Our SaaS business is at $65M ARR with logo churn at 11% and NRR at 104%. We can spend $3M in 2026 on either AI support automation, AI-assisted onboarding, or AI-driven churn prediction. Which bet most likely adds 2-3 points of NRR by 2027?

Advisors debated whether to invest in churn prediction tooling or AI-assisted onboarding, but the more important finding is that nobody actually knows why customers are leaving — exit survey data covers only ~20% of churned accounts and is systematically biased toward less-dissatisfied customers, meaning the company's current churn narrative is likely wrong. The 11% logo churn figure is probably concentrated in small, low-ACV accounts that CSMs can't economically save, which would make a churn alert system an expensive dead end; onboarding fixes, by contrast, operate at a cost-per-account that pencils out and carry a dual payoff — fewer early churns and more accounts reaching the 'first win' milestone that drives expansion.

The Contrarian REFRAME
Nobody's asking the obvious question: why are 11% of logos actually leaving? Not 'what does our churn prediction model say' — why, physically, do customers cancel? Until someone can answer that in one sentence, spending $3M on any AI tool is just expensive guesswork dressed up as strategy. Churn prediction tells you who is leaving, not why — and a warning with no fix attached is just a countdown clock.

Rita Kowalski PUSHBACK
That's the right instinct but you're stopping one layer too shallow. In my experience at this ARR range, companies think they know why customers leave because they run exit surveys — but exit survey response rates are typically under 20%, and the respondents are skewed toward the ones who liked you enough to bother. The 80% who ghost you? Silent data. So you build your 'why' narrative on a biased sample and convince yourself it's insight. I ripped apart a client's entire churn analysis last year and found they'd been blaming pricing for three years — turned out pricing was a proxy complaint covering a product gap that showed up in support tickets nobody had bothered to tag systematically. The real 'why' is almost always locked in unstructured data this company almost certainly hasn't instrumented.

Rachel Wong REFRAME
Rita's absolutely right about the biased data problem — but here's what's actually bugging me about this whole conversation: we're all treating 11% logo churn like it's a monolithic number, and it's almost certainly not. At $65M ARR, logo churn almost always skews toward the smaller, lower-ACV accounts — the borderline fits who squeaked through the sales motion. If that's what's happening here, a churn prediction model fires alerts on $8K ARR logos your CSMs can't economically save, and you've just built a very expensive anxiety machine. The intervention unit economics collapse completely when the churning segment is sub-threshold for high-touch CS. Before any of us can call a winner here, they need ARR-weighted churn alongside logo churn — because if the logos leaving are disproportionately small, AI-assisted onboarding is actually the higher-leverage play: you catch bad-fit customers before the sunk cost calculus kicks in, at a cost-per-account that actually makes sense.

Laurent Jorgensen REFRAME
Rachel's unit economics argument is solid, but she's only running half the NRR math. At 104% NRR, this company's expansion motion is barely keeping its head above water — and in my six years watching accounts, the ones that actually expand are almost always the ones who hit a clear first win inside 60 days. Fix onboarding, and you're not just plugging the churn hole — you're unlocking expansion from accounts that are currently stuck in implementation purgatory, too confused to go deeper, too invested to leave yet. Nobody's counting that upside, but it's real, and it's why the NRR impact from onboarding could be a hell of a lot bigger than 2-3 points.

Advisors debated whether onboarding is the actual causal lever for churn or a proximate symptom — with The Contrarian arguing the 104% NRR expanding accounts undermine the case for a $3M onboarding fix, Laurent countering that those accounts succeed despite broken onboarding (not because of good onboarding), and Rachel framing the real opportunity as the bottom 60% of the cohort who lack internal resources to self-rescue. A secondary dispute broke out over the investment timeline for churn prediction AI: Rachel argued you need a full churn cycle to validate signals (pushing ROI to 2028), while The Auditor pushed back that years of existing product logs and CRM data make that a moot concern — the real barrier is infrastructure and data science talent, not data latency.

The Contrarian REFRAME
Laurent's 'implementation purgatory' framing is interesting but it assumes the bottleneck is actually onboarding — nobody has verified that. Here's what I keep coming back to: this company is at 104% NRR, which means expansion is happening somewhere. Find those accounts. What did their first 90 days look like? If they onboarded the same way as the churners and still expanded, then onboarding isn't the variable — something else is. Spending $3M to fix a process that might not be the causal lever is how companies feel productive while the real problem sits untouched.

Rachel Wong PUSHBACK
The Contrarian's expanding-account comparison assumes onboarding is a binary variable — same process, different outcome — but the real dimension is time-to-value velocity, and it explains why those two cohorts look identical on paper. Your expanding accounts are almost certainly your top-quintile customers: strong internal champions, simpler technical lift, dedicated implementation resources on their side — they'd extract value from mediocre onboarding anyway. AI-assisted onboarding isn't built for them, it's built for the bottom 60% of your cohort who are technically capable but operationally stuck, and that's exactly where the unexploited NRR upside is hiding. And the timeline math nobody's running: a churn prediction model at this scale needs one full churn cycle to train and validate — at 11% annual, that's late 2026 before you can trust the signals, which means you're looking at a 2028 NRR story, not 2027.

Laurent Jorgensen REFRAME
The expanding-account comparison is genuinely useful, but it's being used to argue inaction rather than direction. I ran a cohort analysis at my last company that looked exactly like this — same onboarding playbook, wildly different outcomes — and what we found wasn't that onboarding was irrelevant, it was that our best customers were compensating for a broken process with their own internal resources. They succeeded despite us. The bottom 60% didn't have that internal bandwidth and they quietly disengaged before month four. That's not an onboarding exoneration, that's an onboarding indictment with extra steps.

The Auditor FACTCHECK
Rachel, I need to flag the timeline claim specifically — the assertion that a churn prediction model requires 'one full churn cycle to train and validate' before you can trust the signals. The web research I'm looking at from Databricks and MindsDB explicitly describes training on existing historical customer transaction and behavioral data — not waiting for a new churn cohort to complete. At $65M ARR with 11% annual logo churn, this company has years of historical signal sitting in their product logs and CRM right now. The '2028 NRR story' framing assumes a greenfield data environment that almost certainly doesn't exist here. The real implementation risk isn't data latency — it's the infrastructure investment and data science expertise the research flags as significant barriers for traditional approaches.

Advisors debated whether to invest in churn prediction or AI-assisted onboarding at $65M ARR / 11% logo churn. The Auditor reframed churn prediction as a diagnostic instrument (via SHAP interpretability) rather than just an alert system, Rachel countered that modern deployments ship automated intervention rails that offload triage from CSMs and build a compounding proprietary data moat, while Laurent and Rita pushed back — Laurent arguing that alert systems without CSM capacity create 'anxiety dashboards,' and Rita landing the sharpest challenge: at this ARR profile, churning logos are likely disproportionately small accounts where the unit economics of saving them are already broken, making the entire $3M debate potentially aimed at the wrong problem.

The Auditor FACTCHECK
Rachel, I've got to flag something that changes the churn prediction calculus. The web research I'm looking at — including a systematic review of 240 peer-reviewed churn studies published 2020 to 2024 — specifically highlights SHAP value interpretability as a core output of modern churn models. That means a well-built churn prediction system doesn't just flag who's at risk — it tells you which specific behavioral signals are driving each account toward the exit. That's not a 2028 story, that's the answer to The Contrarian's Round 1 question about why customers are leaving, and it directly addresses the exit survey bias problem nobody's solved yet. The team is treating churn prediction as a retention alert system when it's actually a diagnostic instrument — and that reframes the entire ROI argument.

Laurent Jorgensen WARN
The Auditor, the SHAP interpretability argument is the sharpest thing anyone's said in three rounds — and it still doesn't solve the problem I've watched sink good CS teams. Churn prediction outputs a list. Someone has to work that list. At 11% logo churn on a base this size, you're flagging dozens of at-risk accounts every quarter, and if the CSM team doesn't have capacity to run saves right now — which they almost certainly don't, because 11% churn means they're already triage-mode — then you've spent $3M building an expensive anxiety dashboard that tells your team exactly how much they're failing. AI-assisted onboarding moves in the opposite direction: it reduces the per-account load instead of multiplying it, and I've watched enough burned-out CS reps to know that handing people alert systems without the headcount to act on them is how you lose your best people right when churn gets bad.

Rachel Wong PUSHBACK
Laurent, the anxiety dashboard framing only holds if the intervention layer stays manual — and that's the 2018 version of this product category. Every churn prediction deployment worth its price tag that I've seen in the last 18 months ships with automated intervention rails: triggered in-app health sequences, auto-scheduled EBRs, personalized re-engagement cadences that fire before a CSM even opens their laptop. The CSM works the top ten percent of red accounts personally; the system handles triage on the rest. You're not multiplying workload, you're offloading the commodity saves so your best CS people can focus on the accounts where a human conversation actually moves the needle. And here's the angle nobody in this room has touched yet — the onboarding fix is fully replicable the moment a competitor hires one good VP of CS and runs a decent implementation playbook, but a behavioral churn model trained on three years of your specific product telemetry is a proprietary asset that compounds over time. That's the moat, and that's what changes the due diligence story in 2027.

Rita Kowalski WARN
Rachel, the 'proprietary behavioral model as a moat' argument is the one I hear from every vendor selling churn prediction software, and I've watched it collapse in three separate client engagements. Here's the failure mode nobody's naming: at $65M ARR with 11% logo churn, your highest-priority intervention target is almost certainly a segment you can't afford to save economically — and a sophisticated SHAP output telling you exactly why a $15K ACV account is churning doesn't change the unit economics of dispatching a CSM to save it. The model gets smarter, the math stays broken. What I actually want to know before any of this is whether this company has ever separated ARR-weighted churn from logo churn, because if the churning logos are disproportionately small — which at this ARR profile they almost certainly are — then the entire $3M conversation is aimed at the wrong problem entirely.

Each advisor surfaced the bedrock assumption driving their prior positions: Laurent trusts frontline CSM knowledge over dashboards; Rita treats measurement infrastructure as prerequisite to any valid causal claim; The Auditor believes better information produces better decisions; Rachel argues capital allocation follows pattern recognition, not analysis. The round shifted from debating tactics (onboarding vs. churn prediction) to exposing why each advisor's framework has a specific blind spot — and where it breaks down at a $65M ARR company with 11% logo churn.

Laurent Jorgensen SHARE
Here's the assumption I've never questioned: I believe that the people closest to the work — the CSMs, the onboarding specialists, the support reps — already know why customers leave, and the job of leadership is to stop overriding them with dashboards and start listening to them. Every position I've taken in this debate flows directly from that. I pushed onboarding because my team told me four years ago that customers were drowning in week three, and they were right. I've never once had a churn prediction tool tell me something my senior CSMs didn't already know. Rita's challenge about ARR-weighted churn is where our assumptions diverge most sharply — she takes for granted that the right answer lives in segmented financial data, and she got there because she's spent her career at the portfolio and CFO level, where the numbers are the truth and the frontline signal is noise. I think that view is wrong, but I understand exactly why it feels obvious to her.

Rita Kowalski OBSERVE
Here's what I take for granted and have never seriously questioned: that the measurement layer is prior to everything else. Not strategically prior — ontologically prior. If you can't reliably detect a change, the change doesn't exist for decision-making purposes. Every position I took in this debate — the KPI audit challenge, the ARR-weighted churn push, the skepticism about attributing 2-3 NRR points to any single tool — flows directly from that bedrock assumption. I came up building operations systems, and what I saw over and over was that smart people made confident causal claims about interventions that their measurement stack literally could not support. I stopped trusting confident causal claims without first asking 'can your instruments even see this?' Laurent's assumption is almost the mirror image of mine: he takes for granted that the people closest to the work hold the ground truth, and that dashboards and diagnostic systems are mostly leadership theater that filters out the signal that already exists in human knowledge. That view feels obvious to him because he's right about a narrow version of it — his CSMs did know customers were drowning in week three, and they were ignored. But the assumption breaks down at the edges, because frontline workers only know what they see, and what they see is systematically biased toward customers who stayed long enough to complain. The accounts that churned quietly in month two never told anyone anything. Laurent's framework is built from the customers who survived long enough to be visible.

The Auditor SHARE
My bedrock assumption — one I've genuinely never interrogated until this round — is that the primary failure mode in business decisions is informational: surface a clean, verified signal and better choices follow. Every position I took in this debate flows directly from that premise: the SHAP interpretability argument, the 'historical product logs already exist' pushback on the churn cycle timing question, the reframe of churn prediction as diagnostic instrument rather than alert system — all of it assumes that if you hand people accurate information clearly enough, the right action becomes available to them. Rita, where our worldviews actually diverge is more specific than I expected: you take for granted that measurement instruments, once properly designed, produce outputs that are approximately neutral — but what your framework doesn't account for is that the person who designs the measurement system embeds their own assumptions about what to count, how to aggregate, and what registers as signal versus noise in the first place. You came up building operations systems where the design layer was already settled before you inherited responsibility for the outputs, and that experience makes 'establish the measurement foundation first' feel like an obvious prior truth — when what you're actually taking for granted is that the instrument's underlying design choices are already sound, which is exactly the assumption most worth auditing.

Rachel Wong SHARE
What I take for granted — and I mean bedrock, never-questioned — is that capital follows pattern recognition, not analysis. Every investment thesis I've backed has been a pattern match first and a model second. So when I pushed churn prediction in rounds one through three, I wasn't reasoning from first principles about measurement systems or intervention economics. I was pattern-matching: companies that invest in understanding why they lose customers compound better than companies that invest in fixing the experience of customers they're already losing. That assumption shaped every position I took. Now, The Auditor's assumption — that surfacing clean, verified signal leads to better choices — feels obvious to them because it works in environments where the decision-maker has both the capacity and the mandate to act on information. But in a $65M ARR company with 11% logo churn, the constraint isn't informational. The CSMs already know who's at risk. The real constraint is organizational will and economic incentive to act on what they know. The Auditor built their worldview inside systems where information scarcity was the binding constraint, and so 'surface the signal' always moved the outcome. That background makes the assumption invisible to them — but it breaks down exactly here.

In a moment of collective self-awareness, all four advisors broke from their roles to name what the debate had been obscuring: the measurement infrastructure is likely broken, the advisory process itself has been providing intellectual cover for an executive who hasn't done basic internal homework, and if forced to commit capital today, the only defensible bet is AI-assisted onboarding — because it's the sole option with a feedback loop that closes before 2027. Laurent dissented on timing: no $3M should move until someone spends four hours listening to CSMs, since the real churn diagnosis already exists in frontline heads and has never reached budget authority.

Rachel Wong SHARE
The detail I keep coming back to: the sound of ice settling in a water glass during a pitch, right after a founder can't answer 'who specifically is churning and why.' That silence — that's this entire debate. We've been the panel of smart people that gives the executive permission to act, so when the $3M bet lands wrong, it wasn't their call alone, it was 'the advisors agreed.' Rita's playing the instrument-builder, Laurent's playing the frontline prophet, The Auditor's playing the epistemologist — and I've been playing the pattern-matcher who makes the bet feel obvious, which is the most enabling role of all. Stepping off this stage looks like telling the person that the answer to 'which AI tool' is downstream of a conversation they haven't had with their own CFO about ARR-weighted churn — and until they have that conversation, we're all just dressing up their avoidance in intellectual scaffolding.

The Contrarian OBSERVE
The physical sensation this whole debate gives me is gripping a steering wheel on a car that's still in park — five people revving hard, nobody moving an inch. We've all been playing 'the advisor who sees what others miss,' and I've had the safest role in the room: the contrarian who points at the wrong question and never has to answer the right one. The uncomfortable truth I've been protecting myself from is that 'check if you're solving the right problem' is the most comfortable escape hatch in any boardroom — I get to be right about the meta-problem while contributing exactly nothing to the actual capital decision. Stepping off the stage means saying what I'd actually do: fix onboarding, because it's the only option where I can trace the physical chain of events — customer signs, hits a wall, someone catches them, they stay — and that feedback loop closes before 2027, which is the only one of the three bets where I'm not just buying a smarter way to watch people leave.

Rita Kowalski OBSERVE
The detail that keeps surfacing for me is tactile: it's the feeling of pulling a client's 'churn report' and finding the denominator is wrong — not wrong in a sophisticated way, wrong like someone divided by total logos when they should've divided by logos at start of period, and nobody caught it for eighteen months because the number was moving in the right direction and nobody wanted to look too hard. That's the sensation under this entire debate. We've been arguing about which intervention to fund, but the play we've been performing is called 'Advisors Who Each Secretly Know The Measurement Infrastructure Is Broken But Keep Talking Anyway.' Laurent's the believer who trusts the humans because the numbers have always let him down. The Auditor's the architect who keeps drawing blueprints for a house nobody's confirmed has a foundation. Rachel's the pattern-matcher who moved on from the measurement question two rounds ago because her edge is speed, not rigor. And I'm playing the inspector who won't let anyone through the door — which sounds responsible until you realize I've been standing at the door so long I've forgotten there's a client inside who needs an answer. Stepping off this stage means I say it plainly: if forced to pick with a gun to my head, I'd fund AI-assisted onboarding — not because I trust the NRR attribution, but because it's the only intervention where the failure mode is visible fast enough to course-correct before 2027.

Laurent Jorgensen OBSERVE
The detail that keeps surfacing for me is a specific weight — the weight of a printed churn report sitting in a manila folder on a conference table, untouched for three days. I've held that folder. My team spent forty hours on it. The recurring play we've all been performing is this: every advisor in this room is playing the expert who arrives after the body is already cold, and we're all arguing about which autopsy instrument is most sophisticated. Rita's the instrument designer. The Auditor's the pathologist. Rachel's the investor who funded the hospital. And me? I've been playing the ward nurse who keeps saying 'we told you he was deteriorating on Tuesday' — and nobody in the play ever writes that line into the resolution. Stepping off the stage looks like this: before any $3M moves, someone sits in a room with six CSMs for four hours and just shuts up and listens. No framework. No SHAP output. No measurement architecture. The answer about why customers leave is already fully formed in those people's heads — and the real reason this company is at 11% logo churn is that nobody with budget authority has ever had that conversation.