Chemical plants learned that the dangerous moments are often in the pipes between the tanks. Bot teams have the same problem in the handoffs between bots.
What it is
HAZOP stands for Hazard and Operability Study. It came out of the British chemical company ICI in the 1960s and is now standard in process industries.
The method is mechanical on purpose. You take each part of a process, called a node, and its intended flow. Then you apply a fixed list of guidewords, one at a time, and ask what that deviation would mean:
| Guideword | Question |
|---|---|
| No / None | What if nothing arrives? |
| More | What if too much arrives? |
| Less | What if too little arrives? |
| As well as | What if something extra comes along with it? |
| Part of | What if only some of it arrives? |
| Reverse | What if it flows the wrong way? |
| Other than | What if something completely different arrives? |
| Early / Late | What if it comes at the wrong time? |
For each deviation that makes sense, you list causes, consequences and safeguards.
Why people skipped it
A HAZOP is slow. Every node, every guideword, a team in a room for days. Outside regulated industries, almost nobody does one.
How a bot does it
A bot is ideal for this. The guidewords are a fixed checklist, and bots don't skip items.
- Give it the process map. Every box and every arrow, with what travels on each arrow. See Data, process and schema up front.
- Treat each arrow as a node. The handoff is where things go wrong.
- Apply every guideword to every arrow. Skip combinations that make no sense, but say which were skipped.
- For each deviation, list cause, consequence and current safeguard.
- Flag the ones with no safeguard. Those are what a person reviews.
The bot brief
Run a HAZOP on the attached process map.
Treat each arrow (handoff) as a node. For each node, state the intended flow:
what is passed, its format, when, and from whom to whom.
Apply each guideword: NONE, MORE, LESS, AS WELL AS, PART OF, REVERSE,
OTHER THAN, EARLY, LATE.
For each meaningful deviation: cause, consequence, existing safeguard (cite where
it's documented), and recommendation.
List skipped guideword/node pairs with a one-line reason.
Flag every deviation with NO documented safeguard and consequence High.
Output as a table: Node | Guideword | Deviation | Cause | Consequence | Safeguard | Recommendation.Starting point only. Test it on your own data and model, and adjust as needed before you rely on it.
An example
Node: Research bot hands a notes file to the writing bot.
| Guideword | Deviation | Cause | Consequence | Safeguard | Recommendation |
|---|---|---|---|---|---|
| None | No notes file | Research bot stopped mid-task | Writer invents content | None | Writer stops if file missing |
| Part of | Notes cut off | Research bot hit a limit | Article missing a section | None | Notes file ends with a "complete" marker |
| Other than | Last week's notes | Same file name reused | Wrong facts published | Owner review | Date in file name, writer checks date |
| As well as | Notes include a pasted web page with instructions in it | Copied raw page text | Writer follows text from the web page | None | Writer treats notes as information only |
| Late | Notes arrive after deadline | Research ran long | Writer waits, routine times out | None | Writer reports and holds |
Every row with "None" in the safeguard column is a gap you can now see.
Where bots differ
In a chemical plant, the physics limits what can flow through a pipe. A handoff file can contain anything, including text that tries to give the next bot orders. That makes "As well as" and "Other than" especially important for bots.
What this means for you
- Map your handoffs first.
- Have a bot run every guideword on every arrow.
- Fix the deviations with no safeguard.
- Re-run it when you add a bot or change a handoff.