We rolled AI code assistants into our Gulf software programs six months ago. This was not a trial but an adoption of standard tooling, which would be used by all the developers on all projects. The move at that point seemed to be risky. Artificial intelligence technology was not yet proven to be reliable at scale, the creators were skeptical, and the claim about the productivity remained unclear as to whether it would be met in practice when it was put into action.
The information has already been gathered. According to industry research, 84 percent of developers across the world have embraced AI tools. Our teams in the Gulf have passed this mark of reference much faster than expected, and the metrics of delivery are a vivid example of the changes that take place when AI coding assistance is implemented as a routine and not an experiment.
This is what ended up happening when we implemented AI developer tools on an end-to-end basis on projects in the UAE, Saudi Arabia, and the expanded Gulf region.
The Implementation
We anticipated resistance. The developers are notorious for doubting those tools that are supposed to make their work easier. The earlier efforts to roll out new development tools were met with resistance, slow uptake, and later dropped when the advocates diverted their efforts elsewhere.
AI coders were very different. After the first month, sixty percent of the teams were adopted. By the third month of the year, it had increased to eighty-five percent. The developers who refused to use these tools initially saw their colleagues taking shorter durations to deliver features and quietly adopted the technology unannounced.
This trend was repeated in all teams. Senior developers adopted it the fastest as they saw the direct benefit in offloading repetition. It was followed by junior developers who, using AI help, passed through the learning curve much faster and produced the code that met the team standards without having to be reviewed by the seniors. Mid-career developers who were the most skeptical in the first place changed the last but became the most ardent supporters.
Adoption was not an issue that was compelled by managers but rather by developers noticing their colleagues spending less time on Stack Overflow, fewer tickets, faster, and their work logs being closed sooner. In situations where productivity gains are visible to other workers, tools are self-propagating.
Also read: Why UAE Enterprises Keep Choosing the Hybrid Cloud Strategy
What Changed in the Delivery Timeline Data?
The strongest evidence was due to the comparison of the timelines of delivery before and after the introduction of the AI tools. We were tracking the sprint velocity, feature completion rates, and time-to-production in eight parallel Gulf projects.
The average increase in the speed of sprint was thirty-five to forty percent in two months of application of AI tools. For example, teams that were already performing forty story points per sprint were regularly achieving fifty-five to sixty points per sprint thereafter. The workload was not made easy; teams did not ccorner-cut in fact, they simply wasted less time on mechanical code work.
Time spent developing the features reduced significantly in the case of standard CRUD features. Implementation of a functionality that a regular user needed to use had taken twelve to fourteen days to be finished, but the developer was reduced it to seven to eight days. Authentication integrations that formerly took a full sprint were done in half, leaving space to add additional functionality.
What Developers Really Use AI Tools
A review of the real use of AI code assistants by the developers revealed developments contrary to our anticipations. The self-evident applications were not the ones that had the highest value.
The most significant percentage of the AI-assisted time was spent on the development of test cases. Test writing is a generally disliked activity among developers, even though they are aware of its significance; the fact that test coverage is exhaustive makes them overlook the importance of test writing. AI technologies expedited the test-making process to the point that it was actually done by developers. This resulted in an increase of test coverage between projects to eighty to eighty-five percent without external requirements and pressure.
The production of documentation reflected a very high utilisation that was not expected. The level of inline code commenting, function documentation, and API documentationhasd significantly improved. The AI tools were generating accurate documentation as they developed, and not after. Documentation was no longer written to constantly become outdated, as it took only a little effort to remain relevant.
Collaboration Changes and Team Dynamics
The use of AI devices caused unexpected changes in the functioning of Gulf development teams. Some of the changes led to improved cooperation, whereas some were new challenges that are yet to be resolved.
The systems of knowledge sharing changed. Senior developers spent less time answering basic syntax or framework questions, since junior developers would get such answers through AI applications. This liberated senior staff in architectural discussions and solving complex problems. Junior developers appreciated the feedback loop as it was instant, and they did not feel awkward soliciting colleagues.
The code-review discussions were more strategic. Instead of pointing out the lack of error maintenance or incompatibility of formatting, reviewers focused on business logic, security consequences, and architecture congruency. As a result, reviews were seen as more useful by the reviewers and the ones under review.
Related: What Works in the Gulf When 84% Struggle with Cloud Costs
Challenges That Emerged
All the problems were not corrected with AI developer tools. New issues arose, and these we are currently dealing with.
Overdependence on AI proposals sometimes pproducessolutions that are not the best. The developers sometimes took AI-generated code without fully understanding it, thus resulting in technical debt. To address this, we made it a priority of code reviews to aim at comprehending and architectural appropriateness, as opposed to just the functional correctness.
There was a slight increase in inconsistency in code patterns. Disagreement in the prompt style between developers created functional code that handled problems in a haphazard fashion. We injected common prompt libraries and codes so as to drive AI tools in an attempt to produce more stable outputs.
The issue of security required new revision processes. Dependencies in AI-generated code had an occasional vulnerability to known vulnerabilities or patterns vulnerable to injection attacks. In line with that, the use of automated security scanning was implemented earlier in the development cycle, and the focus of security-related code reviews was increased.
Issues about learning and skill development came up. The junior developers whose productivity was boosted through the use of AI tools were sometimes not able to gain the in-depth knowledge they could have developed through writing additional code manually. We aim at meeting the needs of productivity increase and the need to develop basic skills.
What We Would Do Differently
Six months of data enlightened us on what ought to have been done at the beginning and enabled us to establish practices that produced better results than expected.
At a later stage, we should have put up AI-specific code-review rules. Teams developed their own practices naturally, thereby creating inconsistency. The early quality shortcomings would have been avoided withthe standardisation of expectations related to the review of AI-generated code.
In conclusion
Prompt engineering training produced unexpected value. Groups that were trained in terms of how to approach AI codes of coding strategically had better results and were faster. Early engineering education is what should have been invested in, rather than the developers’ self-discovering ways.
We help persuade the Gulf software teams to go through the process of adopting AI developers and tools at Blesssphere without forgetting the opportunities and challenges involved. The rate of adoption of these tools by the industry is eighty-four percent, which confirms their effectiveness. Whether or not to adopt them is not the relevant question, but how to do it intentionally and set up appropriate guardrails and expectations.
Continue reading: Essential AWS Security Audit Guide: When Data Breaches Cost $4.44M