← All insights
Delivery · Production readiness

From AI prototype to production — the journey EU teams skip

Most AI projects do not fail on the model. They stall between a convincing demo and a system that survives real users, real cost and a real auditor. This is the path LMXAI runs.

LMXAI · Leiden, NL · Published · Revised

Nederlandse versie  ·  Türkçe versiyon

Gap
Demo → production
Usual stall
No eval, no owner
EU extra
AI Act evidence
Partner
LMXAI fractional CTO

Abstract

Most AI projects do not fail on the model. They stall between a convincing demo and a system that survives real users, real cost and a real auditor. This article names the stages of that journey, the points where teams stall (evaluation, observability, security, cost, EU AI Act evidence) and the production checklist LMXAI uses, and sets them against research on why machine-learning systems are hard to deploy.

Keywords production AI · evaluation · observability · MLOps · EU AI Act · checklist

Contents

  1. Background and related work
  2. The journey, named
  3. Where teams stall
  4. The production checklist LMXAI uses
  5. How long it takes
01

Background and related work

The gap this article describes is well documented. A survey of published deployment case studies found practical problems at every stage of the ML workflow, from data management to monitoring and integration[1]. A literature and interview study of MLOps concludes that operationalizing ML products is highly challenging and that many ML endeavours fail to deliver on expectations[2]. Part of the reason is structural: ML systems accumulate hidden technical debt beyond the model code[3].

Interviews with 18 ML engineers identify velocity, validation and versioning as the variables that govern a successful production deployment[4], and a production-readiness rubric turns such findings into explicit tests and monitoring[5]. For retrieval systems specifically, an experience report concludes that validation is only feasible during operation[6], which is why the checklist below insists on traces and a frozen evaluation set before launch.

02

The journey, named

The AI prototype-to-production journey is the work that turns a notebook, a Copilot trial or a hackathon agent into a service with an SLO, an owner, a cost envelope and an evidence pack. LMXAI exists for that interval. The homepage claim is literal: one partner across software, architecture, data and commercialisation, acting as fractional AI CTO.

03

Where teams stall

  1. Unowned demo. It lives on one laptop. Nobody can redeploy it.
  2. No gold questions. Quality is “it felt good in the steering committee.”
  3. Prompt-only security. API keys in a .env, documents in a shared drive, no audit trail.
  4. Surprise invoice. Context and agent loops were never modelled.
  5. Compliance as a slide. “We will be EU AI Act ready” with no inventory or logs.
  6. Handoff vacuum. The vendor leaves; the repo is a zip file.
04

The production checklist LMXAI uses

Stage Exit criterion
Frame Intended purpose, users, data classes, success metric written down
Thin slice One workflow in a real environment, not a slide prototype
Retrieve / tool Index or tools you own; eval set of ≥30 gold items
Serve vLLM or contracted API behind a gateway; traces on
Harden Auth, egress policy, PII handling, human override
Prove SLO + cost sheet + AI Act gap list signed by an owner
Transfer Your org, your runbooks, your on-call — we stay or we leave cleanly

EU-specific: stage “Prove” includes the compliance deliverable — classification, logging, technical-file skeleton — so production and regulation do not become two projects that never meet.

05

How long it takes

A honest thin slice is weeks, not quarters. A workspace or clinical agent that a 500-person org can use is months, because retrieval quality, IAM and works-council constraints dominate the calendar — not model choice. Anyone who quotes a production date from a model card is selling the skipped journey.

Related reading: production RAG and agents, sovereign vLLM, workspace vs Copilot.

References

  1. Paleyes, A., Urma, R.-G., & Lawrence, N. D. (2022). Challenges in deploying machine learning: A survey of case studies. ACM Computing Surveys, 55(6), 1–29. doi:10.1145/3533378 arxiv.org/abs/2011.09926
  2. Kreuzberger, D., Kühl, N., & Hirschl, S. (2023). Machine learning operations (MLOps): Overview, definition, and architecture. IEEE Access, 11, 31866–31879. doi:10.1109/ACCESS.2023.3262138 arxiv.org/abs/2205.02302
  3. Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., et al. (2015). Hidden technical debt in machine learning systems. Advances in Neural Information Processing Systems 28 (NeurIPS 2015). papers.nips.cc/paper_files/paper/2015/hash/86df7dcfd896fcaf2674f757a2463eba-Abstract.html
  4. Shankar, S., Garcia, R., Hellerstein, J. M., & Parameswaran, A. G. (2024). "We have no idea how models will behave in production until production": How engineers operationalize machine learning. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1), 1–34. doi:10.1145/3653697 arxiv.org/abs/2209.09125
  5. Breck, E., Cai, S., Nielsen, E., Salib, M., & Sculley, D. (2017). The ML test score: A rubric for ML production readiness and technical debt reduction. 2017 IEEE International Conference on Big Data, 1123–1132. doi:10.1109/BigData.2017.8258038
  6. Barnett, S., Kurniawan, S., Thudumu, S., Brannelly, Z., & Abdelrazek, M. (2024). Seven failure points when engineering a retrieval augmented generation system. Proceedings of the IEEE/ACM 3rd International Conference on AI Engineering – Software Engineering for AI (CAIN 2024), 194–199. doi:10.1145/3644815.3644945 arxiv.org/abs/2401.05856

Each reference was checked against its publisher, arXiv or official record.

How to cite

LMXAI. (2026). From AI prototype to production — the journey EU teams skip. LMXAI Insights. https://lmxai.com/insights/prototype-to-production.html