[{"data":1,"prerenderedAt":415},["ShallowReactive",2],{"\u002Fen\u002Fdossiers\u002Ftechnical-debt-ai":3,"dossiers-catalogue":382},{"id":4,"title":5,"body":6,"description":338,"extension":339,"image":340,"kind":341,"manufacturer":340,"meta":342,"metaTitle":343,"model":340,"navigation":344,"order":345,"path":346,"publishedAt":347,"relatedDossiers":340,"relatedDrones":340,"relatedServices":340,"seo":348,"sources":355,"specs":340,"stem":379,"translationKey":380,"updatedAt":347,"__hash__":381},"content\u002Fen\u002Fdossiers\u002Ftechnical-debt-ai.md","Technical debt and AI: finally clearing the legacy",{"type":7,"value":8,"toc":326},"minimark",[9,13,21,26,33,44,47,51,54,62,69,73,76,94,112,119,123,126,144,154,161,165,168,183,190,216,223,234,245,249,252,303,306,310,313,316,320,323],[10,11,12],"p",{},"For the past two years, artificial intelligence has been pitched to me in almost\nexactly the same way every time: a productivity lever, therefore a headcount lever.\nThat misses what it actually makes possible inside an information system.",[10,14,15,16,20],{},"AI does not replace a team. It makes affordable a kind of work that never was:\nmethodically taking back old software layers, documenting them, testing them, bringing\nthem up to date. In other words, ",[17,18,19],"strong",{},"clearing a technical debt"," that most IT departments\nhave been carrying for ten or fifteen years, for lack of ever being able to justify the\nspend.",[22,23,25],"h2",{"id":24},"how-much-does-technical-debt-really-cost","How much does technical debt really cost?",[10,27,28,29,32],{},"The most telling figure comes from CIOs themselves. Surveyed by McKinsey, they put\ntheir technical debt at ",[17,30,31],{},"20 to 40 % of the value of their entire technology estate",",\nbefore depreciation. In a large group, that runs into hundreds of millions.",[10,34,35,36,39,40,43],{},"The second figure hurts more, because it describes a suffocation mechanism: ",[17,37,38],{},"30 % of\nthe CIOs surveyed believe that more than 20 % of the budget officially allocated to new\nproducts"," in fact goes to dealing with debt-related issues. You fund innovation, you\npay interest. And ",[17,41,42],{},"60 %"," report that their debt has risen noticeably over the past\nthree years.",[10,45,46],{},"That is the very definition of a vicious circle: the debt consumes the capacity that\nwould allow it to be repaid.",[22,48,50],{"id":49},"why-have-modernisation-plans-been-failing-for-fifteen-years","Why have modernisation plans been failing for fifteen years?",[10,52,53],{},"No CIO is unaware of the problem. This is not a failure of clear-sightedness, it is a\nfailure of arbitration — and the arbitration was rational.",[10,55,56,57,61],{},"Taking back legacy code is massive, repetitive, thankless work, with deferred benefits\nthat are hard to defend in front of an executive committee. Upgrading an application\nacross a major version, rewriting missing tests, reconstituting lost documentation,\nremoving a dead dependency: each task is simple, there are thousands of them, and\nnobody wants to fund them. Against a business demand that ",[58,59,60],"em",{},"is"," visible, modernisation\nloses the arbitration every time. That is why we collectively failed, year after year,\non a subject everyone had correctly identified.",[10,63,64,65,68],{},"What has changed today is not our clear-sightedness. It is the ",[17,66,67],{},"unit cost"," of that\nrepetitive work.",[22,70,72],{"id":71},"what-does-ai-actually-change-on-legacy-code","What does AI actually change on legacy code?",[10,74,75],{},"Two large-scale, published and documented field reports give the measure.",[10,77,78,81,82,85,86,89,90,93],{},[17,79,80],{},"Amazon."," Announcing the results of its code transformation assistant, Andy Jassy\nstated that the average time to move an application to Java 17 fell from roughly\n",[17,83,84],{},"50 developer-days to a few hours",", for a total of ",[17,87,88],{},"4,500 developer-years"," saved\nand around ",[17,91,92],{},"260 million dollars"," in annualised efficiency gains. This is corporate\ncommunication, not an independent study — but the order of magnitude and the nature of\nthe task are verifiable.",[10,95,96,99,100,103,104,107,108,111],{},[17,97,98],{},"Google."," The field report published by its research teams is more sober and\nmethodologically more interesting: across ",[17,101,102],{},"39 internal migrations"," carried out by\nthree developers over twelve months, ",[17,105,106],{},"74 % of the code changes"," were generated by the\nmodel, and the developers estimate the total time was cut by ",[17,109,110],{},"about half"," compared\nwith earlier manual migrations.",[10,113,114,115,118],{},"Note the nature of these tasks: ",[17,116,117],{},"codified, repetitive and verifiable"," work. A version\nupgrade, an API migration, a library replacement. The success criterion is binary and\nthe result can be tested. That is precisely where AI delivers its value — and it is no\naccident that the two companies publishing these figures applied it there, and nowhere\nelse.",[22,120,122],{"id":121},"when-does-ai-make-you-slower","When does AI make you slower?",[10,124,125],{},"This is the part missing from most articles on the subject, and it is the part that\ndetermines whether your programme succeeds.",[10,127,128,131,132,135,136,139,140,143],{},[17,129,130],{},"AI can slow you down."," METR ran a randomised controlled trial with experienced\ndevelopers, on ",[17,133,134],{},"their own repositories",", mature and well known to them. The result:\nwith AI tools allowed, they were ",[17,137,138],{},"19 % slower",". The most instructive part is not the\nnumber but the perception gap: they expected a 24 % gain, and afterwards still believed\nthey had gained 20 %. ",[17,141,142],{},"On complex ground you already master, AI costs time while\ncreating the feeling of saving it."," No programme steered on gut feel can work here.",[10,145,146,149,150,153],{},[17,147,148],{},"AI amplifies, it does not repair."," That is the central conclusion of the DORA 2025\nreport, built on nearly 5,000 responses: AI does not improve an organisation, it\n",[17,151,152],{},"amplifies what is already there",". A solid team gets leverage; a team without tests,\nwithout review and without architecture mostly gains the ability to produce its defects\nfaster.",[10,155,156,157,160],{},"The practical consequence is easy to state, and it is the whole point of an\narchitecture engagement: ",[17,158,159],{},"before accelerating, you need to know what you are\naccelerating on."," A legacy estate with no map, no regression tests and no exit\ncriterion is not handled better with AI — it degrades faster. You then replace old debt\nwith new debt, produced faster and understood less.",[22,162,164],{"id":163},"why-has-obsolescence-become-urgent-in-2026","Why has obsolescence become urgent in 2026?",[10,166,167],{},"Until now, modernising was a choice: the timetable was yours. That is no longer true,\nand it is the real change of 2026.",[10,169,170,171,174,175,178,179,182],{},"For the ",[17,172,173],{},"first time in the report's nineteen-year history",", the 2026 edition of\nVerizon's ",[58,176,177],{},"Data Breach Investigations Report"," ranks ",[17,180,181],{},"vulnerability exploitation as the\nleading initial access vector, at 31 % of breaches"," — ahead of stolen credentials,\nwhich had dominated until then. The same report notes that AI on the attacker's side is\nshrinking the defence window \"from months to mere hours\".",[10,184,185,186,189],{},"The French figures from ANSSI, in its ",[58,187,188],{},"Panorama de la cybermenace 2025",", describe the\nsame pincer movement:",[191,192,193,201,210],"ul",{},[194,195,196,197,200],"li",{},"the rate at which vulnerabilities are published has grown ",[17,198,199],{},"18 % per year since\n2020",";",[194,202,203,206,207,200],{},[17,204,205],{},"about 29 %"," of the vulnerabilities exploited in 2025 were exploited ",[17,208,209],{},"on the very\nday they were published, or before",[194,211,212,213],{},"and above all: ",[17,214,215],{},"more than 6,200 assets in France were still affected, at the end of\n2025, by the main vulnerabilities exploited since 2023 and 2024.",[10,217,218,219,222],{},"That last point deserves a pause. These are not unknown flaws or sophisticated threats:\nthey are published vulnerabilities, patched by the vendor, documented by CERT-FR — and\nstill open two years later. This is not a security problem, it is a problem of\n",[17,220,221],{},"capacity to keep an estate moving",". Which is to say, a technical debt problem.",[10,224,225,226,229,230,233],{},"Add to this that automation now sits on both sides. The campaign catalogued as\n",[17,227,228],{},"C0062"," in MITRE ATT&CK, documented in late 2025, targeted around thirty\norganisations with ",[17,231,232],{},"80 to 90 % of operations carried out autonomously"," by an agent:\nreconnaissance, exploitation, credential harvesting, lateral movement. An attacker\nindustrialising against a defender patching by hand: that gap does not close with\neffort, it closes with tooling.",[10,235,236,237,240,241,244],{},"Finally, the constraint is becoming personal for company leadership. The ",[17,238,239],{},"NIS2","\ndirective widens the French regulatory perimeter from roughly 500 entities to some\nfifteen thousand, with penalty ceilings set by the directive at ",[17,242,243],{},"€10 M or 2 % of\nworldwide turnover"," for essential entities, and accountability explicitly carried by\nthe management body. As I write, the French transposition law has not yet been\npromulgated: the deadline has not passed, it is coming.",[22,246,248],{"id":247},"how-do-you-take-back-a-legacy-estate-with-ai-step-by-step","How do you take back a legacy estate with AI, step by step?",[10,250,251],{},"An AI-assisted legacy remediation programme does not start with AI. It starts with what\nmakes AI usable.",[253,254,255,261,267,285,291,297],"ol",{},[194,256,257,260],{},[17,258,259],{},"Inventory."," What is running, what is still supported by its vendor, what is\nexposed to the internet, what no longer has a known owner. It is the one deliverable\nwhose value is immediate, even if the programme stops there.",[194,262,263,266],{},[17,264,265],{},"Application mapping."," The real dependencies, not the ones on a five-year-old\ndiagram. Without it no batch can be isolated, and without an isolable batch there is\nno remediation.",[194,268,269,272,273,276,277,280,281,284],{},[17,270,271],{},"Splitting into verifiable batches."," A batch qualifies on three criteria: the work\nis ",[17,274,275],{},"repetitive",", the expected result is ",[17,278,279],{},"explicit",", and a ",[17,282,283],{},"test proves"," nothing\nbroke. All three, or the batch does not go to AI.",[194,286,287,290],{},[17,288,289],{},"Safety net first."," Before any transformation: regression tests, a behavioural\nreference, the ability to roll back. This is the item that decides everything else —\nand it is also, today, largely automatable.",[194,292,293,296],{},[17,294,295],{},"Assisted transformation, human review."," The model produces, the engineer decides.\nThe share accepted without modification is measured and tracked; it is not assumed.",[194,298,299,302],{},[17,300,301],{},"Decommissioning."," A modernisation that switches nothing off has repaid nothing:\nit has added a layer. Retiring the old service is part of the batch, not of the next\nphase.",[10,304,305],{},"Steps 1 and 2 are not administrative preliminaries: they are what determines whether\nsteps 3 to 6 produce a gain or a fresh debt.",[22,307,309],{"id":308},"what-has-this-method-been-tested-on","What has this method been tested on?",[10,311,312],{},"I do not sell a theoretical capability. Waypoint360 is a SaaS platform I designed,\nbuilt and operate: authentication foundation, billing, database migrations, continuous\ndelivery pipeline, a design system shared across several applications, large-scale\ninternationalisation. The trade-offs described here are ones I hold on a production\nsystem, not on slides.",[10,314,315],{},"The same requirement applies to this article: every figure quoted points to its source,\nand the sources are dated. On a subject where approximation travels fast, that is the\nminimum owed to a decision-maker committing a budget.",[22,317,319],{"id":318},"where-should-you-start","Where should you start?",[10,321,322],{},"With the inventory, always. It is short, it costs almost nothing, and it produces two\nthings: the list of what is exposed and unmaintained — the part that is urgent — and\nthe list of repetitive, testable batches, the part where AI will pay off immediately.",[10,324,325],{},"If you want to discuss it concretely, the most useful entry point is a status review of\nyour actual estate.",{"title":327,"searchDepth":328,"depth":328,"links":329},"",2,[330,331,332,333,334,335,336,337],{"id":24,"depth":328,"text":25},{"id":49,"depth":328,"text":50},{"id":71,"depth":328,"text":72},{"id":121,"depth":328,"text":122},{"id":163,"depth":328,"text":164},{"id":247,"depth":328,"text":248},{"id":308,"depth":328,"text":309},{"id":318,"depth":328,"text":319},"AI makes legacy remediation affordable — if you know where it pays off and where it costs. A sourced analysis.","md",null,"dossier",{},"Technical debt: what AI actually changes",true,1,"\u002Fen\u002Fdossiers\u002Ftechnical-debt-ai","2026-08-21",{"keywords":349,"title":5,"description":338},[350,351,352,353,354],"technical debt AI","reduce technical debt","legacy modernisation","information system obsolescence","enterprise architecture",[356,359,363,366,370,373,376],{"label":357,"url":358},"McKinsey — Tech debt: reclaiming tech equity","https:\u002F\u002Fwww.mckinsey.com\u002Fcapabilities\u002Ftech-and-ai\u002Four-insights\u002Ftech-debt-reclaiming-tech-equity",{"label":360,"url":361,"date":362},"Verizon — 2026 Data Breach Investigations Report","https:\u002F\u002Fwww.verizon.com\u002Fabout\u002Fnews\u002Fbreach-industry-wide-dbir-finds","2026",{"label":364,"url":365,"date":362},"ANSSI \u002F CERT-FR — Panorama de la cybermenace 2025","https:\u002F\u002Fwww.cert.ssi.gouv.fr\u002Fuploads\u002FCERTFR-2026-CTI-002.pdf",{"label":367,"url":368,"date":369},"Google Research — How is Google using AI for internal code migrations?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2501.06972","2025",{"label":371,"url":372,"date":369},"METR — Impact of Early-2025 AI on Experienced Open-Source Developer Productivity","https:\u002F\u002Fmetr.org\u002Fblog\u002F2025-07-10-early-2025-ai-experienced-os-dev-study\u002F",{"label":374,"url":375,"date":369},"DORA — State of AI-assisted Software Development 2025","https:\u002F\u002Fdora.dev\u002Fdora-report-2025\u002F",{"label":377,"url":378,"date":369},"MITRE ATT&CK — Campaign C0062","https:\u002F\u002Fattack.mitre.org\u002Fcampaigns\u002FC0062\u002F","en\u002Fdossiers\u002Ftechnical-debt-ai","tech-debt-ai","qNHTQePzRS-txgA_4qpyEsi5Ffrv67hNKfxzohkmkRw",[383,388,392,396,398,403,407,411],{"locale":384,"slug":385,"path":386,"title":387,"translationKey":385},"en","application-mapping","\u002Fen\u002Fdossiers\u002Fapplication-mapping","Application mapping: what it must contain",{"locale":384,"slug":389,"path":390,"title":391,"translationKey":389},"chatgpt-visibility","\u002Fen\u002Fdossiers\u002Fchatgpt-visibility","ChatGPT visibility: what actually works, measured",{"locale":384,"slug":393,"path":394,"title":395,"translationKey":393},"it-obsolescence","\u002Fen\u002Fdossiers\u002Fit-obsolescence","IT obsolescence: what it really costs you",{"locale":384,"slug":397,"path":346,"title":5,"translationKey":380},"technical-debt-ai",{"locale":399,"slug":400,"path":401,"title":402,"translationKey":385},"fr","cartographie-applicative","\u002Ffr\u002Fdossiers\u002Fcartographie-applicative","Cartographie applicative : ce qu'elle doit contenir",{"locale":399,"slug":404,"path":405,"title":406,"translationKey":380},"dette-technique-ia","\u002Ffr\u002Fdossiers\u002Fdette-technique-ia","Dette technique et IA : solder enfin le legacy",{"locale":399,"slug":408,"path":409,"title":410,"translationKey":393},"obsolescence-informatique","\u002Ffr\u002Fdossiers\u002Fobsolescence-informatique","Obsolescence informatique : ce qu'elle vous coûte",{"locale":399,"slug":412,"path":413,"title":414,"translationKey":389},"referencement-chatgpt","\u002Ffr\u002Fdossiers\u002Freferencement-chatgpt","Référencement ChatGPT : ce qui marche, mesuré",1787432459048]