Part I — I Built AI into My SDLC: Am I Gaining or Losing?

In first articlein this series we said that AI in the SDLC is no longer experimentation: it consumes tokens, opens pull requests, requires governance.
In second articlewe proposed three measurement workflows
WF-COST for the cost of the AI-enabled process,
WF-USE for the use of AI in the DevOps phases,
WF-RISK for the net value less risks. Three compasses, three formulas, three lenses aimed at the same phenomenon.

Compasses measure, not evaluate.

The question that the income statement of each table - from the CTO to the delivery director, from the founder of a software house to the IT manager of a company that develops in-house - takes home in the evening is more direct:

Am I gaining or am I losing? And when and how do I understand it?

• 5.1 — Does my business increase the net margin on projects?
• Does it reduce the time to market of your products?
• Can you afford short delivery cycles, extreme lean, without paying for it in quality or risk?
• And above all:What is the threshold below which AI in the SDLC is destroying economic value instead of creating it?

This article proposes afourth workflow — WF-LEVA — which is the natural evolution of the previous ones: takes the outputs of the three compasses and relates them to the pure human production baseline, transforming three measurements into a binary verdict. I gain or I lose. And if I earn, how much.

The Fourth Circle remains in the background — the 15 dimensions still apply. But here the look changes:no longer measuring the phenomenon, but deciding whether the business is really involved in it.

The answer is not generic.
It depends on multiple factors. Among the 15 dimensions we can indicate some:

• how mature the business is in adopting and understanding the SDLC process,

• by the type of context in which the business operates,
• how agile IT Governance is,

• how present and consolidated DevOps practices are,
• from how he orchestrated the team of agents,

• and by a couple of architectural preconditions without which the lever doesn't start - it doesn't "start less", it just doesn't start.

Let's see with formulas, tables and reference thresholds how to go from there“Do I gain or lose?” to“How much do I gain/lose?”.

The added value of WF-LEVA compared to the workflows of the previous piece is precisely this:WF-COST, WF-USE, and WF-RISK each measure one dimension of the AI-in-SDLC phenomenon — cost, deployment, net value of risks.
They are compasses.
But none of the three answers the business question.
WF-LEVA brings them to synthesis, taking their output and relating it to the pure human modality. The three compasses tell youWhere are you. WF-LEVA tells youif you're making money.

Part II — The four parameters that decide whether you really make money from it

The formula has four parameters, and each answers a typical objection you hear in the company when talking about AI leverage.

SDLC process maturity

Typical objection: “we have had Copilot for a year, but we don't see the ROI”.

If you are atL1(AI ad libitum, without governance), it is normal. If you are atL5(agentic team orchestrated with FinOps AI + governance built-in), the opposite would be strange. M maturity scales from L0 (AI absent) to L5 (agents orchestrated). It's not a scale of adoption: it's a scale ofability to transform AI into value.

Note of continuity with the previous piece:anyone who has read the WF-USE workflow will recognize it$M$such as the maturity scale implicit in the signature$I{AI}$​. Here we make it explicit as the dominant parameter of the formula: without a declared level of maturity, there is no calculable leverage — there is only story.

Quality of human orchestration

Typical objection: “but our team is at L5, we have agents in production — why doesn't the income statement benefit from it?”

Because starting from L4,$H.$becomes the dominant parameter. It's in$[0, 1]$and measure how much rigor you have put into the human orchestration of the agents: who decides what autonomously, with which guardrails, with which validation. An L5 team with$H = 0{,}2$produces worse than an L3 team with$H.$high.The maturity of the process does not save the mediocrity of the orchestration.

Precondition: rigor of the architectural patterns of the agents

Typical objection: "we chose the best platform, it works great in demo".

Demos don't fail. Production yes. The Claude Code case documented in April 2026 is emblematic:"a missing retry cap let 1,279 Claude Code sessions run 50 or more consecutive compaction failures each. Before the engineering team noticed, the bug had burned roughly 250,000 API calls in a single day". Not a model bug. An error ofarchitectural pattern: The retry cap was not configured. LangChain State of Agent Engineering 2026 reports that“over 60% of agent production incidents related to state management failures”, and that 30% of autonomous agent runs end up in exceptions requiring recovery.[agentmarketcap.ai] [valuestreamai.com]

$P{agent}$ is enabling binary: either you have implemented checkpoint, rollback pattern, multi-layer versioning (code · prompt · template · tool schema), budget-control and retry cap — or the lever doesn't start. It will not startless: it just won't start.[agentmarketcap.ai] [buildmvpfast.com]

Gartner estimates that 40% of agentic AI projects will be abandoned by 2027 — not because the models fail, but because“the pipelines around them did”. They are failed projects$P{agent}$​. [agentmarketcap.ai]

Precondition: cloud native SDLC ecosystem

Typical objection: “we have agents who write a lot of code, but then production goes haywire”.

The most impressive case is told by CircleCI in its ownState of Software Delivery 2026, analyzed by Thoughtworks: 28 million CI workflows, average throughput +59% year over year, butMedian main-branch team throughput down 7%, success rate at five-year low, recovery time increasing.
Reading Thoughtworks is surgical:"AI boosts individual effectiveness and local throughput, but without corresponding investment in the systems those changes flow through, the net effect can be negative. More code enters the pipeline, more of it fails, and recovery takes longer. The result is not acceleration — it is congestion". [thoughtworks.com]

$E{cloud}$is the second enabling binary precondition. Cloud native means: mature CI/CD pipelines (integrated SAST/DAST/SCA), end-to-end observability, IaC, policy-as-code, security-as-code, ephemeral environments to validate AI-generated code before merging.Without Ecloud​, AI speeds up the start of the cycle but slows down the end.The code accumulates in feature branches, it doesn't make it to production. It's thereAI congestion.

The operating rule of the formula:

$If \; P_{agent} = 0 \; or \; E_{cloud} = 0 \; \Rightarrow \; Lever{AI} \leq 1$

You can be at L5, you can have the best orchestrator in the world — if one of the two preconditions is missing, the leverage is not expressed.

Part III — Team A vs. Team B: The Profiter's Paradox

Let's imagine two teams working on the exact same SDLC project.
Same customer base, same stack, same type of product.

Team Auses AI little or badly.
Some developers have Copilot, but no systematic governance. Zero agents in production. The development mode is essentially the pre-AI one, with some local acceleration on individuals.

Team Buses AI a lot. Multi-step agents in production, formal orchestration, significant seat budget and token consumption.

The correct economic question is not “which team spends the most”.AND:“which team produces the most net value for the business — and under what conditions.”
The answer is not trivial.

Let's try to identify some of the dimensions that could come into play in the adoption of AI in the SDLC.

Dimension	Team A(a little/badly)	Team B without Pagent​ + Ecloud​	Team B with Pagent​ + Ecloud​
$C{consumption}$	bass	high	high
$C{human}$	high (full baseline)	average	bass
$C{governance}$	baseline	insufficient	structural
$C{risk}$	baseline	explodes(rollback, remediation, incidents[valuestreamai.com], [getdx.com])	content
$C{SDLC}$	1.0-1.2 × baseline	2,0-3,0× baseline	1.6-2.0 × baseline
Delivery time	baseline	reduced but with high rework	greatly reduced
Code Coverage	baseline	↑↑ (agents write tests)	↑↑↑(auditable tests[docs.github.com], [github.blog])
Change Failure Rate	baseline (10-15%)	>30%(tier low[getdx.com])	0-5%(elite tier[getdx.com])
LevaAI​	1,0	0,4-0,9⚠️	2,3-3,5🚀

The paradox in one line:

Team B without preconditionsspends more than Team A and produces worse. Team B with preconditions spends even more — but the value/cost ratio explodes to triple the baseline.

And it is here, in this table, that we see why the typical conversation in the company —“if we invest in AI we will obtain savings”— is badly placed.AI does not produce savings.AI produceslever, but only within an architecture that knows how to extract it. Out of that architecture, AI produces incremental spend with lower quality.

A contextual data that still today does not receive the attention it deserves: in the Opsera 2026 benchmark on 250,000 developers in 60+ enterprises,“AI reduces time-to-PR by up to 58%, but AI-generated pull requests wait 4.6x longer in review and introduces 15-18% more security vulnerabilities”. And LinearB, on 8.1 million PR:“Acceptance rates for AI-generated PRs are significantly lower than manual PRs (32.7% vs. 84.4%)”. [opsera.ai] [linearb.io]

That is: in the modeTeam B without preconditions, the number of open PRs grows, the speed of each one increases at the start, and the same PRs spend more than four times longer in the review queue — with acceptance rates less than half. It is the empirical portrait ofAI congestionthat Thoughtworks talks about.[thoughtworks.com]

Part IV — The L5 bifurcation: when the human hand decides everything

The L5 level is where the bifurcation becomes maximum. Because at L5, by definition, you already have process, culture, tooling and budget maturity. You have no shortage of conditions. You just lack — or have too much of — the quality of human orchestration.

Same level of process maturity. Same platform. Same volume of product code. The economic result bifurcates into two opposite trajectories.

Red branch — L5 misconfigured

The reference data is impressive and must be read all together:

• Gartner:40% of agentic AI projects will be abandoned by 2027, not due to failure of the models but due to failure of the pipelines[agentmarketcap.ai]
• LangChain:60%+ of agent production incidentsare linked to state management failures[valuestreamai.com]
• 30% of autonomous agent runsends up in exceptions that require recovery[valuestreamai.com]
• LinearB:acceptance rate AI PR 32.7% vs 84.4%of manual PR[linearb.io]

In terms of$Lever {AI}$: $V{prod}$it is very high (the agents produce a lot), but$Q{eff}$collapses ($Rw$ , $CFR$over 30%, frequent rollbacks, opaque compliance), e$C{SDLC\_AI}$explodes ($C{consumption}$out of control,$C{governance}$​ non-existent,$C{risk}$maximum). With$H \approx 0{,}2$and at least one of the two preconditions is compromised, the final number dropsbelow 1. Business loses.
But the team continues to report successes, because the activity graphs are splendid.

Boxout — The Case of $47,000 in 11 Days

November 2025. Two LangChain agents in production enter a conversational loop that no one intercepts. They run continuously foreleven days. The final bill documented by the Zylos Research 2026 analysis:$47,000in API fee, for a single case. It is one of the episodes that led the community to formalize that“an unconstrained agent solving a software engineering task can cost $5–8 per task in API fees alone”and what the agents do“3–10x more LLM calls than simple chatbots”. [zylos.ai], [zylos.ai]

In the WF-LEVA model that case simultaneously weighs on three components of the$C{SDLC\_AI}$: $C{consumption}$exploded,$C{risk}$maximum,$C{governance}$late.
And it reduces to zero - or rather, to negative - the$V{prod}$of that workflow. It's not an isolated case: it's a pattern.

Green branch — L5 well configured

Same ingredients. Only one thing changes:the human hand configured the agent team well in a rigorous cloud native ecosystem.

• Retry caps defined, checkpoint patterns applied[agentmarketcap.ai]
• Rollback engineeringbuilt-innot bolted-on:“rollback capability must be designed into agent systems from the ground up” [agentmarketcap.ai]
• Multi-layer versioning (code · prompt · model · tool schema) pinned[buildmvpfast.com]
• CFR maintained within rangeelite DORA 0-5% [getdx.com]
• Increased coverage with auditable test generation[docs.github.com], [github.blog]
• RACI of the agents clear (who orchestrates what, who validates what)
• Observability in production and tight feedback loops — enabled by$E{cloud}$mature

$V{prod}$very tall,$Q{eff}$very high ($CC \uparrow$, $Rw \downarrow$, $CFR \downarrow$, $Comp \uparrow$), $C{SDLC\_AI}$tall butpredictable and governed. With $H \approx 0{,}9$and preconditions both verified, the leverage flies beyond 3-5x.

The heart of the 2026 debate, in one sentence:

At L5, the value is not produced by the AI. It is produced by the human being who has configured the team of agents well within a cloud native ecosystem. AI is the power. The hand is yours.

Part V — The satisfaction thresholds: the answers to the four questions

These thresholds are the operational summary that the three workflows of the previous piece — WF-COST, WF-USE, WF-RISK — alone do not provide. It is the added value of WF-LEVA: it takes the three measurements and translates them into economic verdicts that can be actioned at the management table.

And we arrive at the moment of truth of the piece. The four questions posed at the beginning deserve four concrete answers, with reference numbers, not with rhetoric.

5.1 — Does my business increase the net margin on projects?

Operational response: +5% to +25% net margin, depending on the maturity segment. With the non-negligible risk of going negative at low maturity.

• L1-L2→ net marginNEGATIVE or flat(-3% ÷ +5%). You pay for seat + nascent governance, you gain ~10-15% individual speed, but the rework eats everything. Thoughtworks/CircleCI says it explicitly on 28 million workflows: throughput +59% year on year,main-branch median throughput down 7%. Translation: the more code that enters the pipeline, the less it makes it into production. Zero margin.[thoughtworks.com]
• L3-L4→ net margin+8% ÷ +18%. “Sustainable but not miraculous” territory confirmed by DX on 400+ companies: median gains 5-15% in PR throughput industry-wide. This is where the three classic compasses (measurement of cost, use, risk) are enough to justify the investment to the CFO.[getdx.com]
• L5 well configured→ net margin+20% ÷ +40%. With a significant caveat:20% of organizations capture 74% of AI value. The fat margin exists, but it is not distributed. It is concentrated in those who have both preconditions verified.[rize.io]

A paradoxical fact to highlight: whosellsAI coding assistant (Cursor, Windsurf) has margins“very negative”— is documented by their own insiders. Who theuse it wellwithin a cloud native ecosystem it has excellent margins.The edge is not in the AI ​​product: it is in the orchestration of AI within a mature process. [linkedin.com], [linkedin.com]

5.2 — Do I reduce the time to market of my products?

Operational response: –10% to –45% lead time, with a sharp jump to L5 well configured.

The precise reference data for 2026:

• Copilotin enterprise: PR time from 9.6 days to 2.4 days — 75% reduction in controlled trials[shno.co]
• Opsera benchmark 2026(250,000+ developers):“AI reduces time-to-PR by up to 58%” [opsera.ai]
• Forrester/Hidden Brains 2026: 40-60% reduction in software delivery time as market promise[linkedin.com]

But the fact that overturns the optimistic narrative:without Ecloud​, time to market can even worsen. Why AI acceleratesbeginningof the cycle (more PRs open, faster) butslow down the end(main-branch decline 7%, recovery time increasing, success rate at five-year lows).[thoughtworks.com]

Operationally translated for business:

• L1→ release times unchanged or worse (paradox effect)
• L3→ -15% ÷ -25% lead time (sustainable)
• L5 well configured→ -35% ÷ -50% lead time, with stable or improved quality

5.3 — Can I afford extreme lean?

Operational response: yes, but only at L5 with both preconditions. And you don't enable lean — bifurcation enables it.

Extreme lean means responding to the market with short delivery cycles, managing small batches, reducing WIP, maintaining high quality. That's exactly what the green branch of the L5 bifurcation enables:

• High deployment frequency with Change Failure Rate 0-5% (DORA elite tier)[getdx.com]
• Code coverage increased by 15-30% with auditable AI-generated tests[docs.github.com], [github.blog]
• Short recovery time thanks to built-in rollback engineering[agentmarketcap.ai], [valuestreamai.com]
• Tight feedback loops because$E{cloud}$​ allows you to measure quickly in production

But on the red branch — L5 without preconditions —extreme lean is structurally impossible. Because every rapid cycle introduces debt that you can't pay off (state management failures in 60%+ of cases). The result ispseudo-lean: short cycles in pipeline start, long blocks in review, unstable production. You are quicker to open PR and slower to send what you open into production.[valuestreamai.com]

The message for business:pushed lean is the natural consequence of a well-configured L5, not a choice you make. If you have orchestrated well, lean comes. If you haven't orchestrated it well, you can only simulate lean — and you pay the bill six months later, in technical debt and production incidents.

5.4 — How high must the leverage be for the income statement to really benefit?

The satisfaction thresholds — the table for the management table:

Segment	Leverage AI​	Net margin	Time to market	Lean pushed	Verdict
L1(AI ad libitum)	0.4 – 0.7	-3% ÷ +2%	unchanged or worse	impossible	⚠️It destroys value
L2(AI in a team but without governance)	0.7 – 0.9	+2% ÷ +5%	-5% ÷ -10%	no	⚠️Cost not justifiable
L3(AI in pipeline governed)	1.1 – 1.4	+8% ÷ +12%	-15% ÷ -25%	partial	✅Sustainable
L4(AI integrated with FinOps)	1.4 – 2.0	+12% ÷ +18%	-25% ÷ -35%	possible	✅Mature investment
L5 poorly configured	0.4 – 0.9	-5% ÷ +3%	unchanged with accidents	pseudo-lean	💀Disaster in disguise
L5 well configured	2.3 – 3.5+	+20% ÷ +40%	-35% ÷ -50%	natural	Real leverage value

The three thresholds that every management should know:

• Leverage ≥ 1.4— threshold ofoperational satisfaction. Below: You're doing AI theater. Above: You are producing economic value.
• Leverage ≥ 2.0— threshold ofstrategic satisfaction. Justify the cost of transition (upskilling, reorganization, governance). Below: support the existing. Above: transform.
• Leverage ≥ 3.0— threshold ofcompetitive advantage. Above this threshold you are producing a differential that competitors will take years to close. But only if you have well orchestrated the agent team in a mature cloud native ecosystem.

Below Leverage 1.0, the business loses.It's not an opinion, it's arithmetic.And this is what explains why a good part of AI initiatives in companies today produce splendid stories and disappointing financial statements.

---

Part VI — What the world says: the productivity paradox is an empirical fact

Until mid-2025, those who questioned the productivity of AI in the SDLC were dismissed as conservatives.
In 2026 the debate has changed radically, because the data has arrived.

The concept that the world has namedproductivity paradox, therefore, says exactly what WF-LEVA formalizes:AI accelerates local activity but can degrade system output if the system is not ready to receive it.

The term“productivity paradox”he was not born with AI. It was coined by the economistErik Brynjolfsson (MIT)in 1993, inspired by the famous observation of the Nobel Prize winnerRobert Solowfrom 1987:“You can see the computer age everywhere but in the productivity statistics”. It described the paradox of the 70s-80s: massive investments in IT, aggregate productivity that doesn't take off.[en.wikipedia.org] [cs.stanford.edu]

In 2026, the term has returned — forcefully — to the debate over AI in software delivery.Thoughtworks, in his analysis ofCircleCI State of Software Delivery 2026, defines it explicitly“empirical facts”:“the productivity paradox that the 2025 DORA Report also identified, and that Thoughtworks has observed across our client engagements: AI boosts individual effectiveness and local throughput, but without corresponding investment in the systems those changes flow through, the net effect can be negative”. Faros AIdedicated an entire research report on 10,000+ developers to the phenomenon — explicit title:“The AI ​​Productivity Paradox Report”. Forbesbrought it to the general public in June 2026, and in May 2026 an academic paper even appeared onarXivwhich proposes a formal variant -“The Productivity-Reliability Paradox”. [thoughtworks.com] [faros.ai] [forbes.com] [arxiv.org]

Stanford AI Index 2026offers the technical context:“on a key coding benchmark — SWE-bench Verified — performance went from 60% to 100% in just one year”. That is: the intrinsic capacity of the models is saturating. The bottleneck shifts entirely tohow you use them, not upwhat they do. The formula$Lever{AI}(M, H \mid P_{agent}, E_{cloud})$becomes almost prophetic: when the model stops being the limiting factor, all the remaining value is in human orchestration and the cloud native ecosystem.[hai.stanford.edu]

Thoughtworks on CircleCI 2026is the empirical confirmation of the paradox: analysis of 28 million CI workflows, throughput +59% year on year, but main-branch median throughput -7%, success rate at a five-year minimum, recovery time increasing. Martin Fowler's reading, quoted by Thoughtworks:“most organizations applying AI to software delivery are still treating it as a coding tool rather than a systems transformation”. Those who treat it as a coding tool are on the red branch of the bifurcation. Those who treat it as systems transformation are on the green branch.[thoughtworks.com]

LinearB State of Software Engineering 2026(8.1 million PRs, 4,800 organizations):"AI PRs wait 4.6x longer before review — but are reviewed 2x faster once picked up. Acceptance rates for AI-generated PRs are significantly lower than manual PRs (32.7% vs. 84.4%)". It is the operational portrait of the$Q{eff}$​ low: apparent speed in generation, real congestion in validation.[linearb.io]

Opsera 2026(250,000+ developers, 60+ enterprises) brings the most surgical note:“AI reduces time-to-PR by up to 58%, but AI-generated pull requests wait 4.6x longer in review and introduces 15-18% more security vulnerabilities”, And“senior engineers capture nearly 5x the productivity gains of junior engineers, exposing a widening execution gap across AI-enabled teams”. This senior/junior asymmetry is a piece of$H.$worth keeping in mind: Effective human orchestration isn't distributed — it's concentrated in senior roles.[opsera.ai]

DORA 2025had already introduced therework rateas the fifth metric, recognizing that the historical four did not capture post-AI instability. And it is the same DORA that publishes its warning against the in June 2026tokenmaxxing:“treating token spend as a performance indicator is a dangerous trap”. That is: anyone who measures token consumption as a success KPI is preparing their own$Leverage < 1$. [plandek.com] [dora.dev]

McKinsey State of AI 2025offers the data that sums it all up:only 39% of organizations report EBIT impact at the enterprise level. Less than four out of ten. The rest live in theROI miragedocumented by Sigma Infosolutions:“most enterprises report that AI makes developers feel faster, yet fewer than a quarter can prove measurable ROI”. [mckinsey.com] [sigmainfo.net]

The 2026 consensus is now clear:AI in the SDLC produces economic value for only a minority of organizations — the same minority it has invested in$P{agent}$​ and$E{cloud}$​ before scaling consumption. The others are paying for the power without having the leverage.

Part VII — What I take away, what I bring forward

Three take-aways to close.

1. The value of the leverage is in the ratio, not in the numerator.Those who only look at the AI ​​cost tell themselves half the story. Those who only look at the value produced tell themselves the other half. Economic value lives in the relationship — and the relationship has a denominator, the pure human modality, which must be written clearly. The man days × daily cost structure of the offer estimation procedure is the natural baseline, and it is the one that every structured software house already uses in presales. All you need is the courage to place it next to the AI-enabled cost and read the report.

2. Architectural preconditions are binary, not nuanced. $P{agent}$​ and$E{cloud}$​ Either they are there, or they aren't there. There is no "P{agent}half" that gives half leverage. A team of agents without retry cap and without rollback engineering is a team of agents that sooner or later burns 47,000 dollars in 11 days. An SDLC ecosystem without mature pipelines and observability is an ecosystem in which AI produces congestion, not acceleration. If one of the two is missing, the leverage is no longer lower: it isabsent. And the income statement is affected asymmetrically — because the AI ​​cost has already been paid.[zylos.ai] [thoughtworks.com]

3. The satisfaction threshold is personal, but the ranges are universal.Below Leverage 1.0 the business loses. Under Leverage 1.4 the business becomes AI theater. Under Leverage 2.0 supports the existing. Above Lever 2.0 transforms. Above Leva 3.0 it acquires a structural competitive advantage. Each management table can choose its own threshold — but it cannot choose to ignore that the thresholds exist and are anchored to market data observable in 2026.

And so, the final answer to the title of the piece:

AI is not the lever of your business. AI is the power. The leverage is you, if you have orchestrated the team of agents well in a cloud native ecosystem.

The fourth circle invoices. This workflow proposal is the basis on which to determine the conditions under which your business profits or loses.

Notes and references

Sources on the formula and the AI/human relationship

• Antúnez R.,AI TCO: The Math of GenAI, April 2026 — precision economics[rene-ace.com]
• NVIDIA,Rethinking AI TCO: Cost per Token is the Only Metric That Matters, April 2026[blogs.nvidia.com]
• FinOps Foundation,Token Economics: Managing AI Value in SaaS Model Token Costs, June 2026[finops.org]
• Rize / InformationWeek,GitHub Copilot ROI: What the Data Actually Shows, May 2026[rize.io]

Sources on architectural preconditions

• AgentMarketCap,Agent Checkpoint and Rollback Engineering 2026, April 2026[agentmarketcap.ai]
• ValueStreamAI,AI Rollback Strategies 2026, May 2026[valuestreamai.com]
• Umapathy A.,AI Agent Versioning and Rollback, April 2026[buildmvpfast.com]
• AgentMarketCap,Self-Healing Agent Pipelines 2026, April 2026 — Claude Code case[agentmarketcap.ai]
• Zylos Research,AI Agent Cost Optimization: Token Economics and FinOps in Production, February 2026[zylos.ai]
• Zylos Research,Token Budgets, Model Routing, and Production FinOps, April 2026[zylos.ai]

Sources on the productivity paradox and thresholds

• Stanford HAI,2026 AI Index Report [hai.stanford.edu]
• Thoughtworks / CircleCI,2026 State of Software Delivery Report, March 2026[thoughtworks.com]
• LinearB,2026 Software Engineering Benchmarks Report [linearb.io]
• Opsera,AI Coding Impact 2026 Benchmark Report [opsera.ai]
• getdx.com,Change Failure Rate in the era of AI, 2026[getdx.com]
• getdx.com,AI Measurement Framework, 2026[getdx.com]
• Shnoco,Time-to-Market Statistics 2026 [shno.co]
• Hidden Brains,Generative AI in Software Development: 40–60% Faster by 2026, December 2025[linkedin.com]
• Wikipedia,Productivity Paradox— entry on the origin of the term (Brynjolfsson 1993, Solow 1987)[en.wikipedia.org]
• Stanford CS,Productivity Paradox— historical background[cs.stanford.edu]
• Farrag SE,The Productivity-Reliability Paradox: Specification-Driven Governance for AI-Augmented Software Development, arXiv 2605.01160, May 2026[arxiv.org]
• Faros AI,The AI ​​Productivity Paradox Report, July 2025[faros.ai]
• Knight L.,Solving The AI ​​Productivity Paradox In Software Development, Forbes, June 2026[forbes.com]
• CircleCI,5 key takeaways from the 2026 State of Software Delivery, February 2026[circleci.com]
• CircleCI / Powell R.,DORA is right: AI is an amplifier, October 2025[circleci.com]

Sources on maturity and market consensus

• Plandek,DORA Metrics in the Age of AI, 2025-26 — rework rate[plandek.com]
• DORA Insights,Finding balance in the era of tokenmaxxing, June 2026[dora.dev]
• McKinsey,State of AI: Global Survey 2025, November 2025[mckinsey.com]
• Sigma Infosolutions,Proven ROI Framework to Measure AI Productivity, May 2026[sigmainfo.net]
• Agrawal M.,The Profitability Paradox of AI Coding Assistants, August 2025[linkedin.com]
• Dubey A.,High Costs and Thin Margins Are Threatening AI Coding Startups, August 2025[linkedin.com]

Sources on Copilot and code quality

• GitHub Docs,Increasing test coverage with GitHub Copilot [docs.github.com]
• GitHub Blog,Quantifying GitHub Copilot's impact on code quality [github.blog]

Internal corpus

• DL-M-77.1 Technical-functional analysis 01— Estimate requirements for offer
• 02-operating-procedures— Deda AI SDLC Toolchain
• The fourth circle in production: AI in the SDLC— Exploras, June 2026 (previous article)[exploras.cloud]