The problem AI success creates

The biggest risk in AI transformation is no longer adoption. It's success.

The narrative to date has been largely about experimentation and upskilling. Get teams comfortable and fluent with AI. Build agents. Prove value. Celebrate wins.

This part works. We've seen it with our own team and with clients: 6x-12x improvements on specific workflows (time-to-complete basis), teams genuinely transforming from task executors to orchestrators with higher quality outputs as the models rapidly improve.

But here's what we're experiencing now, both internally and with the organisations we work with: the challenge isn't getting teams to adopt AI. It's what happens when they do, at scale, over time.

The scale of success

I speak from experience here. Across our team and client work, we've now developed over 600 Claude agents. Many are brilliant, finely tuned agents that compress hours of work into minutes. Some are duplicates of each other. Some have been abandoned. A few do essentially the same thing as three other agents, built by people who didn't know they existed.

We're seeing the same pattern emerge with clients who've completed training. The transformation works. Then the questions start: "How do we know what agents exist across the team?" "Who's maintaining these?" "How do I prove ROI to leadership?"

This isn't failure. It's success creating new problems.

Seen in practice

One example: we recently found four near-identical agents built in parallel across a marketing ops team. Each solved the same brief-writing problem. None of the creators knew the others existed.

Agent sprawl: The new shadow IT

McKinsey's research on agentic AI (Seizing the agentic AI advantage, June 2025) highlights the risk of what they term agent sprawl: the uncontrolled proliferation of redundant, fragmented, and ungoverned agents across teams and functions. When low-code platforms make agent creation accessible to anyone, organisations face a new kind of shadow IT, with agents multiplying across teams, duplicating efforts, and operating without oversight.

That description matches exactly what I've seen happen. Not through negligence, but through enthusiasm.

The data tells the story

The data reinforces this. According to the OutSystems-KPMG survey (Taming AI sprawl starts with agent-aware SDLC management, October 2025), 44% of enterprises already cite increased technical debt and AI sprawl as major sources of risk. Research from Boomi (Navigating the AI Agent Governance Gap, October 2025) is even more striking: just 2% of tech leaders say their AI agents are held fully accountable for their actions and governed in an always-on and consistent manner.

The business case for addressing this is straightforward: duplicated effort and compute waste, abandoned agents creating brittle workflows, inconsistent outputs raising compliance questions, and an inability to prove ROI when the CFO asks. This isn't a future problem. It's a present one that compounds quietly while everyone's focused on the next use case.

The language shift

What I find particularly interesting is the language shift emerging in how we talk about this. Dataiku's analysis (Agent Sprawl Is the New IT Sprawl, 2025) suggested that "agent sprawl is to AI what shadow IT is to enterprise software." The parallel holds. Shadow IT happened not because people were being reckless, but because they were solving problems faster than IT could respond. The same dynamic is now playing out with AI agents, and the consequences of leaving it unmanaged are arguably larger.

A brilliant agent that only one person knows about isn't an asset. It's a liability waiting to walk out the door. The skills that got you here (building AI agents, context engineering, architecting knowledge bases) aren't the skills that get you to the next stage. Now you need discoverability, ownership, governance, measurement.

The emerging agentic enterprise

The MIT Sloan Management Review and BCG research on the agentic enterprise (The Emerging Agentic Enterprise: How Leaders Must Navigate a New Age of AI, November 2025) makes a similar point: strategic oversight, ethical governance, and the ability to orchestrate human-AI teams are becoming the most critical skills as AI agents take on tasks previously performed by humans.

That word, orchestrate, keeps coming up. It's not about controlling AI or slowing it down. It's about conducting the whole thing so it produces something coherent rather than noise.

What mature AI adoption looks like

What does mature AI adoption look like? Based on what we've learned, it's not complicated:

• A single source of truth for what agents exist
• Named owners responsible for quality
• Review cycles that prevent drift
• Metrics that leadership can trust

Not bureaucracy, but lightweight governance that makes teams faster, not slower. The infrastructure that turns ad-hoc brilliance into repeatable capability.

This is why we've built a control centre for the organisations we work with. Not because governance is glamorous, but because without it, the transformation stalls. The early momentum fades. Leadership loses visibility. And the CFO starts asking questions no one can answer with confidence.

The next frontier

The next frontier of AI transformation isn't whether teams can build agents. We've largely solved that. The question now is whether organisations can sustain them, and prove their value clearly enough to secure investment for the next team.