CTO Says AI Is Causing Big Stability Issues
David Ting explains how AI boosts engineering speed while introducing stability risks, and why validation, governance, and expert oversight are essential for reliable software.
David Ting
CTO at Bespin Global
AI Transformation Strategy: David Ting is establishing an AI transformation strategy to enhance operations and develop new products.
AI Developer Tools: Using tools like Claude Code, development productivity reportedly increases by over tenfold.
Human Expertise: Despite AI's advancements, human experts remain vital for ensuring code functionality and quality assurance.
Production Challenges: AI integration can create instability in production environments, necessitating careful monitoring of performance.
Learning by Doing: Practicing coding with AI tools is more effective than traditional training, which quickly becomes outdated.
David Ting is the CTO at Bespin Global and has been integrating AI into workflows for over a decade. But in the last year, he says the game has changed.
We sat down with David to discuss the gains that he's seeing. He said it isn't all gains — stability is a big issue.
Establishing an AI Transformation Strategy
I have actively engaged with AI for over a decade. It's a fast-paced environment, and innovation has sped up even more in the past year.
Want more from The CTO Club?
Create a free account to finish this piece and join a community of CTOs and engineering leaders sharing real-world frameworks, tools, and insights for designing, deploying, and scaling AI-driven technology.
Have an account? Log In
I'm currently the CTO of Bespin Global, where I'm establishing the AI transformation strategy, from streamlining operations to incubating new products and services. We focus on cloud resale and consulting, and we have 1,000 employees worldwide.
I've also started and incubated two internal AI products — a manufacturing AI tool and a security AI tool.
I dedicate at least 50% of my time daily to coding and automating my work and life. This helps me stay current with the latest technical advancements. In fact, I have written over 200k lines of code with AI assistance in the last month.
Why Claude Code Should Be Mandated
We have hundreds of developers, and all of them use Claude Code.
We monitored velocity proactively and found that dev productivity can increase by over 10x. Personally, I've found that 10x is not the limit.
We still use the traditional dev and QA teams with a standard spec-driven workflow. The change is in the tooling.
Even the QA team writes more tests using AI — mostly Playwright tests and additional unit tests.
The dev team uses other tools, too. Personally, I use open-source tooling for most things. I also use OpenAI, Codex, Anthropic API, and Claude. I use all these simultaneously through multiple terminal windows.
How AI Powers End-to-End Engineering Workflows
We write the product spec with AI, perform architecture reviews with AI, and write code with AI.
To do this, you start with a very detailed design specification. Your AI developers will develop it, fill in details, and scrutinize the design (finding issues). Then, an agent works on and delivers each action item.
Once implementation is complete, you need human validation to ensure the software works as designed. Hook observability tools onto the code stack to monitor logs and performance metrics, and give AI a shot to optimize or fix bugs with a very clear target.
When you do this, you will find that one developer can quickly produce 20 to 60 developers' worth of work. Product polish needs expert guidance, but AI does most of the coding.
Why AI Shifts Focus From Code Quality to Output
I kept code quality as a metric that I track, but AI has shifted my focus to the number of implemented use cases and lines of code.
The reason is simple: AI increases the output. So, iterations are extremely quick.
You don't refactor. You let the results dictate where you spend your time.
I kept code quality as a metric that I track, but AI has shifted my focus to the number of implemented use cases and lines of code…You don’t refactor. You let the results dictate where you spend your time.
Why Human Experts Are Still Needed in Engineering Workflows
So, most everything we do is touched by AI. It has improved so much in the last six months alone, but human experts are still essential for guiding development. They ensure that the code functions correctly and the UX design is clean.
AI hallucinates a lot. This creates deep-seated bugs. You need someone who is intimately familiar with architectural best practices to spot these.
That's why we're shifting toward smaller teams of world-class experts. Team strength is no longer dictated by size.
Why AI Creates More Work for On-Call Staff
AI integration has obviously led to cost savings. Our digital assembly line, which is built on agentic AI, has reduced staffing requirements by 75% while increasing delivery speed.
This year, I am also experimenting with automating roles within the company through a team of agents, and I'm receiving some encouraging results.
But these gains are not the whole story.
Quality varies greatly because coding is now nondeterministic. The same prompt yields different results, which can be frustrating.
Avoiding a bug slipping into production demands significant care — prompts require careful construction, and code requires careful testing.
Because of that, people are also working harder than ever — particularly on-call staff.
How AI Affects Production Availability
We added extensive monitoring and guardrails. The Claude API is unstable in production, so we supplemented it with open-source models like Qwen.
Production availability is another issue. Our application has experienced severe quality degradation during peak usage hours.
We plan to run these apps on self-hosted LLMs as a failsafe mechanism.
Why Practical Coding Trumps Training in AI rollout
When I rolled out these changes, people asked for training before using the tools.
I should have required people to use AI in their code from the start. Instead, they took tons of training and courses.
AI is changing so quickly that the material is outdated weekly. Doing is the only way to learn.
Why Adding Developers Adds a Project Tax
CTOs should be redesigning their organizations. Because now, teams of one have maximum velocity.
Productivity depends on two things:
- How well you define tasks for AI
- How well you validate and fix the work the AI produces
Keep the great people who work well with AI. Adding people to a project adds a tax.
Why CTOs Should Assume AI Will Fail
Here's my advice: Be as aggressive as possible with adoption, but validate thoroughly.
You will find success everywhere, but you will experience failures if you lack proper validation and governance structures.
CTOs should dedicate significant time to implementing these structures — and incorporating numerous failsafes.
Assume AI will fail.
Be as aggressive as possible with adoption, but validate thoroughly…Assume AI will fail.
Follow Along
You can follow David Ting's work on LinkedIn.
More expert interviews to come on The CTO Club!
