Clinical teams already know that discipline beats intention. A grading scale turns “the patient seems tired” into something a colleague in another city can act on; a fixed escalation path beats heroics at 3 a.m. Operations teams arrived at the same conclusion through a different door, and the practice they built around it — ChatOps, shared chat as the operating floor, with bots announcing events and carrying out only vetted commands — succeeds or fails on the small habits listed here. The habits below are drawn from the ChatOps reference page at the Ops Atlas, compressed to what a working team can actually enforce. None of them requires a big budget; all of them require someone willing to say no.
Why Clinical Readers Should Care About ChatOps
Hospitals increasingly run their digital operations — monitoring, escalations, follow-up scheduling — the way modern engineering teams do, and the vocabulary is crossing over. A ward that tracks functional status formally, the way the NYHA functional classification turns breathlessness into a shared scale, will recognise every habit below: define the grades, publish them, and let the grades make the argument. Here are the eight.
Eight ChatOps Habits That Earn Their Keep
At a glance, the eight are:
- Grade every command before it exists.
- Make the dangerous ones two-person.
- Match channels by identifier, never by name.
- Keep free text away from the shell.
- Treat refusals as telemetry.
- Report changes, not states.
- Count the cost of notifications.
- Let AI propose while humans approve.
1. Grade every command before it exists
Sort candidate actions into four tiers at design time: inform, read, reversible write, high-impact. A command that merely informs carries almost no risk; anything that writes needs a specific role, a locked-down channel and an audit entry bound to it before launch. Retrofitting risk grades after a bot is live is twice the work and half as honest.
2. Make the dangerous ones two-person
High-impact actions get a second approver — always. One person asks, another signs off, and the platform checks that the two identities differ. Dual authorisation is a standard control in formal security frameworks for good reason; chat makes it easy because the approval can happen in seconds, in the open.
3. Match channels by identifier, never by name
Channels get renamed, deleted and recreated. A bot that routes by display name will, eventually, post something sensitive into a channel that merely inherited the old name. Route by the immutable identifier and the rename stops being a security event. The same logic covers users: authorise the identifier your platform cryptographically vouches for, never the friendly label above a message, because labels are editable by anyone and vouched identifiers are not.
4. Keep free text away from the shell
The moment a bot executes whatever someone types, you have handed every chatter a loaded console pointed at production. Commands should match fixed, pattern-checked forms; anything else is refused, publicly. Refusals are not rudeness — they are the system keeping the allow-list meaningful.
5. Treat refusals as telemetry
Review the refusals weekly. A spike in refused commands tells you one of three things: the allow-list has a gap, someone is probing, or training has failed somewhere. All three deserve attention, and none of them shows up in command-count dashboards.
6. Report changes, not states
Mature incident bots announce what changed — “acknowledged by Priya”, “escalated to secondary” — instead of re-printing the whole world on a timer. Changes are scannable; states are wallpaper. The same bots fail loudly when a step cannot complete, with a fallback path, rather than failing silently and letting a gap hide in the scrollback.
7. Count the cost of notifications
Ungrouped alerting is arithmetic, not opinion: a stream of a few hundred unfiltered weekly alerts, spread across a twelve-person team at ten seconds apiece, burns through some thirteen hours — most of a working day — every week. Group related alerts, suppress the duplicates, and aim for a channel where a large share of posts are things a reader can act on. Track that share as a number if you can — some teams measure what fraction of notifications prompted any action — because a ratio you can see is a ratio someone will improve.
8. When AI joins, it proposes and humans approve
Assistants can draft the response, pick the runbook, summarise the incident — but the approval stays human, permanently, and both the prompts and the tool calls get logged. Prompt-injection research published in 2024 showed how chat-linked AI features can be manipulated through ordinary-looking content; that finding settled the argument. Let the machine suggest. Never let it authorise.
Two Traps That Undo Otherwise Careful Teams
The first trap is approving with emoji. A thumbs-up feels like consent and photographs like evidence, but reactions carry no identity guarantee worth trusting and no workflow behind them; a formal approval action, captured by the gateway with the approver’s identifier, is barely slower and infinitely stronger. The second trap is keeping secrets in the wrong place. Webhooks and tokens pasted into channel topics or configuration files become the quiet master keys of the whole setup — anyone who reads the room holds the keys. Credentials belong in a proper secret store, rotated on a schedule nobody has to remember personally, and a leaked webhook should be treated exactly like a leaked password: revoked within minutes, not discussed within days.
Keeping the Same Standard Over Time
Habits decay without a cadence. The teams that last pair the eight items above with a quarterly scope audit — every allow-list entry re-justified or removed — and a simple honesty metric or two: how much of the change volume is actually covered by the audited pathway, and what fraction of notifications led to action. Publish them somewhere the whole team walks past daily. Clinical follow-up works on exactly this logic, as any reader of a proper cardiac follow-up protocol already knows: the plan is the easy part, and the schedule of reviews is what keeps the plan true. Run your channels the same way, and the tooling becomes almost boring — which, in operations, is the sound of things going right.
