Medicine got serious about checklists when the evidence got too loud to ignore, and the same reckoning has arrived for the software that schedules, monitors and cleans up after itself. The field now has a name — the AutomationOps entry in the Ops Atlas covers it — and with any young discipline come beliefs that spread faster than the evidence. Six of them keep surfacing in team meetings and vendor calls. Each is worth retiring, and the reasons why are more interesting than the myths. You can usually recognise one by its shape:
- It promises a permanent end to a category of vigilance (“you’ll never worry about that again”).
- It treats a rare event as impossible rather than rare.
- It sounds generous to humans while quietly removing their authority to act.
- It sells confidence in proportion to how little you will have to look at the system.
The six below fit at least two of those shapes each.
Myth 1: Automation removes human error
It relocates it. This is not a new observation — an engineer named Lisanne Bainbridge wrote the classic paper on it back in 1983, “Ironies of Automation”, and four decades of incidents have only strengthened the case. When a task is automated, the mistakes do not vanish; they migrate to the moment the automation is designed, the filter written, the target list configured. Design-time errors are quieter than run-time ones — nobody watches a person write a script the way they watch a person administer a drug — and they can lie dormant for months before the conditions let them fire. The practical response is not paranoia but review: treat the design of an automated task with the same seriousness as the task itself, because that is where the error now lives. Teams that internalise this change their calendars — design reviews become as routine as the morning huddle, and a proposed automation gets the same questioning a new protocol would: what does it assume, when is it wrong, who notices, how do we stop it?
Myth 2: An empty AutomationOps target list is harmless
An empty list is one of the most instructive failures in the whole field, because systems keep misreading it. A job that receives zero targets has been told nothing about intent — and jobs that plough on anyway have erased data, sent nothing, or sent everything. Google’s Diskerase service became the canonical warning after an unrepaired empty entry led to real data loss, and a 2025 cloud database outage repeated the lesson at regional scale with an empty record left sitting. The rule the Atlas proposes is refreshingly strict: empty should stop the run and summon a person, every time. No interpretation, no “probably just a quiet day”, no clever defaults — a job that cannot name its targets has no business acting on anything. In hospital terms, it is the difference between “no allergies recorded” and “no allergies” — one is a gap, the other is a finding, and confusing them has consequences.
Myth 3: More approval steps make everything safer
Approval is medicine, and dosing matters. Pile an approval onto every routine job and the approvers drown; the clicks become reflexes, and a reflexive yes approves the dangerous thing just as fast as the safe one. The evidence from operations teams is consistent: approval fatigue is its own risk. The tier model fixes the dose. Read-only work needs no gate. Reversible, single-target changes need a name and a rehearsal. Wide-reaching or destructive jobs earn a second person — soft deletes, staged rollout, the works. Attention concentrated where damage lives beats attention smeared everywhere, and the second opinion keeps its power precisely because it stayed rare. There is a subtler benefit too: when approvals are rare, the approver is senior enough and awake enough to actually read what they are approving. Fatigue does not just slow the queue; it degrades the judgement inside it, which is the one resource no amount of tooling replaces.
Myth 4: Once a task is automated, the work is done
Automation rots, quietly. Filters reference people who left, calendars get renamed, upstream formats drift, and the job that worked beautifully in March misfires in November — still reporting success, because reporting failure was never part of its design. The Atlas treats ownership as continuous: a named person, a review date that actually arrives, a register that gets read. Retired jobs should be retired in fact, not just left running in a forgotten corner with admin rights. If nobody can say what a task did last week, that is not an asset; it is a rumour with credentials, and it belongs on the same list as the fax machine nobody dares unplug. The Atlas’s answer is a register with review dates that arrive on a real calendar, owned by a real person, and an honest retirement path for jobs whose moment has passed. Estates shrink as well as grow, and a team that retires deliberately is a team whose remaining jobs you can actually trust.
Myth 5: AI agents will run themselves safely
The newest generation of autonomous helpers is genuinely impressive — AWS’s DevOps Agent reached its general release in March of 2026, and Microsoft’s SRE Agent is already at work on Azure — but “impressive” and “self-governing” are different claims. The Atlas’s guidance is almost boring in its good sense: admit agents the way any other automation is admitted. Credentials trimmed to the minimum. Read-only status as the starting condition, not a reward. And every proposed action passing the same guards a script would face. An agent with sweeping credentials and its own private pathway is not a colleague; it is an unsupervised locum with root access, and hospitals of all institutions know exactly how that story goes. The encouraging part is that nothing new needs inventing: the same register, the same tiers, the same guard tests that govern scripts will govern agents, provided the agents enter through the front door rather than arriving by personal favour from an enthusiastic vendor.
Myth 6: Fading manual skills are an acceptable trade
When automation absorbs a task, the human skill for doing it by hand decays — that is not moralising, it is observation, and it is why the Atlas says plainly that manual skills need rehearsing. The pilot community drills hand-flying for exactly this reason. In an operations context, the equivalent is simple: the team should still know how to do the automated task manually, because the day the automation is stopped — deliberately, via a kill switch, mid-incident — somebody has to. And the authority to stop things belongs to whoever is on call, without a permission-seeking round trip, with the explanation delivered afterwards. A team that cannot run the process by hand does not have a fallback. It has a hope.
The AutomationOps habit behind all six
Notice the pattern running underneath: every myth fails the same test — it assumes the unattended machinery will keep making sense on its own. The corrective habit is to treat that machinery the way good clinical teams treat protocols: named owners, clear contraindications, rehearsed exceptions, and someone authorised to stop. None of it requires an engineering degree; it requires the same professional instinct that already runs the ward. Concretely, the habit looks like this: rehearse changes before they land, refuse empty target lists, cap how much one run can touch, budget the failures and stop at the line, keep the switches drilled, and score the whole estate in hours returned, rescues needed and guard trips. The Atlas’s own worked example shows why the bother pays: a modest job running forty times a month returned 7.83 net hours monthly, paid for itself in 5.1 months, and held a 95% success rate — numbers a department head can defend at any budget meeting.
Technology’s role in cardiac care grows every year, and the teams who benefit most are the ones who stay sceptical in exactly these ways — a scepticism familiar from our coverage of hospitals using advanced heart monitoring technology, of cardiac ICU guidelines for safety and support, and of heart surgery costs and smart hospital comparison. Machines earn trust the way institutions do: with evidence, rehearsals, and someone accountable when the unexpected arrives.