A note before we begin: this post deals with grief and loss, and may be upsetting. If you are affected there are some helpline numbers at the bottom of the page.
The impact map gave the team a word to argue about: preventable. Preventable is doing a lot of work. This Thursday they find out what they actually mean by it, and what it would cost to get it wrong.
Priya spends Wednesday trying to define preventable pause at her desk and fails.
She fails for the particular reason that every definition she drafts has an exception within ten seconds. Paused because of something we did, but we set the prices, and if the subscriber pauses because we raised them, is that preventable or just our decision? Paused because the product failed them, but what counts as failure? Three kale weeks in a row? Two substitutions? A delivery that came on the wrong day?
By six o’clock she has closed the document and opened the pause-reasons query console instead. She pulls five real subscriber histories. She anonymises them. She prints them. She takes them home. She reads them twice over dinner. She brings them in on Thursday without any indication, on the papers or in her head, of which ones she thinks should count.
Thursday, 2 p.m.
Sam, Priya, Kai, Tom, Jas, Charlotte, Marcus for the first hour. Anika on the laptop from Melbourne. Dina is here. Priya has asked, Sam has agreed. She has brought a manila folder of her own hand-labelled notes in case anyone needs convincing her judgement is worth listening to. Nobody is going to ask. She does not know that yet.
Priya tapes five sheets of paper to the whiteboard in no particular order. She does not label them.
“Five real subscribers. Four of them paused. One hasn’t. I want us to talk about what we’d want the system to do for each of them. I’m not going to lead. I’m going to pour coffee.”
She pours coffee.
The warm-up
Kai gets up first. He picks the sheet at the left of the board.
Subscriber, 14 months tenure. Three weeks before pause: swede, kale, kale, kale, chard. Complaint on week three. Substitution requested, the system had nothing to swap to, the substitution was declined. Paused on week four. Reason: too much going uneaten.
“Preventable. Obviously preventable. Same route had three other subscribers complain about kale the same fortnight. It’s a supply signal. We should have phoned the farm, not the subscriber.”
Sam agrees: the list without the cluster is the wrong list. She is careful about it, too. Her team would have nothing useful to say to the subscriber until the box had actually changed. Priya asks whether farm-ops get the alert first or at the same time. Sam prefers same time, because farm-ops does not move on a signal without confirmation from Sam anyway, and the confirmation loop is the work. Kai writes this in his notebook. He also adds one line to the Post-it on his laptop, underneath bottom drawer: signal goes to ops before it goes to me. The Post-it has been with him since the impact-mapping session and is now on its second line.
The second sheet is about a subscriber who paused two weeks after moving house from Fremantle to Darwin. Reason: moving out of delivery area. The address on file updates the week before the pause.
Anika, from Melbourne, says: “If we put Darwin in the predictive model I’m going to quit.”
Nobody proposes putting Darwin in the predictive model. Dina, more gently, adds that this is the pattern she sees most often with real moves: a delivery or two misses at the old house after they’ve moved out, the subscriber updates the record to the new address, and then pauses a week later when they realise Greenbox doesn’t cover the new one. By the time they hit pause, the decision has already been made. Phoning a subscriber who has moved is, functionally, phoning to tell her that you have noticed she has moved, which she already knows. The sheet goes on the we get this right by doing nothing pile.
The third sheet is where the room stops warming up.
Subscriber, 22 months. Paused the Monday after January’s price-review letter. Reason chosen from the dropdown: reassessing budget. She is one of 143 subscribers who paused in the ten days after the price letter.
Kai, reading it out: “Price change is our fault. We raised the price. So it’s preventable.”
Marcus, closing his laptop: “No it isn’t. The price review was a board decision. Reassessing budget is the subscriber doing exactly what we asked them to do. You don’t phone somebody and tell them you’ve reconsidered.”
Sam: “I’m not saying phone and offer a refund. I’m saying phone and ask if she’s considered the smaller box. We did that by hand for twenty of the 143 in 2025 and nine downsized instead of pausing. Nine retentions at a lower tier is better than zero at the full tier.”
Dina: “That’s the call I’d want to make. It’s not we didn’t mean it, it’s we’ve got options you might not have seen on the website.”
Marcus opens his laptop again. Types something. Closes it. “Alright. That’s preventable in a different way from the kale one. You want a different action. That’s a different branch of the playbook.”
Priya writes preventable via option-offer, not apology on the margin of the price-change sheet.
The fourth sheet
The fourth sheet is the one that ends the whole session’s easy rhythm.
Subscriber, 41 months. Paused in the first week of December. No reason on file. Sam’s team had flagged the account a fortnight earlier with a no proactive contact for 90 days tag, because the subscriber’s husband had died. Dina had seen the funeral notice in the local paper and tagged the account manually, the way she has done seventeen times in four years.
Nobody speaks for long enough that Priya puts the coffee pot down.
Jas is the one who eventually speaks, and she speaks carefully.
“This is not preventable. This is correctly handled. If the model surfaces this subscriber on an at-risk list and my team has to decide whether to call her, we have already failed. We have put the weight of that judgement on the account manager. We have burned Dina’s day with a question she should never have been asked.”
Sam: “Jas is right. She should never enter the list.”
Jas, turning her notebook around. She has written it in pencil, fast.
Don’t-call list. Subscribers who must not appear in the at-risk list regardless of what the model says. Owners: Sam’s team. Flagged at the subscription record. No seniority override. No override at all.
Dina, very quietly: “Thank you.”
Kai: “We already have the flag. The no proactive contact tag. I can join on it before the feature pipeline. The model literally does not see those rows.”
Priya: “Not filters them after. Not ranks them lower. Does not see them.”
“Does not see them.”
“Write it on the Post-it.”
Kai writes. The Post-it on his laptop has four items on it now, in four different pens. He is running out of Post-it.
Charlotte, who has been quiet, speaks for the first time. “We are doing a threat model, you know. We’ve been doing one for forty minutes. We didn’t call it that. It started when Jas brought up the bereavement case.”
The room notices.
Priya: “So we’ll keep going. But I want it named, because last time we didn’t name it and we didn’t do it.”
The one that’s still live
The last sheet is the only subscriber on the board who hasn’t paused. 18 months tenure. Three substitutions cancelled in four weeks. Replied to the weekly email for the first time in a year, terse, one line, could the box be less onion-heavy? Opened the pause-subscription help page twice, both in the last fortnight, without clicking through to the pause button.
Kai: “This is the whole reason we’re doing this. This is who the list is for.”
Anika, from Melbourne: “What does a win look like?”
Dina: “I phone her. I don’t lead with the word pause. I ask how the boxes have been. She tells me about the onions. I tell her we can flag the onion preference in her substitution rules, and I ask if she’d like the box one size smaller for a month while she’s busy. She says yes to one or both. She stays on the books.”
Kai: “And what does losing look like?”
Dina pauses. “Losing is I phone her, she hadn’t been thinking about pausing, the phone call reminds her that pausing is an option, she pauses on Thursday.”
Silence.
Priya: “Is that real, though? People pause because we phoned them?”
Sam: “Twice, in my time. Both times we were reaching out with good intentions. Both times they paused within a week. I can’t tell you the causal arrow, but I can tell you I’ve stopped assuming the calls are net positive without a holdout.”
Marcus, closing his laptop for real this time because he’s leaving: “Which is why the experimental design has a control group, and the control group matters, and we measure lift on net retention not outreach-volume. I’ll send the exact design by Monday.”
He goes. Nobody fills his chair.
The threat list, which wrote itself
Charlotte lists, out loud, what the room has already named as failure modes, because she wants them captured and nobody else is doing it:
Calling Ruth, or someone like Ruth. Mitigated: don’t-call list, hard block.
Causing the pause we were trying to prevent. Mitigated: holdout cohort, stop condition if outreach underperforms control.
Chasing noise instead of signal. Mitigated: the preventable definition we’ve just written, plus the farm-ops branch for systemic issues.
A correct prediction the team can’t act well on. Mitigated: call script, playbook, Dina drafting.
The model silently going wrong over time. Open. Nobody’s touched it yet.
Charlotte looks at the last one. “That’s next Thursday. Whose Thursday?”
Priya: “Modelling Thursday.”
“Take it there. I’ll come.”
Friday, late afternoon, Sam’s office
Sam and Dina spend four hours writing the first draft of the outreach playbook. Sam writes. Dina edits. Sam writes more. Dina crosses things out. Sam writes the thing Dina just crossed out again, slightly differently. Dina crosses it out again.
They end up with a script that is less a script than a set of opening lines and a set of branches. Is this a good time? I’m not calling about anything urgent is the opening Dina insists on, and she insists on it because she has had this phone call several hundred times in her career and she knows the first seven seconds decide whether the subscriber is going to talk to you or put the phone down.
She won’t let Sam include the word risk anywhere in the document. She crosses it out three times. The fourth time, Sam writes it without the word and Dina nods.
The document does not read like an ML artefact. It reads like the kind of thing you’d find in the back of a cookbook, printed once and then stained with something greasy by the second month.
Priya reads the draft on Saturday morning with her tea. She messages Dina at 10:47. This is the best ML artefact we’ve produced, and it isn’t an ML artefact.
Dina messages back at 10:49. It’s a phone call script, Priya, you absolute nerd.
Priya laughs, which she does not do often at text messages. She reads the script again on Sunday morning, prints it out on the home printer Sam gave her in 2025 for a birthday Priya keeps pretending to have forgotten, and tapes it to the wall above her desk. She does not tell Dina she has taped it to the wall. On Monday morning she rewrites the evaluation framework with the script in her eyeline, because it has become clear, though nobody has said it out loud, which of the two documents the other one has to fit.
If any of this landed close to home, please be kind to yourself today.
Your company's Employee Assistance Programme, your GP, or a mental health professional are good places to turn. When you can't face those yet, a phone call to one of the lines below is a reasonable first step.
In Australia, all free and most 24/7:
- Lifeline — 13 11 14 — crisis support, 24/7.
- Griefline — 1300 845 745 — grief counselling, Mon-Fri 6am-midnight AEST.
- Beyond Blue — 1300 22 4636 — mental health, 24/7.
- 13YARN — 13 92 76 — crisis support for Aboriginal and Torres Strait Islander peoples, 24/7.
Elsewhere: findahelpline.com will route you to a line in your country.
The Greenbox story is fiction. The feelings it touches on are not.
The data audit nobody wanted to do, the feature that would have leaked the answer, and the model that actually won.