As AI continues to transform software development, engineering leaders are rethinking how teams work, how tools are used, and how success is measured. At our latest CTO Breakfast Briefing, technology executives gathered to explore these shifts, share emerging practices, and examine the growing influence of agentic systems on the software development lifecycle.

One of the highlights of the session came when Aleksandar Karavasilev, CTO at Damilah, shared the results of internal AI experiments. These revealed practical gains and sparked further discussion around how AI is reshaping team workflows, engineering oversight, and productivity.

AI tools accelerate development, but measuring impact remains complex

Aleksandar opened the session by presenting findings from structured experiments using tools like GitHub Copilot and Cursor. One example showed how engineers used Cursor to analyse over 900 SQL procedures, reducing what would typically take weeks into a few days. Across 80 engineers, the company recorded over 400–500 hours saved in just three months.

While the productivity gains were clear, attendees agreed that traditional metrics—such as story points or cycle time—struggle to reflect the real impact of AI assistance. Some have started running side-by-side comparisons using past sprint data. Others rely on direct feedback from engineers to assess where AI adds the most value.

Agentic AI: Practical applications already in use

The discussion moved beyond basic AI tools to focus on agentic AI—systems that operate autonomously and carry out multi-step tasks with limited human input. Attendees shared concrete examples already in use:

1. Cursor AI for Refactoring SQL

Used to analyse and refactor legacy SQL codebases. In one case, over 900 stored procedures were reviewed using Cursor, accelerating output by up to 40x. The tool helped not only with syntax cleanup but also with identifying duplication patterns, unused procedures, and potential areas for optimisation.

2. Product Development Agents

AI agents that help automate early-stage product work, such as persona creation, solution mapping, and documentation. One attendee described using agents to rapidly develop user personas by aggregating public data, aligning it with internal user journey maps, and drafting functional product outlines that served as a starting point for cross-functional discussions.

3. Multi-Agent Systems (MAS)

Autonomous agents working in coordination—one generating code, another testing it, a third documenting it—mimicking the functions of a development team. In one example, a multi-agent system designed a basic feature end-to-end, then returned human-readable notes and integration steps, helping the engineering lead decide whether to implement or modify.

4. Legacy System Modernisation

Agentic tools helped parse and understand legacy systems (like COBOL), reducing dependency on domain-specific knowledge held by a few individuals. A participant shared how their team used AI to map out the dependencies across an aging monolithic application, flagging risky elements and suggesting components that could be containerised or rewritten.

5. Security-Wrapped Agent Frameworks

To meet regulatory requirements, some attendees described wrapping agents in code that tracks data flow for full traceability. These wrappers were used to ensure that LLM-based agents operating in sectors like healthcare and finance logged all activity, allowing compliance officers to audit and validate their outputs against policy.

6. MCP Servers for Workflow Orchestration

Using Model Context Protocol, attendees built workflows where agents autonomously accessed APIs, gathered data, and triggered actions without manual steps. One use case involved pricing intelligence: the AI queried multiple APIs for pricing data, summarised the output, sent updates via internal messaging tools, and triggered reports—automatically and on schedule.

Reviewing AI-generated code requires human oversight

Participants expressed caution about relying too heavily on AI-generated code. Some described scenarios where AI wrote code, created the tests, and approved the changes—resulting in logic flaws that went unnoticed.

In response, organisations have implemented stricter review processes:

• Senior engineers are required to approve all AI-generated pull requests
• AI-generated reviews are disallowed
• AI support is used during reviews, but not in place of human reviewers

Human accountability and measurement challenges

Participants also discussed optimal approaches to handling bugs in AI-generated code – as, they generally agreed, no AI tool is capable of producing perfect code.

In particular, “code bloat” was highlighted as a major issue. In other words, as AI operates so rapidly, it’s easy to generate huge amounts of code – and inevitably the more code that exists, the more bugs there will be.

Because of this, there was a consensus that:

• It’s essential to ensure human involvement at every step of the development process.
• “People must remain accountable for what they produce,” said one attendee.
• Code reviews – both by humans and tools such as CodeRabbit – are particularly important.

Clear rules and guardrails enable safe adoption

Governance came up as a priority. Attendees explained how their organisations define usage boundaries and assign accountability:

• Technical leads or architects determine how and where AI tools can be applied
• Guardrails are introduced to prevent misuse in code, testing, or deployment
• Some use lightweight wrappers to log all AI input and output, enabling auditability

This helps organisations embrace AI without losing control over quality, process, or compliance.

Privacy, security and regulatory requirements shape implementation

Security and compliance featured heavily in the conversation. Organisations are under pressure to ensure that AI tools don’t expose sensitive information or create audit gaps. A few attendees discussed:

• Creating data wrappers to monitor agent activity
• Preventing tools from leaking PII
• Building traceable workflows that can meet third-party audit standards

These measures allow AI to be used in sensitive industries like finance and healthcare.

Managing AI Costs and Infrastructure choices

As AI becomes more integrated into workflows, its operational cost has come under scrutiny. Attendees noted the rising expenses of using large language models through public APIs and the unpredictability of long-term pricing.

Alternatives currently under consideration include:

• Hosting lightweight private LLMs
• Running AI models locally on edge devices
• Replacing SaaS-based solutions with internal AI-driven systems built using MCP

These strategies aim to make AI adoption more sustainable while maintaining data privacy and performance.

AI is redefining engineering workflows

Attendees agreed that AI is changing more than just speed—it’s influencing how engineers work. By offloading repetitive or time-intensive tasks, AI enables developers to focus on problem-solving and innovation.

However, participants raised concerns about the learning curve for junior engineers. If AI handles too much, new developers risk missing foundational knowledge. Several organisations now combine AI tooling with mentorship programmes and hands-on training. One attendee noted that AI should be treated as a peer contributor: fast, efficient, but imperfect—always requiring human oversight.

Rethinking SaaS business models in the age of AI

Attendees discussed how AI—particularly LLMs and generative tools—are forcing product and commercial leaders to reconsider SaaS business models. One speaker described the tension between using AI to accelerate a content-driven business while also facing existential risk from the same technology. As AI agents shift how people search, consume, and generate content, traditional SEO-led monetisation strategies may become less effective. Another participant noted that the ease of creating AI-powered workflows has made it harder to differentiate products, urging businesses to focus on their unique value propositions. Several attendees agreed that pricing models would need to evolve—moving away from flat-fee SaaS toward usage-based or outcome-driven approaches—especially as LLM compute costs rise and multi-agent architectures become more common.

Final Thoughts: Success requires structure, not just tools

AI tools have moved from experimentation to execution. They’re helping engineering leaders solve long-standing problems, accelerate timelines, and explore new ways of working. But the session also made it clear that success depends on structure—on clear governance, trusted oversight, and shared learning.

Our next CTO Breakfast will focus on how organisations scale their AI practices across teams, while maintaining quality, security, and cultural cohesion.

We hope you’ll join us.