Skip to content
Briebug

Developer Resource Center

Sharing our recent findings, expertise, and insights with the community.

I Was Wrong About AI and Software Delivery

Angular AI by Updated 7 min read

I predicted AI would bring marginal gains. In September 2026, working mostly alone, I merged ten times the monthly commits of a six-person team.

Updated November 12, 2026: I was wrong. Here is what happened.

In September 2025 I wrote that most teams would see marginal gains from AI, and I was wrong. On our largest active platform rebuild, a team of about six engineers averaged roughly 60 merged commits a month. In September 2026, working mostly alone, I merged 601.

What I predicted

A year ago I argued that AI would not fix the reasons software projects fail. I wrote about “workslop”: AI output that looks finished but is hollow. I warned that AI helps a weak team write bad code faster. And I said that most teams would see marginal gains. You can read the original article below, unchanged.

What happened

The numbers come from a platform we have rebuilt since December 2024 for ReachWell. You can read about it in the ReachWell case study. Disclosure: ReachWell is a Briebug affiliate (a related party), not an arm's-length client.

The table shows merged commits on the platform’s main development branch. The first row is the whole team. The rows from June 2026 are my own commits.

Merged commits on the platform’s main development branch
PeriodWhoMerged commits
December 2024 – May 2026Team of about sixAbout 60 a month on average
June 2026Me120
July 2026Me158
August 2026Me307
September 2026Me601
October 2026Me460

Two engineers joined for about a month in September and October 2026. Their work is not in my own numbers.

How honest is this measure?

Commits measure volume, not value. A small commit and a large one count the same.

The raw numbers also flatter me. I work 80 to 100 hours a week.

So adjust for hours. Six engineers at about 40 hours each is about 240 hours a week, for about 60 commits a month. I work about 90 hours a week, for about 600. That is still more than 20 times more output per hour.

This is one engineer’s experience. It is not a study. But it is too large a change for me to keep the prediction I made.

Our own website

The same change shows on this website. The site moved to a new codebase on October 31, 2026, and we merged more than 70 changes in the next 12 days. On November 12, 2026, an SEO and AI-search audit went from findings to 12 live, tested changes in one day.

What made the difference

The tools alone did not do this. Five practices did.

  1. Agents do whole tasks. An AI agent reads the code, makes the change and runs the tests, from start to finish. It is not autocomplete.
  2. A test comes first. Every change starts with a test that fails, and the agent must watch it fail.
  3. Every change gets a review. An AI code reviewer reads every pull request, and every push gets a self-review first.
  4. Standing rules are written down. The agents read the rules at the start of every session. The rules come from our mistakes: each mistake becomes a rule.
  5. Evidence comes before opinion. We reproduce a problem at the layer where it was reported before we write a theory about its cause.

What I got right in 2025

AI amplifies your baseline. Without discipline, you get workslop, and you get bad code faster. That was true in 2025, and it is still true. The gains came from AI paired with that discipline, not from the tools alone.

What this means for your team

Briebug is rolling out AI training for three audiences:

  • Executives: strategy and ROI.
  • Managers: process and governance.
  • Developers: working with agents, test-first.

If you want these gains without the workslop, talk to us about bringing this to your team.


The original article (September 30, 2025)

Everybody is excited about AI. The headlines tout how models can write code, generate tests, scaffold applications, and even “automate software development.” The promise is seductive: faster delivery, fewer bugs, more features, happier clients, larger margins. Yet for most organizations, meaningful gains in speed and quality from AI remain frustratingly elusive.

Despite pouring resources into AI pilots and tooling, many teams struggle to translate that promise into real results. Tools get adopted, but productivity gains appear marginal. Technical debt still accumulates. Architects still argue. Bugs still slip into production. In practice, teams still wrestle with requirements shifts, architectural complexity, coupling, maintainability, and team knowledge gaps.

The “Workslop” Problem: AI’s Productivity Paradox

A recent phenomenon called “workslop” is shining a light on one of AI’s underappreciated risks—output that looks okay, but fails to move the needle. Workslop refers to AI-generated content that masquerades as good work but lacks depth, context, or substance.

  • TechCrunch warns about coworkers producing AI-generated “slop” that forces recipients to fix or re-interpret the work. [TechCrunch]
  • The Register reports that 40% of U.S. workers say they’ve received workslop in the last month—polished but hollow output that wastes time. [The Register]
  • The Harvard Business Review / BetterUp collaboration coined the formal definition and highlights how that kind of output clogs communication, erodes trust, and erodes real productivity. [Harvard Business Review]
  • One survey estimates workslop costs organizations about US $186 per employee per month in wasted time. [American Bazaar]
  • Some researchers see workslop as one factor behind stats that 95% of organizations report “zero return” on AI investments. [TechCrunch]

Why Software Projects Still Fail (And Why AI Can’t Fix Those Root Causes)

Before imagining that AI will rescue every project, it’s worth revisiting why software projects fail—and which failures are beyond what AI can auto-fix. Common failure modes include:

  • Unclear or shifting requirements
  • Poorly managed scope creep
  • Lack of domain understanding
  • Overly tight coupling and monolithic design
  • Insufficient architecture planning / up-front design
  • Technical debt accumulation
  • Weak testing, quality assurance, or lack of continuous integration / delivery
  • Team skill gaps, poor alignment, communication breakdowns
  • Operational, deployment, reliability, and scaling issues

The Limits of AI: Knowledge, Architecture, and Craftsmanship

AI is powerful, but it is not magic. A few key limits to keep front of mind:

  • AI does not increase domain or architectural knowledge
  • Quality of suggestions depends on training data
  • Security, compliance, concurrency, performance edge-cases [arXiv]
  • Hallucinations, inconsistencies, and context loss
  • Emerging complexity in AI systems themselves [arXiv]

The Risk: Developing Bad Code Faster

One of the more subtle dangers is that AI can help teams produce bad code faster. If your team’s baseline is flawed—coupling, copy-paste, weak abstractions—AI simply amplifies that. Worse: when developers rely on AI suggestions without sufficient scrutiny, you may end up with increased technical debt, hidden coupling, or architectural entanglement that is harder to disentangle later.

What to Do Now: Adopt AI Thoughtfully (Don’t Let It Adopt You)

If you’re reading this hoping to “turn on AI and watch quality soar,” that’s not how this will play out. But you can get meaningful gains—if you follow a structured path. Here’s a recommended adoption approach (drawing from our earlier Briebug playbook):

  • Begin with internal enablement and pilot projects
  • Invest in architectural & AI-roots education
  • Train your team in best practices for architecture and software development
  • Scale a standardized, safe delivery playbook
  • Elevate adoption to increased capability and productivity

Why Experts Matter (And How Briebug Can Accelerate Your Journey)

Many organizations stumble because they treat AI like a “features toggle” rather than a strategic shift. That’s where a partner with domain expertise and proven patterns can accelerate adoption and help you avoid landmines. Here’s what expert partners bring:

  • Real-world playbooks and best practices for integrating AI into enterprise software delivery
  • Guidance on architecture guardrails, modular design, abstraction layers, and trend evolution
  • Help in training your team (skills, mindset, prompt-engineering, code review for AI)
  • Oversight in defining governance, security, and compliance rules around AI use
  • Experience in incremental rollout, pilot design, and scaling AI safely

It’s Time to Act—Don’t Be Late

AI in software delivery isn’t coming—it’s already here. But the winners in the next two years will be those who adopt thoughtfully, pragmatically, and with discipline. The laggards—those who treat AI as a gimmick or hand over code generation without guardrails—risk producing more “workslop,” more debt, and stalled productivity.

If you’re ready to take control and accelerate your software delivery, schedule a Discovery Call with us at Briebug. Let’s discuss your projects, your goals, and how you want your engineering stack and team to evolve over the next two years. Together, we’ll build a roadmap to get you there—faster, safer, and smarter.