In short. With AI, a convincing application can be built in a few days. It proves that a team can produce, and nothing more; it says nothing about what the system will cost to maintain, to secure and to unwind. What has become scarce is judgement, the kind that decides, declines and answers for the result. Seven questions let an executive recognise it in their teams and their suppliers without reading a line of code.
An executive committee has two proposals on the table for rebuilding its supplier portal. The first arrives within three weeks, with an application that already works: screens, a connection to the ERP, a dashboard. The second runs to six pages and shows no screen at all, but it sets out what will be built and what will not, which choices will be hard to undo, and where access to suppliers’ bank details will live. The committee picks the first. You can see what you are buying.
Eighteen months later the portal holds three copies of the supplier master data, an approval workflow nobody knows how to change, and bank details stored somewhere the security team only discovers during an audit. None of it was badly coded. None of it, in truth, was ever decided: the choices that commit a company for years had made themselves, screen by screen, taken by a tool or a habit in place of a person. We call this decision debt. It does not show up in the tests, which pass. It shows up the day something has to change, or the day an auditor asks.
Why the demo no longer proves anything
For a long time a working demo was a strong signal, for a simple reason: building was expensive, so building was proof. That reasoning no longer holds, and the measurements of 2025 and 2026 say something more precise than “AI speeds things up”. The DORA report, which has tracked the performance of software teams for a decade, finds that AI raises throughput while degrading stability, and it draws a conclusion worth pinning to an office wall: AI does not fix an organisation, it amplifies what the organisation already is. In a randomised trial, the METR institute found that experienced developers took longer with AI assistance than without it, while remaining convinced afterwards that they had saved time. Veracode measured that nearly half of all code-generation tasks produced a catalogued vulnerability, and that the rate was not improving from one model generation to the next. In March 2026 Amazon lost a day’s worth of orders in a six-hour outage; whether you blame generated code or, as the company does, human error, the response was the same, two mandatory reviewers across three hundred and thirty-five critical systems. Felt speed, then, is not an indicator. The quality of decisions is, and it can be read without an engineering degree, provided you ask the right questions.
1. “If we had to reverse this choice in two years, what would it cost?”
A good answer separates what can be undone in a week from what commits you for years. It puts a figure on the impact of going back, in orders of magnitude, along four axes: the days of work to migrate, the volume of data to move, the number of systems and teams that depend on it, and what leaving the contract costs, notice and penalties included. A library, a screen or a report format can be swapped in a week; the data model, the identity provider, where the data lives, or a contract that makes leaving expensive, commit you for years. Architects speak of two-way doors and one-way doors, and judgement means moving fast through the first kind and taking your time at the second. The warning sign is a team that argues as long about a user-interface component as about how its data is split, or one that tells you everything is reversible because it is in the cloud.
2. “What did you decide not to build, and why?”
A good answer contains a list, because every component you add will have to be understood, maintained, patched, monitored and defended. IT departments that do not sell software learned this long before AI, with the first wave of cloud “as a service” offerings and then the SaaS products that followed: every bespoke build is a ball and chain, and “we don’t do custom any more” became doctrine. AI only strengthens the case, since generated code is harder to maintain than it is to write. The custom work that remains is the kind nobody can avoid: integration, security, the catalogue and automation. Now that code is almost free, the temptation is nonetheless back to reopen the door: one more internal tool, a local copy of a master dataset, a bridge “for now”, as if custom work had cost something to write rather than to carry. The most useful refusal is still the plainest, which is to use what the ERP already does rather than rebuild it alongside. Nor is the temptation confined to IT any more. Business teams now build their own applications with assistants, much as they once built spreadsheets that ended up running a critical process, but with a far larger footprint of access, data and dependencies. Banning this does not work. What works is a path: a sandbox to experiment in, a catalogue of what already exists, and a stop at architecture before any tool touches money, personal data or access rights. The warning sign is an answer with no refusal in it, because a decision with no rejected alternative is not a decision, it is a habit.
3. “What debt are we taking on, and what will trigger paying it back?”
A good answer is recorded in a decision register, in precise terms: this shortcut, for this reason, repaid when this volume, this date or this second business unit arrives. Every architecture carries debt and that is no flaw; chosen debt gets repaid, suffered debt gets discovered at the worst moment. Three new kinds of debt have appeared since 2025 and have been given names. Comprehension debt is the gap between how much code a system holds and how much of it any human genuinely understands, and it grows in silence because the tests pass. Intent debt is the loss of the why, of the reasons and constraints nobody wrote down. Upstream of both sits the decision debt of the first paragraph. All three are repaid by the same act, writing before building, and the most useful question in a steering committee has become this one: for each of our critical components, who in this organisation could rewrite it in the middle of an incident? The warning sign is a sentence you have heard before, “there is no debt, we did everything properly”.
4. “Where do secrets and access rights live, and who can cut them off within the hour?”
A good answer comes with the very first diagram, never in “phase two”. It says who reaches what and under which identity, where the keys are and who can revoke them, and what happens if a third-party component is compromised. None of this is new, and security has always been a precondition rather than a layer. What AI changes is the scale and the actors: the same abundance that makes code cheap multiplies components, access paths and flaws, and the agents themselves are now handed rights, often broader than a developer’s. In July 2025 a consumer coding agent wiped a vendor’s production database despite an explicit code freeze, then fabricated data to hide the damage; the tool had the access, and nobody had decided that it should not. The warning sign is security “planned for a later stage”, or the absence of anyone who can say how an access is cut off within the hour.
5. “Who reviewed what was produced, and where is that review recorded?”
A good answer fits in one sentence: nothing reaches production until a named person has reviewed it and can explain it, and that review leaves a record of who asked, who reviewed, what changed and why. Kubernetes wrote this into its contribution rules in 2026: if the author cannot explain a generated change, the pull request is closed. This is no longer a nicety; it is evidence you will be asked to produce. The NIS 2 directive makes the approval of security measures a personal responsibility of management; the Cyber Resilience Act has required the reporting of exploited vulnerabilities since September 2026 and will require a software inventory by the end of 2027; the AI Act has imposed transparency since August 2026. Three texts, one question: who decided, who reviewed, where is the proof. The warning sign is “the tests pass”, and nothing more.
6. “Who on our side will be able to evolve this system without you?”
A good answer names someone and says how that person learns. A system that only its author understands belongs, in effect, to its author, and dependence on a provider, internal or external, has become a risk in its own right. The skill needed to understand, meanwhile, is being formed less and less. A randomised trial published in early 2026 showed that engineers who delegate the writing to an assistant understand markedly less of what they ship and debug it less well, while those who use the assistant to understand keep the skill. At the same time, French IT services firms are cutting their intake of junior developers and youth employment in software is falling wherever it is measured. The judgement of 2035 is being formed today or not at all, and a client who insists that someone on its own side can evolve the system, while accepting that a supervised junior works on it, is doing something for both. The warning sign is the promise “you will have the documentation”.
7. “What would make you change your mind?”
A good answer is written in advance, as a volume, a cost or a regulatory shift that would justify revisiting the decision. This is the question that separates a judgement from a conviction, and a supplier who has done the work answers it gladly, often with relief, because it shows you are looking at what matters. The warning sign is the phrase “it’s the industry standard”.
The page that holds it together
The seven answers fit on one page, and that is the page we ask for before the first line of code of any structural choice: the context, the options considered, the choice made, what reversing it would cost, the condition that would reopen it, and the name of the person accountable. Now that assistants and agents take part in building and running systems, the security part of that page gains an AI section, which is a review gate in its own right: which agents are involved, under which identity and with which rights, on which data, and who reviews what they produce. Existing frameworks, from ANSSI’s recommendations to the OWASP Top 10 for agentic applications, help fill it in; they cover access and human oversight, not reversibility, debt or what you decide not to build, which remain a matter of judgement. That is what the seven questions add: not one more standard, but a grid that covers what those frameworks leave out. Architects call it an ADR, an architecture decision record. Ten or so pages a year, reviewed by leadership, are worth more than an architecture binder nobody opens, and they are also the documents you will produce on the day of an audit.
To find out where you stand, you do not need a programme. Among the choices of the past two years, find the three that would cost the most to reverse, and ask for that page for each of them. If nobody can write it, you have learned what you needed to know. Then put the seven questions into your next request for proposals; the answers, and sometimes their absence, will tell you more than any demo.
Output has become abundant. What has not is someone who can say why, who can say no, and who answers for the result.
Sources
- DORA, State of AI-assisted Software Development, September 2025; METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025; Veracode, 2025 GenAI Code Security Report.
- Digital Trends (from Business Insider), AI code wreaked havoc with Amazon outage, March 2026; The Register, July 2025 incident.
- Jason Gorman, Comprehension Debt, September 2025; Addy Osmani, Comprehension Debt, March 2026; Margaret-Anne Storey et al., From Technical Debt to Cognitive and Intent Debt, April 2026.
- Kubernetes, Pull request guide, AI section, 2026; Thoughtworks, AI-accelerated shadow IT, 2025-2026.
- How AI Impacts Skill Formation, randomised trial, January 2026; KPMG / Numeum, Grand Angle ESN & ICT 2026; Stanford, Canaries in the Coal Mine, 2025.
- Directive (EU) 2022/2555 (NIS 2), Article 20; Regulation (EU) 2024/2847 (Cyber Resilience Act); Regulation (EU) 2024/1689 (AI Act), Article 50.
- ANSSI, Security recommendations for a generative AI system, 2024; CERT-FR, bulletin CERTFR-2026-ACT-016 on AI agents, April 2026; OWASP, Top 10 for Agentic Applications 2026, December 2025.