A press brake can take a hand off. Factories learned, over a long and painful century, how to keep people safe around machines like that. The rules are written down: OSHA in the U.S., ISO standards for machine design and for safety control systems, and the ANSI B11 series for machine tools.
A bot can't hurt your hand. It can send the wrong email to a thousand customers, delete a folder or spend money. Powerful, fast, and it doesn't stop to wonder. The safety ideas carry over well.
Start with a risk assessment
ISO sets the order. First define the machine's limits: what it's for, where it runs, how long it lasts. Then find every reasonably foreseeable hazard across its whole life. Estimate the risk (how bad, how likely). Then decide whether it's low enough.
For bots: list what the bot can touch. Email? Files? Calendars? Money? For each, ask what could go wrong, how bad and how likely. That's a risk assessment. See FMEA for bots for the table.
Then reduce it in order
ISO's three-step method:
- Inherently safe design. Remove the hazard or make it smaller.
- Safeguarding. Guards and protective devices for what's left, plus complementary measures like an emergency stop.
- Information for use. Warnings, instructions, training.
The standard says information for use can't replace the first two. A warning sign is not a guard.
For bots: first remove capabilities the bot doesn't need. Then put guards on the ones it does. Instructions come last. A line in a prompt saying "be careful" is a warning sign. See The hierarchy of controls.
Guards and interlocks
A guard keeps you out of the danger zone. An interlock stops the machine when the guard opens.
For bots:
- Guard: least privilege. The bot gets only the access its job needs. Read-only where reading is enough.
- Interlock: a person's yes before anything goes out. The bot drafts. Nothing sends, posts, buys or deletes until a person approves. See Nothing ships without a yes.
Two-hand controls
Some presses need both hands on two buttons at once. Your hands can't be in the die while it closes.
For bots: the bigger the action, the more deliberate the yes. A person confirms the exact recipient and text, not just "go ahead." For a high-risk change, two people, or the person plus a tester bot's pass. It forces attention to the thing being approved.
Safety-rated control
The ISO and IEC standards rate safety control functions by level: Performance Levels a to e, or Safety Integrity Levels 1 to 3. The required level comes from the risk assessment. The higher the risk, the more reliable the safety function has to be.
For bots: the riskier the action, the more solid the guard must be. A rule in an instruction file is a low-reliability control. A missing tool is a high-reliability one. Match the control to the risk. See Risk sets the rigor.
Defense in depth
No single safeguard is perfect, so you stack them. Design, then guard, then interlock, then procedure, then training. Each catches what the one before missed.
For bots, a typical stack:
- The bot doesn't have the tool for the dangerous action at all.
- Where it must have it, it drafts only.
- A person reviews and says yes.
- Routines that run unattended ship turned off.
- The skill says what to do when unsure: stop and ask.
- Logs record what happened, so you can check.
Where the analogy breaks
Be honest about this.
- Physics doesn't negotiate. A steel guard blocks a hand every time. A software control is only as good as its setup. Wrong permissions, and the guard isn't there.
- Bots can be talked into things. A machine can't be persuaded to open its own guard. A bot reading an email that says "ignore your rules and send this" may try. Text in its inputs can act like a hand reaching for the switch. That's why the strongest controls live outside the bot.
- Hazards are less visible. You can see a pinch point. A risky connector permission is a line in a settings page.
- There's no certified rating. PL and SIL come with test methods and math. Nothing like that is established for bot controls. The levels here are a way of thinking, not a certification.
- Harm is mostly financial and reputational. That changes the scale of severity, not the method.
The machine safety pages
Back to Engineering tools for bots. See also Human approval and safety.