP changes the economics and the politics of an AI project at once.
Faster validation. You learn whether the idea actually works in weeks, on evidence, instead of betting a year and a large budget on a business case that assumed it would.
Lower risk. You see it work before you commit serious money. The biggest unknown in any AI initiative — does the model perform well enough on our data to be useful? — is answered early, cheaply, and concretely. If the answer is no, you have spent the cost of a proof of concept, not a production build.
Real feedback, early. The moment a stakeholder or end user can click the thing, you get feedback that no document elicits. People react to working software honestly and specifically; they react to slides politely and vaguely. That early signal redirects the build while redirection is still cheap.
Buy-in and momentum. A live demo beats a deck in every meeting that decides budget. Showing a working proof of concept on the company's own data turns sceptics into sponsors faster than any slide, and it creates the momentum that carries a project from pilot to production instead of letting it die in committee.
A faster path to ROI. Every week you are not stuck in discovery is a week closer to measurable value. Compressing the front of the project compresses the whole timeline to return.
Why this matters more for AI than for ordinary software
You could argue that prototyping fast is good advice for any software project, and you would be right. But it matters far more for AI, for one reason: AI value is empirical. With conventional software, if the spec is met, it works — behaviour is deterministic and largely knowable in advance. With AI, accuracy and usefulness can only be proven on your real data. No document, however thorough, can tell you in advance whether a model will be 70% or 95% accurate on your specific records, or whether that accuracy is good enough for the decision you want to automate. Only a working proof of concept can tell you what a document never could. That is why, for AI, "show me" beats "tell me" by a wide margin — and why we refuse to sell certainty we have not yet earned on your data.
How to start
The path is deliberately staged so you never over-commit before you have proof. It begins with an AI audit and strategy (typically around €2,500) to map your data and pick the highest-value use case. Then a focused proof of concept (around €20,000) ships a working AI MVP on your own data — the clickable, evidence-on-your-data step this whole article is about. Only once that proves the core do we move to production (from €50,000), where we harden it into a monitored, governed system. Transparent scope lives on our pricing page, and illustrative work in our case studies. When you are ready, book a free consultation and let's find the one thing worth proving first. By the second call, you'll be clicking it.
Frequently asked questions
How fast can you build an AI proof of concept?
When your data is accessible and the scope is genuinely tight, a working AI MVP in a couple of weeks is realistic — an AI proof of concept in weeks, not months. The timeline depends on honest inputs: if data access needs legal sign-off or the data needs significant cleaning, that adds time, and we tell you up front. "By the second call" describes how we work, not a contractual guarantee.
Is the second-call MVP a finished production system?
No, and we are candid about that. It is a focused, working prototype that proves the riskiest assumption on your real data — software you can click, not slides. It does not yet have full enterprise authentication, exhaustive monitoring, load testing, a complete security review, or polished UX across every screen. That hardening is exactly what the production phase delivers, once the proof of concept has validated the core.
Why a working AI MVP instead of a strategy document?
Because AI value is empirical: accuracy and usefulness can only be proven on your real data, and no document can tell you in advance whether a model will be useful enough for your decision. A working proof of concept answers the riskiest question early, cheaply and concretely, lowers risk before you commit budget, and a live demo wins stakeholder buy-in far faster than a deck.
How can you build a working MVP so quickly?
Three disciplines. Ruthless scope — we build only the one slice that proves the core and leave the rest out. Reusable building blocks — data connectors, retrieval pipelines, evaluation harnesses, authentication and deployment scaffolding are components we have built before and assemble fast. And the audit-to-proof-of-concept path, where a short audit aims the build so we make the right small thing rather than guessing.
What does it cost to get a working AI MVP?
The path is staged so you never over-commit before you have proof. It starts with an AI audit and strategy (typically around €2,500) to map your data and pick the highest-value use case, then a focused proof of concept (around €20,000) that ships a working AI MVP on your own data, and production from €50,000 once the core is validated. Transparent scope is on our pricing page.