ATAILA Newsroom · Budapest · 2026-08-19
Most AI projects never reach production. Ours start there.
The Hungarian business news portal Portfolio.hu has published KPMG’s Global Tech research together with an interview with Tamás Kórász, head of KPMG’s consulting division. The finding is a contradiction worth sitting with: AI investment can pay back twice over — and most projects never get the chance, because they never reach production. We see that gap every week. It is the exact problem this platform was built to remove.
The article we are responding to
Portfolio.hu · 2026-08-19
What the article reports
KPMG surveyed 2,500 managers across 27 countries; the figures the article quotes are blunt.
Kórász’s explanation of why is the useful part. Projects fail when companies use AI to paper over gaps in their data or operations — in his words, approaches that try to cover data-system or operational shortcomings with an AI solution “usually end in failure”. They fail when the data sits scattered across outdated systems, when there is no governance, when staff are not prepared, and when the work is driven by fear of missing out rather than a real business case. The projects that succeed pick a problem where pattern recognition genuinely helps, fix the data first, leave room to experiment, and redesign the workflow around the result.
He also names the security angle: AI use pushes company data into cloud environments “scattered and hard to track”, and the next risk is AI agents — what permissions they get, and who answers for them.
Figures as reported in the article (KPMG Global Tech research, Gartner, KSH, GKI); see the original for full context.
We agree with almost all of it
This is not a rebuttal. The diagnosis matches what we see every week: a prototype that impressed everyone in the demo, built on a slice of data in someone’s notebook, with no path to the systems, the security review or the operations team that production demands. The 60% that never ship are not failures of AI. They are failures of delivery.
Where we part ways with the industry’s usual conclusion is what to do about it. The standard answer — “more maturity, more governance, more change management” — is all true, and all of it is something a 20-to-200-person company cannot staff. We built a platform so it does not have to.
What ATAILA does about each of these
The article's finding
Most projects stall between the prototype and production.
Our answer
Production is not the last step of an ATAILA project — it is the first thing we provision. Every project starts with real DEV, UAT and PROD environments, and a governed Release Manager carries each change through them. There is no separate “go-live project” that never gets funded.
How the Release Manager works →The article's finding
AI is used to cover up gaps in data and operations.
Our answer
We start with one workflow and the data it actually touches — not “AI for the company”. That is exactly why we turn away work where the data is not there yet. A narrow, real problem reaches production; a broad one does not.
Who ATAILA is for — and who it is not →The article's finding
Data lands in the cloud “scattered and hard to track”.
Our answer
Our compute is private and EU-resident end to end. Your documents, embeddings, logs and the application itself live on infrastructure you control — not spread across a public AI provider and three SaaS tools. That is also what makes governance an answerable question.
The sovereign EU cloud underneath →The article's finding
Nobody owns what happens after the pilot.
Our answer
We operate what we ship — monitoring, backups, security updates, model lifecycle — under a written SLA. The project does not end at handover, because there is no handover.
What “operated” includes →Our commitment, in one sentence
A project that runs on the ATAILA platform goes to production. Not as a hope, as architecture: the production environment, the governed release path into it, and the team that operates it afterwards are all part of the platform from day one — so “reaching production” stops being a separate project that can quietly run out of budget.
We can say this because the platform proves it on our own products: the open*.hu family — OpenMath and OpenChef today, more coming — was built by ATAILA Coder on private AI and runs in production on the platform; the newest one went live in hours by reusing modules the platform already had. The full picture is on the references page.
One thing we will not claim
We cannot make a bad idea pay back twice. If a workflow is not worth automating, or the data genuinely is not there, we say so — before you spend anything. That is the whole point of the questions we ask up front. What we can do is remove the reason most good ideas die: the long, unfunded, unowned gap between a working prototype and a running system.
If you have that one confidential workflow you keep meaning to automate — the one you would never paste into a public AI — tell us about it. We will tell you honestly whether it will reach production. Then we will take it there.
Source: Vállalati AI-beruházások: kétszeres megtérülés is elérhető, de a projektek többsége még éles üzemig sem jut el — Portfolio.hu, 2026-08-19
← Back to the Newsroom Press inquiries: contact us