Your journey to AI-Native, with Moderta.

Whether you are transforming an established organization or building a new company from scratch, we guide the journey end-to-end. We bring senior advisory, a research-backed method and our own AI platforms for research, software delivery and governed orchestration.

Two journeys Transform · BuildThree platforms Research · Delivery · Orchestration
New companyBuild AI-nativeEstablished orgTransformAI-nativeorganizationResearch platformJira AI dev pipelineModerta OrchestrationAdvisory & expertise + Moderta platforms
01 · WHY A STRUCTURED JOURNEY

Make AI adoption a structured journey.

Access to AI is no longer the constraint. Model access alone does not create business value. The research shows that results depend on what surrounds the technology: redesigned workflows, ready data, clear governance and people who know how to work with AI.

88%
of AI proofs-of-concept never reach wide-scale deployment. Only 4 of every 33 make it to production
IDC with Lenovo, 2025[5]
42%
of companies abandoned most of their AI initiatives in 2025, up from 17% the year before
S&P Global Market Intelligence, 2025[4]
60%
of AI projects without AI-ready data will be abandoned through 2026, Gartner predicts
Gartner, 2025[3]
~6%
of companies are AI "high performers". They are nearly 3× as likely to have fundamentally redesigned workflows
McKinsey, State of AI 2025[2]
70%
of the effort in successful AI programmes should go to people and processes, 20% to technology and data, and 10% to algorithms
BCG "10-20-70" principle[1]
78%
of AI users bring their own AI tools to work. 60% of leaders say their company lacks an AI vision and plan
Microsoft & LinkedIn Work Trend Index 2024[7]

How our journey addresses each failure mode

Failure modeWhat the research showsHow the Moderta journey responds
Pilot purgatoryMost POCs never reach production[5][4]Every pilot has baselines, exit criteria and a scale decision (Phase 3)
Data not readyMost organizations are unsure their data practices are ready for AI[3]Fix the data for priority use cases first (Phase 2), not everything at once
Tools without redesignWorkflow redesign is among the strongest contributors to EBIT impact[2]Lighthouse workflows are redesigned end-to-end, not just given AI tools
Technology-only focus70% of the effort should go to people and processes[1]Four parallel workstreams: strategy, platform, people and governance
Shadow AI & no planEmployees adopt AI faster than leadership plans for it[7]A shadow-AI audit and AI policy in Phase 1, then one governed AI layer
02 · CHOOSE YOUR STARTING POINT

Two journeys, one destination

Established organizations and new companies face very different problems. One has to change systems, processes and habits while keeping the business running. The other has a blank sheet but limited resources and no room for expensive architectural mistakes. We run a tailored journey for each.

03 · JOURNEY A · TRANSFORM

From scattered pilots to an AI-native operating model

Our transformation journey follows Moderta's five-step approach (Understand, Assess, Design, Enable, Evolve), organized into five phases. Each phase moves four workstreams forward together, because technology alone does not produce value. The bottom row shows where each Moderta platform is used.

1 · Understand & AssessWeeks 0–62 · Design & FoundationsMonths 2–43 · ProveMonths 3–64 · ScaleMonths 6–185 · EvolveOngoingStrategy & valueValue map &use-case portfolioTarget state &costed roadmap2–3 lighthouseworkflowsPortfolio acrossfunctionsNew products &pricingPlatform & dataData & architecturereadinessKnowledge layer,model gatewayFirst agents inproductionOrchestration layer,LLMOpsLearning loops,model routingPeople &ways of workingSkills baseline,shadow-AI auditAI champions,training planRedesign teamworkflowsNew roles, operatingrhythmHuman + agentteamsGovernance & riskAI policy, riskclassificationIdentity, permissions,audit designEvals & approvalgatesContinuousmonitoringEarnedautonomyModertaplatformsResearchOrchestrationJira pipelineJira pipelineOrchestrationOrchestrationJira pipelineAll platforms
Figure 1. Transformation roadmap. The four workstreams advance in parallel, and each phase has explicit exit criteria. Durations are indicative and are confirmed after the readiness assessment.
PHASE 1

Understand & Assess

Indicative: 4–6 weeks

What we do

  • Executive interviews and value mapping
  • AI readiness assessment across data, architecture, security, skills and governance
  • Inventory of current AI use, including shadow AI
  • Use-case discovery and prioritisation by value, feasibility and risk

You receive

  • AI readiness scorecard
  • Prioritised use-case portfolio
  • Risk register and AI usage policy draft
  • Business case for phase 2

Moderta platforms

AI Research Platform speeds up market, competitor, vendor and regulatory research for your context, with every finding cited and reviewed by our consultants.
PHASE 2

Design & Foundations

Indicative: months 2–4

What we do

  • Target-state AI architecture (knowledge, models, orchestration, governance)
  • Data readiness fixes for priority use cases
  • Identity, permissions and audit design
  • Governance aligned to NIST AI RMF / ISO/IEC 42001[8][9]

You receive

  • Reference architecture and roadmap
  • Governed knowledge layer for first domains
  • Model gateway with vendor flexibility
  • AI champions network and training plan

Moderta platforms

Moderta Orchestration Jira AI Dev Pipeline are configured on your knowledge, roles and permissions. We index your code templates, packages and past Jira work into the pipeline's knowledge base.
PHASE 3

Prove: lighthouse workflows

Indicative: months 3–6

What we do

  • Redesign 2–3 high-value workflows end-to-end, not just add tools
  • Deploy the first role-based assistants with human approval gates
  • Pilot the AI development pipeline with one or two engineering teams
  • Set up evaluation, monitoring and cost tracking

You receive

  • Workflows live in production with measured baselines
  • Evaluation suite and quality dashboards
  • Documented playbook to replicate
  • Scale-up decision backed by data

Moderta platforms

Jira AI Dev Pipeline Moderta Orchestration AI Research Platform
PHASE 4

Scale across the organization

Indicative: months 6–18

What we do

  • Roll proven patterns out across functions
  • Expand assistants to leadership, sales, operations, HR and finance
  • Bring every engineering team onto the AI delivery pipeline
  • Establish LLMOps, model routing and cost governance

You receive

  • One governed AI layer across the business
  • New roles: AI product owners, knowledge stewards, evaluation engineers
  • Operating rhythm for AI portfolio management

Moderta platforms

Moderta Orchestration Jira AI Dev Pipeline
PHASE 5

Evolve: AI-native

Ongoing

What we do

  • Increase agent autonomy where evaluation data shows it has been earned
  • Redesign team structures around human and agent teams
  • Explore AI-enabled products, services and pricing
  • Run quarterly reviews of value, risk and cost

You receive

  • A continuously learning organization
  • Fractional AI leadership on demand
  • Ongoing platform improvements
Why the journey is designed this way
  • Four parallel workstreams reflect BCG's finding that most of the value comes from changing people and processes, not from algorithms[1].
  • Workflow redesign is a phase of its own (Phase 3) because it is one of the factors most strongly linked to EBIT impact in McKinsey's research[2].
  • The people workstream starts in Phase 1. Employers expect 39% of core skills to change by 2030, 85% plan to prioritise upskilling and 77% plan to upskill specifically for AI[6].
  • Governance is built in from the start, aligned with the NIST AI Risk Management Framework (Govern, Map, Measure, Manage)[8] and ISO/IEC 42001, the first AI management-system standard[9].
04 · JOURNEY B · BUILD

Launching a company that is AI-native from day one

New companies have one big advantage: nothing to undo. They can design workflows, teams and architecture around AI before habits form. Our build journey helps founders make the high-leverage decisions early and gives them an engineering engine that lets a small team ship like a large one.

1234BlueprintAI-native coreLean launchScale by learningWeeks 0–4Months 1–3Months 3–6Month 6+• AI-first business model• Value loop & data strategy• Human vs agent work design• Governance-by-design• Market research (cited)• Reference architecture• Knowledge & model gateway• Code templates & packages• AI delivery pipeline• Evals from the first commit• Small senior team + agents• Role assistants for ops, sales, finance• Feedback capture in product• Unit economics per task• Learning loops improve product continuously• Model routing for margins• Earned autonomy for agents• Hire for judgement, not throughput
Figure 2. The build journey. The architecture, delivery pipeline and governance are designed in from the start, so the company scales through learning loops rather than headcount. Durations are indicative.
STEP 1 · BLUEPRINT

Design around AI

We test the business model against the "remove the AI" question. We also decide which work humans own and which agents carry, and where proprietary data and feedback will build a moat. AI Research Platform

STEP 2 · CORE

Stand up the stack

The AI-native reference architecture is right-sized for a startup. On day one we deliver templates, internal packages and a knowledge base the delivery pipeline can reuse. Jira AI Dev Pipeline

STEP 3–4 · LAUNCH & SCALE

Ship lean, then compound

Role-based assistants run back-office work under human approval. Usage signals flow back into prompts, retrieval and models, so quality and margins improve together. Moderta Orchestration

The evidence for building AI-native from day one
  • In Y Combinator's Winter 2025 batch, about a quarter of startups reported codebases that were roughly 95% AI-generated (founders' own estimates)[10].
  • According to Stripe, the top 100 AI companies reached $5M in annualised revenue in 24 months. The top 100 SaaS companies of 2018 took 37 months[12].
  • Startup teams are getting smaller. Carta data shows average Series B headcount falling from 53 to 45 and a median seed team of 4. Funding conditions explain part of this trend as well as AI[11].

Speed only compounds if quality keeps up. That is why Journey B puts quality gates, evaluations and governance in place from the first commit (see the evidence on AI-generated code below).

05 · OUR PLATFORMS

Three Moderta platforms that speed up the journey

Advisory tells you what to do. Platforms help you do it. We built three platforms on the AI-native principles in our research on AI-native companies: retrieval first, human approval at critical points, cost-aware model selection and full auditability. The diagram shows where each one sits in the AI-native reference architecture.

GOVERNANCE & IDENTITY6 · Experience: Jira, chat, role workspaces5 · Orchestration & agents4 · Models: tiered by complexity3 · Knowledge: RAG, templates, history2 · Data & integration: Jira, Git, docs1 · Infrastructure: cloud, CI/CDModerta OrchestrationLayers 2–6 + governance · every roleJira AI Dev PipelineLayers 2–6 · engineering deliveryAI Research PlatformLayers 3–6 · evidence & knowledgeShared by all three: retrieval-first design,human approval gates and a full audit trail.
Figure 3. Moderta platforms mapped onto the AI-native reference architecture. Each platform covers a different area of work, and all three run under one governance model.
06 · PLATFORM
R

Moderta AI Research Platform

Multi-source research, verified and cited, with human review

AI decisions should start from evidence. Our research platform runs AI research agents that break a question into sub-questions. They search many sources, cross-check claims, and produce a structured report in which every finding links to its source. Consultants review each report before it informs a decision.

Questionbusiness contextPlansub-questionsAnalyst & industryAcademic researchRegulation & standardsVendor & market dataYour internal docsVerifycross-check, gradeSynthesizecited reportExpert reviewhuman sign-offunverified claims → re-search
Figure 4. Research platform flow. Agents search in parallel across source types. Claims that cannot be verified are sent back for another search or flagged, and a human expert signs off every report.
Readiness & strategyIndustry benchmarks and peer practices for your AI roadmap
Technology evaluationVendor, model and tool comparisons with evidence
RegulationEU AI Act, NIST, ISO 42001 and sector rules for your use cases
Market & due diligenceMarket sizing and competitor analysis for new ventures
Why the platform verifies and cites everything
  • Finding information is costly. Interaction workers spend about 19% of their time tracking down information[29].
  • AI makes skilled people faster, within limits. In a preregistered experiment with 758 BCG consultants, those using AI on tasks inside AI's capability frontier completed 12.2% more tasks, worked 25.1% faster and produced work of more than 40% higher quality. On a task outside the frontier, they were 19 percentage points less likely to reach the correct answer[28]. This is why an expert reviews every report.
  • Unverified AI research is unreliable. A test of eight AI search engines found that more than 60% of answers cited news sources incorrectly[31]. Stanford found that specialised legal AI tools hallucinated in roughly 17–33% of queries[30]. Our verify stage and mandatory citations are designed for exactly this risk.
07 · PLATFORM
J

Jira AI-Driven Development Pipeline

Turns Jira from a task tracker into an AI-assisted software delivery orchestrator

The pipeline integrates with your Jira workflow and automates ticket refinement, code generation, testing and pull-request creation. It uses retrieval-augmented generation (RAG) over your own code, templates and past tickets, chooses the cheapest model that can handle each task, and keeps engineers in control at two explicit checkpoints.

1 · Jira workflow with human checkpoints

BacklogSprintAI RefinementAI ReadyAI DevelopmentCode ReviewDone⚡ auto-trigger on sprint start👤 Human approvesrefined requirements⚡ auto-trigger on AI Ready👤 Developer reviewsand approves the PRrequirements, recommendations,clarifying questionsbranch, code, tests,PR, status in Jira tab

2 · AI engine: retrieval first, generation second

Jira eventticket → AI ReadyIngest & classifytype, complexityRAG retrievalsemantic searchSimilarsolution?Match → reuse code, tests & docsadapt with fast, low-cost modelNo match → assemble contextknowledge base + advanced reasoning modelyesnoRefinePlanGeneratefrom templatesQuality gatestsc · lint · tests · securityCreate PREscalateto developerpassfail → AI fix & retest (bounded attempts)attempts exhausted
Figure 5. AI engine flow. The pipeline searches for existing solutions before generating anything new. Generated code must pass automated quality gates before a pull request is opened. Failures go through a limited number of fix-and-retest attempts and are then escalated to a developer.

3 · RAG knowledge base

Vector DBembeddings + metadataInternal packagesCode templatesPast Jira tasksArchitecture docsTest librariesExtract → Chunk → Embed → Index → Continuously update

The knowledge base stays in sync with approved code, completed Jira issues and engineering documentation. Merged PRs feed approved patterns back into the knowledge base for future tasks.

4 · Cost-optimised model selection

TaskApproachModel tier
Simple / similarReuse RAG output, minimal generationFast, low-cost
MediumEnhance existing solution, generate codeBalanced
Complex / newArchitecture reasoning, planning, implementationAdvanced

Models are selected by task complexity, how much reusable knowledge exists and expected cost, rather than sending every task to the most expensive model.

5 · Template-based generation

AI fills your approved templates for controllers, services, modules, UI components and tests, so generated code starts from your engineering standards and is checked before review.

6 · Outputs & visibility

Git branchIssue-linked implementation branch
Pull requestGenerated code, test results and change summary
Jira updateAI Development tab with status, decisions, logs and artifacts
Team notificationAlerts developers when human review is needed

7 · What the research says, and how the design responds

AI coding assistance is one of the most-studied areas of AI at work. The results are positive overall but uneven, and that variation shaped the pipeline's design.

Measured effect of AI coding assistance on developer productivityControlled and field experiments; positive = faster or more outputControlled task, GitHub Copilot (2023)+55.8% fasterField experiments, 4,867 devs (2025)+26% tasks completedGoogle enterprise RCT (2024)~21% fasterExperienced OSS devs (2025)19% slowerSources: Peng et al. 2023; Cui et al. 2025; Paradis et al. 2024; METR 2025. Studies differ in task type and method, so the bars are not directly comparable.
Figure 6. Gains depend on context. Large gains appear on well-defined tasks[13][15][14]. Experienced developers working in large, familiar codebases were measured 19% slower, even though they believed they were 20% faster[16]. METR's 2026 follow-up suggests a speed-up but calls the estimate unreliable[17]. We measure the impact on your teams rather than assuming it.
Design choiceEvidenceWhy it matters
RAG over your own code and past ticketsRetrieving repository context improved code completion by more than 10% over in-file baselines[22]. Retrieval helps only when the retrieved context is high-quality[23]A curated index of approved code, templates and solved tickets, not your entire repository history
Model routing by complexityRouting cut LLM cost by more than 2× without quality loss, and by over 85% on one benchmark at 95% of GPT-4 quality[24]. Model cascades matched GPT-4 at up to 98% lower cost in the best case[25]Cheaper models handle simple or similar tasks. Advanced models are reserved for genuinely new work
Automated quality gates45% of AI-generated code samples failed security tests across more than 100 LLMs[27]. Duplicated code blocks rose about 8× in 2024[26]. More AI adoption has been associated with lower delivery stability[18][19]Compilation, linting, tests and security checks must pass before a pull request exists, with limited fix-and-retest attempts
Templates & internal packagesThe 2024 rise in copy-paste and fall in refactoring point to maintainability risk[26]Approved patterns and reusable modules guide generation and are checked during review
Two human checkpoints46% of developers distrust the accuracy of AI tools and only 3% highly trust it[20]. 30% have little or no trust in AI-generated code[19]Humans approve requirements before AI builds anything and review every pull request before merge
End-to-end workflow, not a plug-in68% of developers save more than 10 hours a week with AI, but 50% lose more than 10 hours a week to organizational inefficiencies[21]. DORA finds AI "amplifies" existing strengths and weaknesses[19]Automating refinement, implementation, testing and status updates in Jira removes friction around the code, not just in it
Agentic implementationSWE-bench evaluates coding systems on repository issues. Verified is a human-filtered subset of 500 tasks; results depend on the agent, tools and evaluation setup.[34]We start with well-specified tickets and evaluate results on your repositories. Requirements approval and human PR review remain mandatory.
08 · PLATFORM
O

Moderta Orchestration

One governed AI layer. Purpose-built help for every role.

Moderta Orchestration runs role-specific assistants for leadership, engineering, sales, operations, HR, finance and compliance. Each assistant is connected to the company knowledge, workflows and permissions that role is allowed to use. Assistants draft, prepare and coordinate the work. People keep control of approvals and critical decisions.

Role-awareEach assistant is scoped to its role's permissions and approved knowledge
Workflow coordinationPrepares cutovers, proposals, invoice reviews and more
Human approvalDrafts go to human review, with no unsupervised production access
AuditabilityEvery action is recorded, and models can be swapped
The evidence for role-based, governed assistants
  • Assistants raise productivity, most of all for less experienced staff. In a study of more than 5,000 customer-support agents, AI assistance increased issues resolved per hour by 15% on average (14% in the 2023 working paper), with gains of up to 34% for novice and low-skilled workers[32].
  • Human approval matters at the edge of AI's capability, where unchecked AI use made consultants' answers less accurate[28].
  • A governed layer is the answer to shadow AI. 78% of AI users bring their own tools to work[7]. Giving them approved, permission-aware assistants is safer than banning AI.
09 · WAYS TO ENGAGE

Start where it makes sense for you

You don't have to commit to a full transformation on day one. Most clients start with an assessment or a focused pilot and expand once they have results.

Indicative: 4–6 weeks

AI Readiness Assessment

Scorecard, use-case portfolio, risk register and a costed roadmap. Backed by the AI Research Platform.

Indicative: 8–12 weeks

AI Delivery Pilot

The Jira AI Development Pipeline on one or two teams, with your templates indexed, baselines measured and a decision on whether to scale.

Indicative: 6–18 months

Transformation Programme

The full Journey A, or Journey B for new companies, with Moderta Orchestration as the governed AI layer.

Flexible

Fractional AI Leadership

A fractional CTO, AI architect or security architect embedded with your team, plus independent technical reviews.

10 · MEASURING SUCCESS

What we measure, and when

We set baselines before anything changes, so results are measured rather than assumed. This matters because people's sense of AI's impact can differ sharply from what is measured[16]. Software delivery is tracked with DORA metrics[19]. Typical measures:

AreaMeasuresSource of truth
Business valueCycle time, cost per task or case, revenue per employee, EBIT contribution of AI use casesFinance & operations data
Software deliveryChange lead time, deployment frequency, change failure rate, failed deployment recovery time; PR rework and reuse as additional measuresJira, Git and CI/CD logs
AI qualityEvaluation scores, human override and rejection rate, escalation ratePlatform evaluation logs
CostModel spend per task and tier mix (low-cost, balanced, advanced)Model gateway
Adoption & peopleActive users by role, training completion, employee sentimentPlatform analytics, surveys
RiskPolicy violations, audit findings, incident rateAudit trail, governance reviews
11 · FAQ

Common questions

Do we need perfect data before starting?

No. Phase 1 identifies which data matters for your highest-value use cases. We fix that data first rather than launching an organization-wide data programme before any value is delivered.

Will the Jira pipeline replace our developers?

No. Developers approve requirements before AI starts work and review every pull request. The pipeline takes on repetitive refinement and boilerplate implementation, so engineers spend more time on design, review and complex problems.

Are we locked into one AI vendor?

No. Our platforms use a model gateway and route each task to an appropriate model tier. Models can be swapped as prices and capabilities change.

How do you handle security and compliance?

We design for permission-aware access, least privilege for agents, human approval for consequential actions and full audit trails. Governance aligns with recognised frameworks such as the NIST AI RMF and ISO/IEC 42001, and with the EU AI Act where it applies.

We are a startup. Is this too heavy for us?

Journey B is right-sized: a lightweight blueprint, a lean reference stack and the delivery pipeline from day one. Our aim is to help you avoid costly rework later, not add process.

Let's map your journey to AI-native.

Tell us where you are today, whether that's scattered pilots, a legacy estate or a blank sheet. We'll recommend the most useful first step.

Book a consultation →
SOURCES

References

  1. BCG, "Scaling AI Requires New Processes, Not Just New Tools" (2026), source of the 10-20-70 principle. https://www.bcg.com/publications/2026/scaling-ai-requires-new-processes-not-just-new-tools
  2. McKinsey, "The state of AI in 2025: Agents, innovation, and transformation". https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025
  3. Gartner press release, "Lack of AI-Ready Data Puts AI Projects at Risk" (Feb 2025). https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
  4. S&P Global Market Intelligence, Voice of the Enterprise 2025, reported by CIO Dive. https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/
  5. IDC with Lenovo (2025), reported by CIO.com. https://www.cio.com/article/3850763/88-of-ai-pilots-fail-to-reach-production-but-thats-not-all-on-it.html
  6. World Economic Forum, Future of Jobs Report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025/digest/
  7. Microsoft & LinkedIn, 2024 Work Trend Index. https://blogs.microsoft.com/blog/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work/
  8. NIST, AI Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework
  9. ISO/IEC 42001:2023, AI management systems. https://www.iso.org/standard/81230.html
  10. TechCrunch, "A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated" (Mar 2025). https://techcrunch.com/2025/03/06/a-quarter-of-startups-in-ycs-current-cohort-have-codebases-that-are-almost-entirely-ai-generated/
  11. Carta, State of Startup Compensation H2 2025. https://carta.com/data/startup-compensation-h2-2025/
  12. Stripe 2024 annual letter, reported by TechCrunch (Feb 2025). https://techcrunch.com/2025/02/27/stripe-ceo-says-ai-startups-are-growing-faster-than-saas-ever-did-and-calling-them-wrappers-misses-the-point/
  13. Peng, Kalliamvakou, Cihon, Demirer, "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot" (2023). https://arxiv.org/abs/2302.06590
  14. Paradis et al. (Google), "How much does AI impact development speed? An enterprise-based randomized controlled trial" (2024). https://arxiv.org/abs/2410.12944
  15. Cui, Demirer, Jaffe, Musolff, Peng, Salz, field experiments at Microsoft, Accenture and a Fortune 100 firm (2025). https://economics.mit.edu/sites/default/files/inline-files/draft_copilot_experiments.pdf
  16. METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (Jul 2025). https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  17. METR, uplift study update (Feb 2026). https://metr.org/blog/2026-02-24-uplift-update/
  18. Google Cloud, 2024 DORA Accelerate State of DevOps Report. https://cloud.google.com/blog/products/devops-sre/announcing-the-2024-dora-report
  19. Google, DORA 2025 "State of AI-assisted Software Development". https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report
  20. Stack Overflow Developer Survey 2025, AI section. https://survey.stackoverflow.co/2025/ai
  21. Atlassian, State of Developer Experience Report 2025. https://www.atlassian.com/blog/developer/developer-experience-report-2025
  22. Zhang et al., "RepoCoder: Repository-Level Code Completion Through Iterative Retrieval and Generation" (EMNLP 2023). https://arxiv.org/abs/2303.12570
  23. Wang et al., "CodeRAG-Bench: Can Retrieval Augment Code Generation?" (2024/2025). https://arxiv.org/abs/2406.14497
  24. Ong et al., "RouteLLM: Learning to Route LLMs with Preference Data" (ICLR 2025). https://arxiv.org/abs/2406.18665
  25. Chen, Zaharia, Zou, "FrugalGPT" (2023). https://arxiv.org/abs/2305.05176
  26. GitClear, AI Copilot Code Quality 2025 research. https://www.gitclear.com/ai_assistant_code_quality_2025_research
  27. Veracode, 2025 GenAI Code Security Report. https://www.veracode.com/resources/genai-code-security-report/
  28. Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (Harvard Business School / BCG; Organization Science). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
  29. McKinsey Global Institute, "The social economy" (2012). https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-social-economy
  30. Stanford HAI / RegLab, legal AI hallucination benchmarking (2024). https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
  31. Columbia Journalism Review / Tow Center, "AI search has a citation problem" (Mar 2025). https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
  32. Brynjolfsson, Li, Raymond, "Generative AI at Work" (QJE 2025; NBER WP 31161, 2023). https://www.nber.org/papers/w31161
  33. Jimenez et al., "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" (2023). https://arxiv.org/abs/2310.06770
  34. SWE-bench team, official SWE-bench Verified benchmark and evaluation methodology. SWE-bench Verified

Note: Survey statistics reflect each source at its publication date. Forecasts (e.g. Gartner) are predictions, not observed outcomes. Productivity studies vary in method and context. Phase durations on this page are indicative planning estimates, not research findings. Diagrams are original illustrations by Moderta Native.