AI agents meet the enterprise reality check
Autonomy demos well and deploys badly. The companies making it work are the ones narrowing scope rather than widening it.
The agent pitch is straightforward: describe an outcome, let software plan and execute the steps. In controlled demonstrations it works. In production it meets systems that were never designed to be driven by a probabilistic planner.
Where deployments stall
- Permissions: most internal systems have coarse access models, so an agent either has too little authority to finish or too much to be safe.
- Observability: when a ten-step process fails at step seven, teams need a trace, not a transcript.
- Recovery: the hard engineering is not the happy path, it is the retry, the rollback and the handoff to a human.
What working deployments have in common
The successful implementations look modest on paper. A single workflow. A bounded set of tools. Explicit human approval at the point where an action becomes irreversible. Logging detailed enough that an operations team can audit yesterday's decisions this morning.
Narrow agents in well-instrumented systems outperform general agents in poorly instrumented ones, every time.
The pattern echoes earlier automation waves. Robotic process automation promised general office autonomy and delivered value only where processes were already documented and stable. The lesson survived the change of technology.
About the author
Senior Writer, Software & Startups
Daniel Okoye covers software business models, developer tooling and early-stage company building. He has spent a decade analysing SaaS pricing and go-to-market strategy.
