AI coding tools can help developers finish some tasks faster, but faster code generation is not the same as faster software delivery. The gains depend on the task and workflow; enterprise agents can also be held back when the information they need is difficult to find, unavailable under current permissions, or disconnected from development tools. In one MIT Technology Review Insights report hosted by Google Cloud, organizations said AI could access an average of 45% of their enterprise data. That figure describes reported access, not a proven formula for how much faster a team will work.
Do AI agents actually make software development faster?
Sometimes, and the evidence needs to be read by what it measured. A code-completion assistant, an autonomous coding agent that can use tools, and an enterprise agent retrieving company information are different interventions. Likewise, task time, task throughput, developer perceptions, and end-to-end delivery are different outcomes.
| Evidence | What was measured | What it does—and does not—show |
|---|---|---|
| Microsoft Research, 2025 | Across randomized field experiments at Microsoft, Accenture, and an anonymous Fortune 100 company, involving 4,867 developers, AI code-completion tools were associated with a 26.08% increase in completed tasks; the reported standard error was 10.3%. | This is field-experiment evidence about code-completion assistance and completed tasks across those settings. It is not a universal estimate for autonomous agents or end-to-end delivery. |
| Microsoft Research, 2023 | In a controlled experiment, recruited developers using GitHub Copilot completed a bounded JavaScript HTTP-server task 55.8% faster. | This result applies to that specific task and experiment, not software projects generally. |
| GitLab / The Harris Poll, 2026 | In a survey, 78% of technology professionals said developers were writing and committing code faster after AI-tool adoption; 85% agreed AI had shifted the bottleneck from writing code to reviewing and validating it. | These are respondents’ reports and agreement, not independently measured causal speedups. |
These findings cannot be averaged into a single “AI makes developers this much faster” figure. Their populations, tasks, interventions, and measures differ. The 2025 field experiments offer causal evidence in their studied settings; the survey figures capture reported experience.
Why can’t AI agents access all of a company’s data?
Enterprise access is not simply a matter of connecting an agent to every database. Data may sit in separate systems, lack consistent structure or context, be stale or difficult to discover, or carry permissions that should not be widened for an automated tool. A useful agent needs the relevant information in a form it can interpret, while access remains scoped and accountable.
#1 Best Overall
A report from MIT Technology Review Insights, hosted by Google Cloud in a partnership context, says AI can access an average of 45% of enterprise data. The page also reports that 55% of executives said their current data systems actively prevent them from scaling agentic AI. The opened report page does not state its publication year. These are survey findings, not proof that opening access to the remaining data by itself would accelerate development.
Access should be evaluated for usefulness as well as breadth: can the agent find current, task-relevant information, understand its context, and act only within appropriate permissions? A percentage technically reachable says little about whether the data is usable or safe for a particular task.
Why faster code generation may not mean faster delivery
Generated code still has to be reviewed, tested, integrated, secured, and maintained. If those steps become the constraint—or if suggestions create rework—the time saved while writing code may not reduce the time from work starting to a reliable release. GitLab’s survey found 28% of respondents said their software development lifecycle tools were fully integrated with shared data and workflows. That is a reported adoption figure, not a measure of delivery performance.
DORA’s 2025 report draws on a survey of nearly 5,000 technology professionals and more than 100 hours of qualitative data. It frames AI as an amplifier of an organization’s existing strengths and dysfunctions. This is a broad industry-research interpretation, not a randomized causal finding: capable teams may use AI to reinforce effective practices, while fragmented workflows or weak validation can make added output harder to absorb.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →GitLab Chief Product and Marketing Officer Manav Khurana put the control concern this way in the company’s June 23, 2026 release: “AI coding tools have delivered on their promise of speed. But the events of the past few months, including supply chain attacks, reliability issues, and regulators tightening expectations around AI traceability and provenance are making clear that speed without control is a liability, not an advantage,”. This is a vendor executive’s interpretation, not an independent research result.
What enterprise AI usage figures can—and cannot—tell you
OpenAI’s enterprise figures describe activity among its own customers, not the market as a whole. In that customer base, Codex accounted for 64% of combined Codex and ChatGPT output tokens as of June 2026. OpenAI also reported that frontier firms generated 8.3 times as many output tokens per active user as typical firms in June 2026, up from 2.6 times in January 2026. Token volume indicates product usage; OpenAI cautions that it is an imperfect proxy for business value. None of these measures establishes that a firm delivered software faster or achieved better results.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate agents in your own development workflow
A useful pilot tests the whole workflow rather than only how quickly an agent produces code. Set a baseline, specify the tasks and participating team, and choose outcomes before deployment. Track task completion separately from end-to-end flow, and include the time and quality costs of review, testing, integration, defects, rework, and maintenance.
- Define the intervention: distinguish autocomplete or chat assistance from an agent that takes actions, and from an agent that retrieves enterprise data.
- Measure the intended outcome: record task time or throughput separately from release lead time and quality; do not treat them as interchangeable.
- Check the context: determine whether the agent can discover current, relevant project or business information, rather than counting only connected systems.
- Scope permissions: grant access appropriate to the task, log actions, and keep human review for consequential decisions or actions.
- Report the setting: state the team, task type, observation period, tool, and baseline so results are interpretable and repeatable.
A 2026 paper by University of Washington-associated researchers reports 85.1% overall accuracy and 94.4% accuracy for high-confidence predictions in a 205-participant study of a framework for predicting permission preferences. That result does not validate automatically granting permissions in production. Permission prediction is not a substitute for policy, authorization controls, or review.
Quick Recap
Sources and scope
- Microsoft Research: The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers, June 2025.
- Microsoft Research: The Impact of AI on Developer Productivity: Evidence from GitHub Copilot, February 2023.
- MIT Technology Review Insights: Scaling AI agents with trustworthy data, hosted by Google Cloud; publication year not stated on the opened page.
- DORA: 2025 State of AI-assisted Software Development Report.
- GitLab / The Harris Poll: GitLab Research Reveals Organizations Are Generating AI Code Faster Than They Can Control It, June 23, 2026; corporate release of survey findings.
- OpenAI: Enterprise signals: What frontier firms are doing differently, updated August 12, 2026; usage figures refer to June 2026 unless stated otherwise.
- Towards Automating Data Access Permissions in AI Agents, 2026.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
