Operations: What We Actually Do About It

It is Monday, five past eight, and Jade has the live list open for the first time. Sixteen accounts, ranked, each with a score and a stated reason beside it. Ten modules of work sit behind this screen — a budget split, a business case, a requirements document, sixty hand-labelled accounts, a rebuilt data pipeline, a model, a governance review, a deployed system — and it is worth saying plainly what all of that has achieved so far.

Nothing. Not one customer has been saved. The model has narrowed the field from 400 accounts to sixteen names, which is genuinely valuable and is not the same as value. The saving happens, or does not, in what a person does between now and Friday. This is the step most AI projects treat as an afterthought — the bit that happens after the demo, once the interesting people have moved on — and it is where most of their value dies.

What lands on Jade’s desk

The list itself comes from Kofi in module 10 — running nightly, landing in the tool her team already work in, with the cause field mandatory rather than optional. That field exists because Maya insisted in module 5 that a score without a reason is not an alert, because Sofia wrote it into the requirements in module 4, and because Elena made human review a condition of approval in module 9. Three people, at three different desks, months apart, all defending the same small thing. It is the difference between a rep opening a call with “your risk score is 82%” and opening it with “your export has been broken since March”.

Two of Daniel’s funding decisions from module 2 are sitting in this Monday too, and neither is visible on the screen. The $50k of training money means Jade’s six reps have already been taught to read a reason and run a playbook, so they work at the top of their capacity range rather than the bottom — an untrained team spends the first hour of every intervention deciding what to say. And the $100k he put into product engineering means that when Corvex Manufacturing’s broken export comes up this morning, there is an actual fix on an actual plan to escalate to, rather than a sympathetic noise.

The length of the list is Priya’s decision from module 8, made in a room with Maya in it: a threshold tuned to produce about sixteen alerts a week, because six people can run ten to sixteen prepared interventions in a week and a list nobody can finish is just guilt with a dashboard. Jade owns the supply side of that arithmetic. She did not choose the demand.

What a CS operations manager actually does

Customer success operations is the least glamorous job in this course and the only one where the money is actually saved. The role owns the machinery behind the relationships: which accounts get attention this week, what a good intervention looks like, how it gets recorded, and what the team stops doing to make room. Everyone upstream produces artefacts. Jade produces phone calls.

The craft is in refusing two obvious failures. The first is treating the score as an instruction — sixteen alerts, sixteen calls, all broadly the same call. The second is letting the list become optional, which is what happens within a fortnight if the calls it produces embarrass the people making them.

The software on Jade’s desk

The first screen carries the lesson. The at-risk list appears inside Gainsight, the platform Jade’s team already opens forty times a day, rather than in a new tool with its own login. Kofi built it that way because Sofia specified it that way, and both were right: a daily list living somewhere else is a habit that has to be created, and habits that have to be created mostly are not.

Salesforce supplies the pre-call context — contract value, history, what was promised at the last renewal — because opening a conversation without it is how a rep ends up offering a discount to a customer whose problem is a broken export. The intervention itself happens in Outlook and on the phone, which is worth stating plainly in a course about AI: the entire apparatus of the previous ten modules exists to make one email and one call better targeted. Then the outcome is recorded back in Gainsight with a single click. That last, dullest step is what makes the next module possible at all.

The decisions Jade makes on Monday morning

Choosing the response from the cause, not the score

The instinct everyone brings to a ranked list is that the response should scale with the number: big score, big gesture. It is wrong, and watching Jade work the queue shows why. The score decides who. Only the cause decides what.

Bramford Logistics sits at the top, and deserves to. The champion who bought the product left three weeks ago, weekly users have fallen from 61 to 24 since June, the renewal is nine weeks away and the replacement is reviewing all software spend. The usage collapse is the symptom; the cause is that the one person inside Bramford who cared has gone. Jade’s response is an executive relationship call — Northwind’s commercial director, their new decision-maker, a conversation about what the next year looks like. It is the most expensive action on her board and she spends it here, because a new decision-maker needs a relationship, not a webinar. The tempting cheap alternative, a training invitation aimed at the falling login numbers, answers the symptom and leaves the cause exactly where it was.

Corvex Manufacturing scores high for the opposite reason: they use the product constantly, because their production line depends on it, and they have raised 23 tickets about the same broken export since March. Their last survey was 3/10. The response is a technical escalation to product with a date attached — and thanks to Daniel’s $100k, there is a real fix to attach a date to. Note what Jade does not do. A discount here would be an insult: Corvex have not said the product is expensive, they have said it does not work, and offering money off is an admission that nobody read the tickets.

Juniper Health has failed two card payments and used 12 of 50 seats all year. They are not disengaged; they are paying for a team that was reorganised away. The right response is an honest conversation about a smaller contract, and it is the decision that takes the most nerve, because Jade is proposing to reduce revenue on purpose. The alternative — a discount that keeps the fifty seats on paper — is worse in a way that only shows up later. It papers over the mismatch, and mismatches resurface. Next year the same customer is still paying for thirty-eight seats they do not use, the discount has expired, and now they leave altogether. The honest smaller contract loses a slice of $27k once. The papered-over one loses all of it, twelve months later, with a customer who has learnt that the price was negotiable all along.

Aldgate Media never finished onboarding: two of three setup sessions cancelled, and the workflow module they bought the product for was never switched on. The response is a training and onboarding session, not a sales call. There is nothing to sell — they have already bought the thing. What they need is an hour of somebody’s time and a room of their people.

Fenwick Retail Group has a new CFO demanding 30% off or the contract goes out to tender, with the renewal in six weeks. Three teams use the product steadily and the survey says 7/10, “good tool, feels expensive”. This is a commercial negotiation wearing a churn score, and the response is a contract conversation: what could Northwind trade a concession for — a longer term, a seat commitment, different payment terms — rather than a panic discount granted for nothing. Give 30% away in a phone call and you have not saved an account; you have published a price list to every customer who hires a new CFO.

Deciding to leave three accounts alone

This is the decision that separates an operation from a call centre, and it is the hardest to defend in a meeting, because doing nothing looks identical to not getting round to it.

Hale & Porter LLP are down 45% in August, and were down in August last year, and the August before. Half the firm is at court recess. They renewed early twice without negotiating. Jade’s response is no action, and it is a decision rather than laziness: ringing a happy customer to save them spends a call and, worse, spends credibility. There is no version of that conversation that improves the relationship.

Thornbury County are 22 months from renewal in year one of a three-year contract signed through formal procurement. Their low usage is a real problem and Jade writes it into next year’s account plan, where it belongs. It is not a problem she can solve this quarter with a phone call, and putting it at the top of a list every morning does not fix the adoption — it consumes the slot Bramford needed.

Osprey Analytics are thriving, have built their own templates and have asked about fifteen more seats. They are not a retention problem at all; a call about how to keep them would simply baffle them. Jade routes the seat question to their account manager, which costs her team nothing this week.

Deciding that an unlogged intervention did not happen

Jade’s third decision is the one her reps push back on, in Slack, every week: do I really have to log this? The answer is yes, and the reason is worth feeling rather than just knowing.

In module 12, Nadia has to answer whether any of this worked. To do that she needs to know exactly which accounts were contacted, with which response, on what date, and which were left alone. If half the interventions happen in people’s diaries and never reach the record, she is not comparing a treated group to an untreated one — she is comparing a group she cannot identify to a group she cannot define. That is measuring fog. The retention number will move, because retention numbers always move, and nobody will be able to say whether it was the programme, the export fix, or the season. Six months of work then resolves into an argument between confident people.

This is why Kofi’s one-click logging from module 10 matters more than it sounds. The difference between a fifteen-second log and a five-minute form is the difference between a complete record and a half-empty one, and a half-empty record is worse than none, because it looks complete.

Writing the playbook down rather than improvising the week

The last decision is to turn all of the above into something repeatable: cause mapped to response, in writing, with the capacity rules attached. Sponsor loss goes to executive outreach. Product defect goes to escalation with a date. Over-licensing goes to a right-sizing conversation. Failed onboarding goes to training. Seasonal and contractually locked go to nothing, on purpose.

A playbook is not bureaucracy; it is how six people make the same decision as each other while Jade is on holiday. It is also the thing that improves. As the log fills up, the pattern of which responses actually held accounts becomes visible, and that pattern feeds back twice — into the playbook, and eventually into what the model should be flagging in the first place. That loop only exists if somebody was writing things down when it was inconvenient.

Where this goes wrong

It looks like this. The list goes live, the team wants to be responsive, and every alert gets the same thing: a friendly call and a temporary discount. It is expensive, because a fifth of those accounts were never leaving and have now been trained to expect money off. It is patronising to the healthy ones. It is useless to Corvex, who wanted an engineer, not a voucher. And because nobody recorded which accounts got what, it is impossible to learn from — so when the board asks in six months whether the programme worked, the honest answer is that nobody can tell.

What Jade hands on

What leaves her desk is two things. The first is the intervention playbook: a cause-matched response for every alert type, the capacity rules for scaling it beyond a pilot, and the explicit permission to do nothing. The second is six weeks of logged actions — every account, every response, every date, and the accounts deliberately left alone.

It goes to Nadia, the Product Analytics Lead, in module 12, with a question Jade asks in exactly these words: tell me honestly, did any of it work, or were those customers staying anyway? What her team recorded is the raw material of that answer. Nadia can be as rigorous as she likes and she cannot recover data that was never captured; the quality of the evaluation is capped, right here, by the discipline of six people logging calls on a Friday afternoon.

Match the response to the cause

Read each account, decide what you would send, then tap the card.