Product boxes love the word “automated”. Automated backups, automated security scans, automated reordering, automated everything — and yet almost nobody who buys these products asks the follow-up question that matters: who is accountable when the automation does something wrong at 3 a.m.? There is now a name for the discipline that asks it. AutomationOps, as the Atlas defines it, is the craft of running operational automation the way you would run any production system: picking what deserves to be automated, making each repeat safe, capping the harm one run can cause, and staying accountable until the job retires. This explainer keeps the language plain and the definitions precise, because this is one topic where sloppy words cost real money.

An unattended job is a colleague who never sleeps, never complains, and never asks whether this month’s instructions still make sense. Governance is how you manage such a colleague.

What AutomationOps actually covers

The Atlas draws the boundary around run-time actions: things that act on live systems while nobody supervises. Restarting a crashed service, cleaning out temp directories, failing over between systems, rotating credentials, pushing one configuration to a hundred machines — all of that is in scope. What stays out is just as clearly marked. Writing infrastructure templates is InfraOps territory. Multi-step workflow engines belong to WorkflowOps. Commands fired from chat channels are ChatOps. Network device changes sit with NetworkOps, and measuring how much repetitive work exists belongs to SRE. The relationships are feeder, not competing: those neighbouring practices spot candidates and send them across, and once a job has been registered and guarded here, it flows back as a controlled, owned asset. If an artefact touches a live system unattended, it belongs on this register — whatever logo sits on the tool that triggered it.

AutomationOps: a label with a crowded past

Buyers should know the name is contested, because vendors have used it for other things. UiPath put a product of this name into general availability in April 2021, focused on orchestrating its robot fleet. Robocorp taught a methodology under the same banner; that material was archived in 2023. The Atlas borrows the word for something wider and vendor-neutral: an editorial label for the whole practice of governing operational automation, whatever tool performs it. When a product page claims to “do AutomationOps”, the honest reading is that it automates some slice of the discipline — a scheduler, an approval gate, a logging pane. That can still be useful. It just is not the whole system, and a reviewer’s job is to find out which slice. For anyone grading products, the practical consequence is simple: ask which definition a vendor means before scoring the claim.

The five controls, in precise terms

Across incident write-ups and operations guides, the same five safeguards keep recurring. Each can be stated in a single sentence, yet together they cover most of the failure modes on record:

• Dry run. A rehearsal mode that predicts what a change would do before doing it. Caveat worth printing: in one widely deployed configuration tool, check mode skips plain command and shell probes unless a flag is set, so a spotless rehearsal can hide exactly the step that would have failed.
• Target validation. A run handed an empty target list refuses to start — empty never means all. The rule earned its fame through Google’s Diskerase service, which erased data after an unrepaired empty entry and produced a post-mortem that changed industry practice.
• Blast-radius cap. Work is sliced into batches, canary first; tools expose this as ordered serial processing with a maximum failure percentage per batch, so a bad run touches four machines instead of four hundred.
• Failure budget. A run halts itself once errors cross a set count, rather than retrying its way through the whole fleet on a bad night.
• Kill switch. A pause control per automation plus one global halt, drilled twice a year for the higher risk tiers, because a switch nobody has practised pulling is decoration.

Risk tiers: consequence sets the treatment

The Atlas proposes four grades, drawn strictly by what a job can break. Tier 0 is read-only work — reports, alerts — which can irritate but not destroy. Tier 1 covers reversible changes to a single target. Tier 2 is reversible change across many targets — think a configuration push to a whole fleet — and earns dry runs, batching and health checks as standard equipment. Most scheduled jobs in a typical estate live here, which is why tier 2 hygiene decides whether the overall estate is trustworthy. Tier 3 is for the irreversible and the data-destroying, where the requirements jump: a second person must approve, deletions become soft, and rollouts proceed in stages. That top grade maps onto NIST’s dual-authorisation family — AC-3(2) plus CM-5(4) — which tells you the idea has serious institutional pedigree. For a buyer, the tier of a vendor’s own automation features is a fair question to ask before purchase.

Idempotency and request IDs

One property separates calm automation from nerve-shredding automation: whether running the same job twice is harmless. Well-built jobs describe a desired state — “ensure running”, not “start” — so a repeat finds nothing to do. Actions that cannot work that way, because they create, charge or send, take a caller-supplied request ID recorded atomically, which makes the duplicate harmless too. Google’s own repair jobs could run every fifteen minutes without fear of damage precisely because of this property, and the Molecule testing framework keeps an idempotence check in its default sequence so regressions surface early. When a vendor cannot say whether their automated actions are idempotent, that silence is a finding.

How to interrogate an automation claim

Reviews of software usually benchmark speed and price; for automation features, the more useful test is a short interrogation. Ask these, in this order, and score answers honestly:

1. Show me the rehearsal mode, and tell me which probes it skips.
2. What happens when the target list comes back empty?
3. How many targets can one run touch before it pauses?
4. After how many failures does a run halt itself?
5. Where is the kill switch, and who is authorised to pull it?
6. What identity does the automation use, and how narrow is it?
7. Who is named as owner, and when is the next review date?

Three vague answers out of seven is not a mixed review — it is a pattern, and patterns travel. One encouraging signal to look for in documentation: vendors who publish their guard trip rate, showing how often their own safety brakes engage, are demonstrating the controls work rather than merely claiming they exist.

The math that decides whether it was worth it

Finally, the question a reviewer rarely asks and a buyer always should: what does this automation return? The Atlas’s arithmetic is refreshingly plain. Take the monthly run count, multiply by what a human used to spend in minutes, knock off the time still consumed by checking and upkeep, and divide by sixty. The remainder is net hours returned. In the worked example the Atlas uses, a job running forty times a month nets 7.83 hours, breaks even at 5.1 months against its build cost, and holds a 95% run success rate. Those three numbers say more about a product than any feature grid, because they price in the maintenance tail that marketing never mentions. Notice, too, what the formula refuses to do: it never credits saved time before subtracting upkeep. A job that banks thirty minutes per run but demands an hour of fixing each month can easily net zero, which is why honest arithmetic always includes the tail.

The habit this discipline builds — asking who owns the unattended machinery, what stops it, and what it returns — transfers neatly to every purchase decision. It is the same scepticism our guides recommend when finding reliable and unbiased product reviews, the same instinct that serves you well when finding and booking rental vehicles, and the same questioning that pays off when comparing motorcycle rental options across China: demand to know what happens when the promise meets a bad day. Automation is staff, not magic — and staff, sensibly, get references.