Bernd is the CTO and founder of Dynatrace, a unified observability and security company that helps simplify enterprise cloud complexity.
Fewer than one in 10 enterprise applications is fully observable today, according to a January 2026 Neurones IT Asia report, and a March 2026 SolarWinds survey found that 77% of responding IT teams lacked full visibility across hybrid environments. When software development was primarily human-driven, missing context slowed teams down and increased operational overhead. Now that AI agents are increasingly creating, deploying and operating code, the same visibility gap is a hard ceiling on how fast any organization can safely scale.
Most enterprises are already operating at two speeds simultaneously. In the first mode, AI augments existing workflows, helping engineers move faster and increase productivity. That remains the dominant pattern today. In the second mode, increasingly capable AI agents take on larger portions of software delivery and operations, while humans focus on goals, specifications and oversight.
For most enterprises, the question isn’t whether this shift will happen but how they should introduce it while managing risk, compliance and existing operational dependencies. Since intellectual property migrates, it stops living in the codebase and starts living in the spec because the code itself becomes a generated output. This transition doesn’t happen in a single project. Existing systems carry compliance obligations, integration dependencies and technical debt that can’t be automated away overnight.
The pragmatic path is to start with innovation teams and lower-risk applications, prove the model, build institutional confidence and let that trust expand to more business-critical workloads over time.
Agents Moving Fast—And Blind
AI agents can write, deploy and optimize software faster than any human team, but they have no awareness of what happens once that software is running in production. They can’t see whether the changes they made caused a slowdown, triggered an error cascade or violated a compliance boundary. They make decisions at machine speed without understanding the downstream consequences. Speed and blindness at the same time.
The probabilistic nature of GenAI becomes far more challenging when multiple agents operate in sequence and are interconnected. A single agent that’s 95% accurate sounds impressive until you put 10 of them in sequence, at which point cumulative accuracy drops to roughly 60%. Every wrong action can trigger the next, producing outages, regulatory exposure and financial loss that compound before any human is aware something went wrong.
Running critical operations at this scale requires a deterministic foundation where the same inputs produce the same conclusions, grounded in verified facts, live system context and a trusted data layer.
Why Observability Becomes Non-Negotiable
As agent adoption scales up, an increasingly important question emerges: How do you actually know what your AI systems are doing? Not in principle but in production, in real time, across your entire hybrid estate. Without that visibility, automation is very difficult to trust, govern or extend beyond pilot projects.
The same way a CIO wouldn’t accept running a large engineering organization without management reporting, you can’t run a large fleet of AI agents without understanding their actions and outcomes. Making AI activity transparent, accountable and auditable is a precondition for moving from experimentation to enterprise-wide deployment.
Going a bit deeper into technical matters, agents are only as effective as the context they receive, and large language models can’t reason over the full volume of telemetry that modern digital systems generate. Organizations need a data layer that transforms metrics, logs, security data and all other data types into trusted, actionable context. What separates production-grade agentic systems from proofs of concept are agents developed and grounded on a deterministic foundation.
A New Maturity Metric And The Road To Autonomy
How organizations measure success is also shifting. For teams running agent-led operations, one metric gaining traction is the percentage of completions that require no human intervention. The lower that number, the more effectively the AI is working. However, getting that number down safely without creating new risks depends entirely on continuous insight into what agents are doing and what their actions produce in production.
The path forward is a staged progression:
1. Automating well-defined, bounded tasks.
2. Moving to supervised autonomy where AI proposes a complete plan that a human reviews and approves before anything executes.
3. For organizations that have built the right foundations, the final step is full autonomous operation on appropriate workloads.
Most enterprises sit somewhere between the first and second stages right now. Full autonomy is still ahead, and at every stage, the humans remain in charge of setting goals and making consequential calls.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
