The illusion of control in complex systems

Most dashboards are built to give leadership a sense of control over a complex system. Stephen Townshend — who led platform services at IAG in New Zealand at the time of this conversation, and hosts the Slight Reliability podcast — thinks that sense of control is often an illusion.

He talks through the executive dashboard his team built that nobody could actually act on, why engineers quietly build "shadow networks" that route around official escalation, why clean boundaries between teams and systems can be a comforting fiction, and why mean time to recovery is a more misleading metric than most teams realise.

Chapters

• 0:01:28 — Stephen's path from performance engineering to SRE
• 0:04:15 — The executive dashboard nobody could act on

If we're not using data to make decisions, then it's not valuable.

• 0:07:47 — The two kinds of value in observability data

Having clear boundaries between different components in these complex systems that we build — it's an illusion. It gives us a sense of control, but we don't actually have any.

• 0:10:24 — Why documented rationale beats perfect documentation

It's very easy to track things like mean time to recovery, which have now been shown to be mathematically flawed.

• 0:14:14 — Shadow networks: how big orgs really route around blockers
• 0:18:39 — Why clear boundaries between teams can be an illusion
• 0:20:22 — MTTR and the problem with mean-based incident metrics
• 0:24:20 — Where Stephen still doesn't trust AI with a decision

Key takeaways

• Townshend's own motto, after pushing back on a dashboard nobody could act on: "if we're not using data to make decisions, then it's not valuable."
• Engineers in large organizations build informal "shadow networks" of people they trust to actually help — bypassing the official escalation chart because it gets bogged down in bureaucracy.
• Clear boundaries between teams and systems can be a genuine illusion: in orgs where delivering one feature requires six teams, staying strictly in your lane actively works against success.
• Mean time to recovery, one of the most widely tracked incident metrics, has been shown by independent research to be statistically unreliable — incident durations are heavily skewed, which makes the average a misleading summary.
• Even Townshend's own "single pane of glass" project has a performative element: executives get visibly reassured by a dashboard about a system they're paying millions for, whether or not it's the most useful thing his team could be measuring.
• His approach to AI mirrors a broader pattern in the conversation: useful as a smart search layer across fragmented documentation, much harder to trust with a decision that actually matters.