Cluster: from prototype to production

From PoC to production: where AI projects break, and how to avoid it.

Most AI prototypes stay a demo. The jump to production requires architecture, data, and governance decisions taken at the beginning, not at the end.

The demo works, production does not.

In 2026 the first AI prototype costs almost nothing to build. Lovable, Cursor, v0, Google AI Studio, Bolt have collapsed the barrier to entry. The problem is no longer getting started, it is reaching a system that survives real data, real users, and real business requirements. That is where the budget goes: industry sources (McKinsey and Gartner, 2022-2024) estimate that 70-80% of AI initiatives never move past the pilot. The exact number varies. The direction does not.

We call it the prototype cliff: the drop between the perfect demo and the first user who does something unexpected.

Where AI prototypes break

  1. 1. On the data. The PoC runs on a clean, often synthetic dataset. The first contact with real production data, incomplete, with drifting schemas, with duplicates, degrades accuracy and response times.
  2. 2. On the architecture. The monolithic prompt that does everything in one call works for one user and does not scale to a hundred. Without a separation between model, orchestration and interface, every change touches everything.
  3. 3. On security. Keys in plain text, sensitive data passed to the model, no trace of who asked what. The security team says no and the project stops.
  4. 4. On measurement. The prototype has no model quality metrics (accuracy, latency, cost per call). Without numbers, every production conversation is opinion.
  5. 5. On ownership. Nobody inside the company has the mandate to maintain it after release. A system with no internal owner dies at the first maintenance request.

Vibe coding moved the problem, it did not solve it

Generating a prototype in an afternoon is now normal. That compressed the creative phase and grew the next one: figuring out whether that prototype can hold the weight of an organisation. The speed advantage flips to a disadvantage the moment the prototype becomes, without review, the foundation of a production system.

Industry blogs in 2026 have reported meaningful shares of AI-generated code carrying security vulnerabilities. That is not a reason to stop using the tools. It is a reason not to treat the output as production-ready.

With agents the gap is wider

An agent does not produce an answer, it produces a sequence of actions. Every step can fail on its own and an intermediate error can generate a final output that looks correct. Whoever judges only the output does not see it.

Microsoft Research found that 61% of failures in multi-agent systems arise at the handoff points between agents, not inside a single agent. It is the point almost nobody tests at the prototype stage, and the one that gives way first in operation.

On top of this comes a cost item that does not show up in the prototype. An agent that runs on ten requests a day can become unsustainable at ten thousand. Cost per run must be measured during the sprint, not discovered after release.

How we stress-test an agent

How to design a prototype so it survives production

You do not need to build it differently. You need to start it differently. Five decisions at hour zero change the trajectory.

  • API-first architecture. Separate model, orchestration and interface from day one. Every layer swappable.
  • Real data, immediately. Even a small sample of real data tells you more than a complete synthetic one.
  • Metrics from sprint one. Accuracy, latency, cost per call. If you do not measure them, you are not managing them.
  • Governance as a checklist. Security, privacy, audit trail as sprint items, not a separate downstream project.
  • Named internal owner. Whoever takes it after go-live joins from the first standup.
When it makes sense to start
  • You have a use case with a measurable outcome.
  • You can access a real data sample within two weeks.
  • An internal owner will take the project after the pilot.
  • You want a go / no-go answer inside a month, not in six.
When it is better to stop
  • The business value is unclear, or not yours to decide.
  • The data you need does not exist or is not accessible.
  • An off-the-shelf product already solves the problem.
  • No one in the organisation has the mandate to own it.

We do not prototype at random.

protot.ai is the rapid-validation stage of the ZeroFive chain. AI Rating and AI Strategy measure maturity and set the route, protot.ai validates the use case with a working prototype, AI Agents and AI Shift take to production and govern what has passed validation.

See how we workAgents and agentic processes

Frequently asked.

What founders and product teams ask before starting a validation sprint.

Why do so many AI PoCs never reach production?
Almost always three reasons together: the use case was never validated with business criteria, the architecture was chosen to impress in a demo rather than to hold real data, and governance was deferred to later. When later arrives, rewriting costs more than starting over.
What is the prototype cliff?
It is the drop between the demo that works in a meeting and the system that has to hold users, data and integrations. Projects that optimised for the speed of the first output, not the trajectory to production, fall off it.
Does vibe coding help or make the problem worse?
It helps the speed of the first prototype. It hurts the trajectory when generated code becomes the basis of a system with no human review, no tests, no security controls. It stops being an advantage the moment responsibility moves from the individual to the team.
What does an AI prototype need to reach production?
Access to real data, an API-first architecture that separates model, orchestration and interface, quality metrics on the model, a security and compliance plan, and a named internal owner who will keep the system after release. If any of these is missing, the prototype stays a prototype.
How much time does governance add to a prototype?
Very little, when you think about it at design time. A lot, when you glue it onto a prototype that is already written. In our sprints, governance is a checklist that runs in parallel with the build, not a phase after it.

Before you write the code, validate the use case.

A 30-minute discovery call to see whether your AI idea is worth prototyping.

Book the discovery call