What the research says about AI and engineering context
Teams have adopted AI quickly. The published research suggests the bottleneck has moved from writing code to finding the right context, trusting the answer and keeping systems stable. Each figure below comes from the study linked under it and is quoted within the scope its authors give it. After each finding, the links point to the use-case pages that deal with that problem in depth.
How to read these figures
These studies measure different things in different ways, so read each figure on its own terms rather than adding them together or comparing them directly.
- Each study has its own scope. One analyses changed lines of code, one is a controlled study and several are surveys. Their populations, time periods and definitions differ.
- Correlation is not causation. Where a report links a trend to AI adoption, that is an association the authors observed or estimated, not proof that AI caused it.
- Survey figures report what respondents say they experience: their use, trust, frustrations and time lost. They are not measurements of their work.
- The METR study is one setting: experienced developers working in repositories they knew well, using early-2025 tools. The authors say it does not show that AI slows down most developers.
The "What it points to" notes are our reading of each finding, not conclusions drawn by the studies' authors.
AI writes new code instead of reusing existing code
GitClear analysed 211 million changed lines of code from 2020 to 2024. During 2024, code blocks with five or more duplicated lines increased eightfold, and 2024 was the first year in which copy-pasted lines outnumbered moved lines (code that was refactored and reused). The report links the shift to AI assistants, which make inserting new code easy and, with limited context, are less likely to suggest reusing an existing function.
What it points to: an agent can only reuse what it can find. CodeZero lets it search the organisation's indexed repositories, by name and by topic, before it writes something new.
Developers use AI, but do not trust its accuracy
In Stack Overflow's 2025 Developer Survey, 84% of respondents use or plan to use AI tools, yet more developers distrust the accuracy of those tools (46%) than trust it (33%). The single biggest frustration, cited by 66%, is "AI solutions that are almost right, but not quite", and 45% say debugging AI-generated code is more time-consuming.
What it points to: "almost right" is what an answer looks like when it is built on guesses. Both products show what an answer is based on (file references, sources, and whether a match was explicit, identity, exact, name, suggested or unmapped) so it can be checked before it is used.
Feeling faster is not the same as being faster
METR ran a randomised study with 16 experienced open-source developers on 246 real issues in repositories they knew well. With early-2025 AI tools, they took 19% longer. Beforehand they had expected a 24% speed-up, and afterwards they still believed AI had made them 20% faster. The authors are explicit that this is one setting and does not show that AI slows down most developers. METR has since marked these results as out of date: its February 2026 follow-up with late-2025 tools points to a speed-up, but the authors call that data an unreliable signal.
What it points to: time goes into checking, correcting and re-prompting. Reducing that means giving the agent the context up front and reviewing a plan rather than a large diff.
More AI, less stable delivery
Google's 2024 DORA report finds that as AI adoption increased, it was accompanied by an estimated 1.5% decrease in delivery throughput and an estimated 7.2% reduction in delivery stability, for every 25% increase in AI adoption.
What it points to: changes made without knowing their impact show up as failed deploys and incidents. CodeZero asks what a change affects across repositories before it is made; Vörr shows what depends on a service at runtime.
Finding information is the top time sink
Atlassian's State of Developer Experience 2025 surveyed 3,500 developers and managers. Half report losing 10 or more hours a week to organisational inefficiencies, and the leading friction point is finding information (services, docs and APIs), ahead of adapting to new technology and context switching between tools. Nearly all respondents (99%) report saving time with AI, and 68% save more than 10 hours a week; Atlassian argues those gains are offset by the same friction.
What it points to: the time AI saves on typing is lost on finding things. Shared, indexed context turns "where is it and who owns it" into a question with an answer.
Operations data lives in many tools
Grafana Labs' 2025 Observability Survey (1,255 responses) found that companies use an average of eight observability technologies, with 101 different ones in use across respondents, and that Grafana users configure an average of 16 data sources. The most frequently cited obstacle, at 39%, is the complexity and overhead of setting up and maintaining tooling.
What it points to: an assistant that can see one tool misses what the others record. Vörr queries every connected source that can answer, reports each result with its source, and maps the different names one service has across those tools.
Product facts
| What it covers | |
|---|---|
| CodeZero index | The organisation's repositories on GitHub, GitLab or Bitbucket: committed code on the default branch, including repositories nobody has checked out |
| CodeZero relationships | Imports and packages; HTTP calls matched to routes in other repositories; services named in deployment config; queue producers and consumers; tables and collections read or written. Inferred links carry file:line evidence, and HTTP and config links also carry a confidence level |
| CodeZero languages | Symbols are parsed for TypeScript, JavaScript, Python, Go, Java, Ruby and C#; other languages are searchable as text |
| Vörr cloud | AWS, Google Cloud and Azure, scanned with read-only credentials, with dependencies, exposure, blast radius and scan coverage |
| Vörr live signals | More than 30 tools across metrics, logs, traces, errors, alerts and CI/CD, for example Datadog, New Relic, Grafana, Prometheus, Loki, Splunk, Elasticsearch, Jaeger, Sentry, Jenkins and Argo CD, plus CloudWatch through AWS and GitHub Actions, GitLab CI and Bitbucket Pipelines through your code connection |
| Boundaries | Vörr is read-only towards your systems. CodeZero's agent cannot approve its own plan, and unattended runs refuse commit, push and destructive commands |