Last month, GOV.UK quietly released results from its AI coding assistant trial. No fanfare. No ministerial press conference. Just solid data showing that 1,000+ developers across 50 departments had saved nearly an hour a day using AI tools.
After writing about the £54,000 Microsoft Copilot debacle that delivered “extremely small” productivity gains, this felt a little different. The same government that bungled one AI rollout in one area had simultaneously nailed another.
So What Did They Do Differently?
The AI Coding Assistant (AICA) trial ran from November 2024 to February 2025. Unlike the scattergun Copilot approach, this was targeted: 2,500 licences offered to developers who actually code for a living. The results were substantial.
Time savings averaged 56 minutes per working day. That’s 28 working days saved per developer annually. Not 2.2 hours per week like other trials—nearly an hour every single day.
More importantly, the satisfaction metrics told a different story:
- 72% said the tools offered good value for their organisation
- 58% wouldn’t want to go back to working without AI assistance
- 65% completed tasks faster, 56% solved problems more efficiently
Compare that to the general Copilot trial where only 30% used it daily and most couldn’t identify when AI was making things up.
What They Got Right: The Fundamentals
They picked the right people. Instead of randomly selecting civil servants and hoping for the best, they targeted developers—people who already understood code review, debugging, and quality control. These users had the skills to evaluate AI output critically.
They chose appropriate tools. GitHub, Copilot and Google Gemini Code Assist aren’t perfect, but they’re purpose-built for coding tasks. The tools matched the job, unlike general-purpose chatbots being asked to revolutionise everything.
They measured what mattered. Rather than vague productivity promises, they tracked specific metrics: time saved on code creation, analysis, and review. The data shows developers saved 24 minutes daily on coding and analysis alone.
They maintained quality standards. Only 15.8% of AI-suggested code was accepted without edits. That’s evidence that developers were doing their jobs properly, reviewing and improving AI output rather than blindly accepting it.
The Human Element Preserved
One of the most striking differences with this trial is that it didn’t try to replace developers. The focus instead was on amplifying their existing skills.
The 39% of users who reported committing AI-suggested code were still making informed decisions about what to accept, modify, or reject. The AI became a sophisticated autocomplete, not a replacement programmer.
As The Gen AI Academy’s experts put it:
“The best AI implementations don’t eliminate human judgement—they give humans better raw material to work with.” Erik Schwartz
“The majority of successes I’ve seen from companies on this journey comes from starting with the users, what they struggle with or are lacking, and enabling them through training and guidance to incorporate the right tools. With clear ground rules and objectives, they can contribute to and measure against” Hugo MC Pinto
This trial succeeded precisely because it preserved the human element that makes good software development possible: critical thinking, quality review, and contextual understanding.
Why This Matters Beyond Government
The coding trial offers a blueprint for successful AI implementation anywhere:
Start with skilled users. Don’t expect AI to magically make unskilled people skilled. Give it to people who already understand the domain and can evaluate the output.
Match tools to tasks. Stop trying to use general AI for everything. Specialised tools work better for specialised jobs.
Measure specific outcomes. “Increased productivity” is meaningless. “24 minutes saved on code creation” is actionable data.
Expect human oversight. If 85% of AI output needs editing, that’s not a bug—it’s working as intended.
The Uncomfortable Truth About AI Success
The government’s coding trial succeeded because it was boring. No grand promises about transformation. No claims about replacing entire departments. Just a straightforward question: can AI help developers write code faster?
The answer was yes, with proper implementation, training, and realistic expectations.
Most organisations fail with AI because they’re trying to solve the wrong problem. They want AI to fix their dysfunctions, eliminate their training requirements, or transform their culture. The government’s coding trial worked because it had a simple goal: make good developers slightly more efficient.
What’s Next?
The trial results don’t influence future procurement – apparently that decision sits elsewhere in government. But the success provides a template that other departments (and organisations) should study carefully.
The contrast between this trial and the Copilot failure centres on implementation, user selection and having realistic and testable expectations of what AI can actually do.
When I wrote about the Copilot trial, several people asked whether I thought AI in government was doomed to fail. This coding trial suggests it isn’t – but success requires doing the work properly rather than hoping technology alone will solve organisational problems.
The government got AI right by treating it as a sophisticated tool requiring skilled users, not as magic that transforms anyone into an expert. That’s a lesson worth remembering for anyone implementing AI, whether in Whitehall or your local startup.
Helena McAleer is the co-founder of thegenAIacademy.com . She connects organisations implementing AI with real-world experts who know how to deliver results the right way – and yes, she still uses the em dash!
