Whitepaper · July 2026
The Last Mile of
AI-Native Software
AI has collapsed the cost of building software. It has not collapsed the cost of running it. This paper supplies the chapter most AI strategies are missing: who takes AI-built software live, who runs it, and who answers the auditor.
13 pages · free · no email required. Cite freely with attribution.
Operations running across Europe. Sweep the lights.
of companies now abandon most of their AI initiatives, up from 17% one year earlier.
S&P Global · 2025 · n=1,006of agentic AI projects will be cancelled by end of 2027: cost, unclear value, inadequate risk controls. A forecast, not a fate.
Gartner · Jun 2025of large-scale IT budgets already goes to operating and maintaining existing systems. Every new app lands on this ledger.
US GAO · Jul 2025Coding agents produce working applications in hours; business units that never wrote software now commission it in an afternoon. Yet AI programs stall at a rising rate, and the cause is operational, not technical. Regulation has made it a board topic: DORA, NIS2, and the EU AI Act hold the organization accountable for the operation of AI-built systems, regardless of who built them. Most AI strategies still have no operations chapter. This paper is that chapter, written to be adapted.
The AI-Native Operations
Maturity Model
Five levels, from "it works" to "it runs, governed". Locate your organization, pick a target level, and let the gap set the priority. One question suffices: for the last application a business unit wanted, how many calendar days passed between "it works" and "it runs in production, with a record"?
Operated portfolio
The software is run for you: deployed, fixed, patched. Your people just approve.
- L1
Backlog
Building is fast; going live is a ticket to central IT. Months, every time.
- L2
Paved road
Golden paths, but for the dev team only. Everyone else is back at level 1.
- L3
Governed self-service
IT sets guardrails once; any team ships inside them. Days, not months.
- L4
Operated portfolio
The software is run for you: deployed, fixed, patched. Your people just approve.
- L5
Proven autonomy
Level 4, plus proof: the run documents itself, and the exit is tested, not promised.
Level 3 governs the launch; level 4 takes over the run. For a portfolio that grows by tens of applications a year, the difference compounds. First published by nyxory, July 2026. Cite freely with attribution.
Eight requirements.
Written to be testable.
Whether the answer is an internal build, a service provider, or a product, a strategy paper can adopt this list as its procurement baseline.
- Accepts software from any source
- Governance as an envelope, not a queue
- Operations taken over, not tooled up
- Every action on the record, per application
- Humans approve; the layer executes
- Cost governed per application
- Sovereignty at the operational tier
- Exit tested, not promised
The paper closes with a drafted operations chapter, vendor-light, written to be lifted into your AI strategy paper as it stands.
Tobias Nentwig
Founder & CEO, nyxory. Writes about the operations side of AI-native software: what it takes to run what agents build, governed and on the record.
Judge the model against your own application.
Bring one real project to the Idea-to-Production Workshop and leave with it live, operated, and on the record. Then decide what level 4 is worth to your strategy.