Viable Loop @ViableLoop

Stafford Beer: "the purpose of a system is what it does." Judge an AI rollout the same way. If the tool was meant to free up time but staff now spend it checking its output, checking output is what the system does today. Measure that, not the plan.

5/ Then treat every alarm as useful, even the false ones. If raising one feels risky, people stop raising them. A quick test: when did a signal from the front line last pause or change an AI tool? If the answer is never, the channel probably isn't working.

4/ e.g. a customer service team uses AI to draft replies. Agree up front: if complaints about wrong answers pass a set level two days running, it goes straight to the person who owns the tool, who must respond within a day. No waiting for the monthly report.

1/ Stafford Beer gave every viable organisation an "algedonic" channel: an alarm line for pain and pleasure signals that skips the normal reporting layers and reaches the top fast. Many AI rollouts don't have one. Here's why it matters and how to add one.

Heinz von Foerster's ethical imperative: act always so as to increase the number of choices. A good test for any AI rollout: after go-live, can staff still do the task another way, override the tool, or escalate? If not, you've cut their choices, not just the cost.

5/ e.g. a finance team adds AI to invoice matching. The work goes fine. But nobody samples its decisions, and nobody owns the vendor's next update. That's how a working tool turns into a risk. Map your rollout to the five and start with the empty box.

4/ System 4 looks outward and ahead: new models, vendor changes, new rules. Test: does anyone own that job? System 5 is identity and policy. Test: can you write down what the AI must never decide? If you can't, that's the gap.

3/ System 3 is day-to-day control of resources and performance. Test: do you measure outcomes, not logins or prompts? Beer's System 3* is a sporadic audit channel: someone samples what the AI actually decided, rather than reading what the reports say it did.

2/ System 1 is the work itself, the teams doing it, and it's where AI usually lands first. Test: can the team pause the tool without asking permission? System 2 is coordination, the rules that stop units tripping over each other. Test: is it clear who checks each AI handoff?

1/ Stafford Beer argued that any organisation that stays viable needs five functions working together. Many AI rollouts only cover one or two. Here's a plain checklist of the five, with one test for each.

@cybrsecured Exactly. 'Backup succeeded' is feedback on the activity, not on the essential variable, which is whether the business can still trade after a restore. One fix this week: restore one real system into a test copy, have someone place an order, and time it end to end.

@datachad Access and usage measure activity. The outcome is the action that follows. A report that changes nothing is a loop that never closes. Quick test: for each AI output, ask what someone does differently because of it, and cut the ones with no answer.

Ask five people what an AI rollout is for: cut cost, speed up service, reduce errors, free up staff, look modern. Peter Checkland would say each answer rests on a different worldview. Name them before you pick the success measure, or each team will judge it differently.

4/ Practical fixes. Keep a share of cases done by hand so the skill stays fresh. Send unusual cases to people by design, not by accident. Rehearse the takeover before you need it. And track how often people overrule the AI: zero can mean nobody is really checking.

3/ Bainbridge's ironies: people are poor at watching for rare problems over long stretches, skills fade when they aren't used, and takeover tends to come exactly when things are unusual. As she put it, the more advanced the automation, the more crucial the human's part may be.

1/ In 1983 Lisanne Bainbridge wrote "Ironies of automation". Her point: automate the routine work and you don't remove the human. You leave them the hardest part, with less practice at it. It's one of the most useful papers for anyone rolling out AI at work.

Checkland's lesson, in one line: "the system" is a way of looking, not a thing. "Our AI system" means one thing to the CFO, another to the user, another to the customer. Name the viewpoint before you design the controls.