How We Were Leaving Expansion Revenue Behind

The Shape of the Gap

Here is how expansion was happening before this project: a CSM had a good relationship with a contact, the contact mentioned growth plans during a QBR, the CSM looped in an AE, and a deal happened. Or an account hit a limit and submitted a support ticket, and someone noticed and escalated it commercially. Or an inbound request came through. All of these are real. None of them are a strategy.

The problem with luck-based expansion is not that it produces zero revenue. It is that it produces wildly inconsistent revenue and creates no learning. You cannot optimise a process you cannot observe. You cannot predict revenue from a motion that depends on which CSM happens to remember which account on which day. And you cannot build a repeatable expansion function on the back of conference conversations and lucky renewal calls.

We had good CSMs. The issue was not their effort or judgment. The issue was that the data informing their judgment was invisible to them. Usage patterns, API consumption trends, feature adoption curves, seat utilisation by team, all of it sitting in a data warehouse that the commercial team had never been given access to in any actionable form.

What the Data Actually Showed

We segmented the customer base by usage intensity across four dimensions: feature breadth, seat utilisation rate, API consumption relative to plan limits, and workflow automation depth. Nothing exotic. All of it available in the warehouse already. Then we asked a simple question: which accounts are in the top quartile on at least three of these four dimensions, and when did they last have a commercial conversation?

The answer was 23% of the base, zero outreach in the prior six months. These were not edge case accounts or tiny customers. Several of them were among our largest contracts. They were using the product deeply, hitting plan limits in some cases, and the commercial team had no idea because nobody had built the bridge between the product data and the people who needed to see it.

We estimated the expansion ARR sitting in that cohort using three conservative assumptions: a 35% attach rate (meaningfully below our historical rate for accounts that did receive outreach), our median expansion deal size, and a six-month selling window. The number was uncomfortable enough that the CFO scheduled a follow-up meeting. That follow-up meeting was the green light for this project.

The Architecture Gap Behind It

Why was the data invisible to the commercial team? Because the standard SaaS data architecture is directional in a way that nobody talks about enough. Data flows from your product databases into the warehouse. The warehouse feeds dashboards and analytics tools. Insight is assumed to travel from dashboards into human decisions. The problem is that the people responsible for commercial action, AEs and CSMs, do not make decisions from dashboards. They make decisions from their CRM. If the insight does not reach the CRM in an actionable form, it does not reach the decision-maker.

Reverse ETL solves this by running the pipeline in the other direction: warehouse to CRM, not just CRM to warehouse. You score your signals in the warehouse, where you have the compute and the data model to do it properly, and you write the result directly into CRM fields where it can trigger workflows, create records, and alert people, exactly as any other CRM event would. The AE does not need to log into a BI tool. They do not need to run a query. They see a notification in the system they already live in, with the signal context pre-processed and the recommended action already attached.

That is the whole architecture. The complexity is not in the technology. It is in deciding what to score, how to score it, and what the output should trigger. That is a RevOps problem, not an engineering problem. The engineering is the easy part.

What an Expansion Qualified Lead Actually Is

Defining the Object Before Building the Pipeline

Before we wrote a single line of pipeline code, we spent three weeks answering one question: what, precisely, is an Expansion Qualified Lead? Not a health score. Not a renewal signal. Not a CSM gut feeling. A specific, bounded definition that any AE or CSM could apply consistently.

We landed on four criteria that all had to be true simultaneously. First, the signal had to represent a specific commercialisable gap, not general positive sentiment. Second, it had to be rooted in product behaviour, not relationship context, because relationship context is not measurable at scale. Third, it had to point at something the account did not already own. And fourth, the estimated expansion had to exceed 15% of current contract value, because below that threshold the cost of an AE-led sale does not pencil.

That fourth criterion created the most friction. CS leadership wanted everything flagged. Their argument was that even small expansions matter for NRR. Our counter-argument was simpler: if you train AEs to treat EQLs as optional noise by filling their queue with sub-threshold opportunities, you lose the high-confidence ones too. You get one shot at establishing that EQLs are worth acting on quickly. Do not waste it on deals that should have been handled through a self-serve order form.

The Signal Architecture: What We Tracked and Why

Six signals. That is all. Every one of them connected to a specific commercial hypothesis and a specific type of conversation. The discipline required here is harder than it sounds: you are not cataloguing interesting product behaviours. You are selecting the subset of behaviours that correlate with a willingness and readiness to spend more money. Those are not the same list.

The export signal is the one that will tempt you to automate and should not be automated. High data export volume is ambiguous in a way that no scoring model can resolve without human context. It might mean a team embedding the product deeper into their infrastructure. It might mean a team preparing to leave. Call it wrong in either direction and the damage is real. Route it to a CSM for a judgment call every time. The two minutes it takes is worth it.

Scoring, Weighting, and the EQL Threshold

Score thresholds alone do not work. This is the mistake most first versions of this kind of system make, and we made it too. An account scoring 85 out of 100 because it is at 92% seat utilisation sounds like a strong EQL. It is not, if the account is 60 days into onboarding or has a contract value of $4,000. A score tells you signal intensity. It does not tell you commercial viability.

The fix is hard gates that sit above the scoring model. Two gates in our case: a minimum contract value floor (accounts below a threshold ACV were routed to a CS-owned self-serve expansion path, not AE-led), and a minimum account tenure of four months (new accounts show high usage that reflects curiosity and exploration, not value realisation). Both gates had to be passed before the score was even evaluated. Adding them took the EQL accuracy rate from 61% to 94% across the pilot cohort, which is a dramatic jump for a relatively simple change.

The accuracy metric here is worth defining precisely, because "accuracy" means different things in different contexts. We measured it as the percentage of EQLs assigned that converted to at least one substantive commercial conversation, defined as a booked call or a proposal sent, within 30 days of assignment. Not closed. Not interested. Just: was there a real conversation? If an AE receives an EQL and decides on review that it does not merit outreach, that is still recorded as a non-conversion for accuracy purposes. This kept us honest about signal quality rather than flattering ourselves with loose definitions.

How the Data Actually Moves

The Reverse ETL Pipeline End-to-End

Once the signal logic was agreed, the engineering was the easy part. Four stages. Transferable to essentially any modern SaaS data stack. The only things that change company to company are the tooling choices in the middle, and those are commodity decisions.

Three decisions in the pipeline took iteration to get right, and all three are easy to get wrong on first build.

Run scoring hourly, not on every event. Real-time scoring sounds better than it is. In practice it creates score oscillations as individual events push accounts above and below the threshold, which produces noise in the CRM and trains AEs to distrust the signal. Expansion selling is not a seconds-matter motion. Hourly cadence is commercially indistinguishable from real-time and far cleaner operationally.

Use incremental sync, not full refresh. A full refresh on a large account database writes to CRM records on every cycle, including records where nothing has changed. This creates spurious workflow triggers, fires duplicate notifications, and confuses the AE experience. Incremental sync writes only new or changed records. It keeps the notification volume meaningful.

Build the dedup check before you go live. We did not, and within two weeks of launch one account had generated seven EQL records because its seat utilisation was oscillating across the threshold on consecutive hourly runs. The dedup logic is simple: check for an existing open EQL on the account before creating a new one, and enforce a 30-day suppression window after any EQL closure regardless of status. It takes an afternoon to build and saves weeks of cleanup.

What the AE Sees in the CRM

Most Reverse ETL write-ups stop at the pipeline. The pipeline is the straightforward part. What determines whether the system produces revenue is what the AE sees when the record lands in their CRM. If they open it and feel confused, or if they have to open three other records to understand what they are looking at, the EQL will sit untouched. You have roughly one notification to establish whether this is worth their time.

We designed the EQL record around three questions an AE needs answered before they pick up the phone: why is this account flagged right now, what is the estimated opportunity, and what is the opening move? The first two come from the pipeline directly. The third required an extra lookup layer that mapped each primary signal type to a short, approved conversation framing, written in product value language rather than commercial language. An AE should be able to read the EQL record and know what they are going to say in the first thirty seconds of the call before they dial.

The opening angle field was the one Sales leadership pushed back on. Experienced AEs do not want a script. Fair point. We were not building a script. We were eliminating the step where the AE contacts the CSM for context before reaching out, which was adding an average of 18 hours to time-to-first-contact. The opening angle is not a constraint on judgment. It is pre-processed context that lets the AE skip a time-consuming prep step and still show up to the call knowing what they are talking about. AEs who tried it kept using it.

The Part That Had Nothing to Do With Data

Who Owns Expansion Leads When You Have Both AEs and CSMs

Who owns an expansion lead on a live customer account? This question will consume more calendar time than any technical decision in the project. Budget for it accordingly.

The answer we arrived at was: it depends on the deal, and the deal type is determined by the system, not by whoever gets there first. EQLs below our expansion ARR threshold route to the CSM with a simplified commercial process and a low-friction order form. EQLs above that threshold, or involving a new product module rather than a seat or tier expansion, route to the AE with a mandatory CSM involvement flag. Both types use the same EQL object. The routing field on the record determines the path.

Getting there took six weeks. The breakthrough came from a meeting format we borrowed from a previous project: everyone in the room writes down their position before anyone speaks. When you externalise your position before the discussion starts, you cannot shift it retrospectively to match what you heard. The negotiation that follows is about the gap between written positions, which is a much more tractable problem than two people trying to win a live argument. It was the most efficient six-hour investment of the entire project.

Compensation, Attribution, and the Incentive Structure

The compensation model is not an afterthought. It is load-bearing. An EQL system that creates opportunities without telling people how they will be credited for working them creates perverse incentives. AEs will protect ownership instead of collaborating. CSMs will either compete or disengage. Neither outcome is what you built the pipeline for.

Three clean rules. AEs who own and close an EQL get full quota credit. CSMs get an expansion attainment bonus tied to their book's total expansion ARR, which makes the EQL system something they want to succeed rather than something they fear. Co-owned deals above the threshold use a documented contribution split, with the ratio written into a structured CRM activity log rather than negotiated after the fact.

That last part, the structured activity log, was the Finance team's condition for approving variable attribution. They wanted an audit trail, not a conversation about who remembers doing what. It adds roughly ninety seconds per deal closed. Not a single person has complained, because the alternative is a split dispute and nobody wants those.

The Numbers, Honestly Presented

The Headline Metrics

Two quarters post-rollout. Here is what we saw, with the attribution caveats clearly stated because overclaiming on causation is the fastest way to lose credibility with Finance and the fastest way to build a system you cannot learn from.

The NRR number needs a footnote. The same two quarters saw an onboarding improvement that raised baseline engagement across the board, and a pricing structure change that improved upsell economics. The EQL system did not operate in a controlled environment. Attributing the full 24-point lift to it would be wrong. My modelled estimate is roughly 60% of the improvement is defensibly EQL-driven. That is still a large number. Present it with the caveat, not without it. The people who will scrutinise it are the same people who will fund the next iteration.

Which Signal Performed Best

Seat utilisation was the highest converting trigger, and the reason is not interesting from a data perspective but is important to understand from a sales perspective. When a customer is at 91% seat utilisation and three people are sharing logins, they already know they have a problem. They have been deferring the purchase decision, not avoiding it. The EQL surfaces it. The AE calls. The conversation is short because the customer is not being persuaded of anything they do not already believe. Most of these close in one or two calls.

API rate approach produced the largest average deal sizes because the expansion path is usually a tier upgrade rather than incremental seats. Technical buyers understand capacity constraints quickly and tend to move fast once they decide. Feature adjacency was the trickiest commercially, because the AE needed genuine product knowledge to frame the conversation correctly, but it produced the best post-close satisfaction scores. Customers who expand because they were already gravitating toward the adjacent product do not feel sold to. They feel helped.

What Actually Changed

The coverage rate is the number I care about most. 31% to 89% of expansion-ready accounts receiving a commercial conversation. That is the structural gap, largely closed. Everything else, the NRR improvement, the deal count increase, the faster close rates, is downstream of coverage. Fix coverage and you unlock the rest.

Something else changed that does not show up in the metrics. The CS-Sales relationship on expansion got visibly better. Before EQLs, CSMs felt AEs appeared in their accounts only when there was money on the table and then disappeared. AEs felt CSMs withheld account access to protect relationship control. Both feelings were grounded in real experiences. The EQL record gave both roles a shared object to coordinate around. It made the process legible to both sides at the same time. Escalations about expansion jurisdiction dropped by roughly 70% in the six months post-rollout. That improvement is not in the NRR model. It is real.

Honest Retrospective

Nine Signals Was Too Many

The pilot had nine signals. Within three weeks, AEs were ignoring EQL notifications. Not because the signals were wrong. Because the accuracy rate was 58%, which meant roughly four in ten EQLs were generating a call where the account was not actually ready for a commercial conversation. Once an AE learns that a notification type is wrong 40% of the time, they treat it as optional. That pattern is very hard to reverse.

We cut to six signals for the full rollout, tightened every threshold based on pilot outcome data, and accuracy jumped to 94%. The signal count reduction mattered, but so did the tightening. The lesson is not "use fewer signals." It is "do not launch signals you have not validated." Every unvalidated signal you launch erodes trust in the ones that are working.

The Scoring Models Drifted Without a Review Cadence

We built the models, launched, and did not formally review them for six months. In that time: two new product feature areas shipped, the ICP shifted slightly toward larger accounts, and a pricing change altered what seat utilisation thresholds meant commercially. The models were still running on the old picture. Accuracy had quietly slipped by about 8 percentage points before the weekly KPI review flagged the trend.

Quarterly model reviews are now on the calendar, tied to the product roadmap cycle. If the product adds a new module, the feature adjacency signal needs to be updated. If pricing changes, the seat utilisation threshold for a commercial trigger may need to change with it. Scoring models are not set-and-forget infrastructure. They are assumptions about your product and customer base that need to be revisited every time those assumptions change.

The CSM Notification Was an Afterthought and Acted Like One

We built the AE workflow first and added the CSM notification later. It showed. The CSM notification had less signal context, fired after the AE had already been alerted, and gave CSMs no window to add relationship context before outreach happened. The result was CSMs finding out an AE had contacted their account from the account itself. That is not a small friction. That is a trust problem between the two roles you most need to cooperate.

We rebuilt the CSM notification to fire simultaneously with the AE assignment, with equivalent signal detail, and with a 24-hour review window on lower-priority EQLs before AE outreach is triggered. That window was used in 12% of cases. In every one of those cases, the CSM flagged a relationship context that would have made the outreach land badly. Twelve percent sounds small. It is not, when the cost of a misjudged outreach on a live account is relationship damage that takes months to repair.

Compensation Communication Was Underinvested

We spent the better part of six weeks designing a compensation model. We spent two weeks communicating it. Wrong ratio. First month post-launch: three compensation disputes, all stemming not from the model being wrong but from edge cases that had not been walked through explicitly in the rollout sessions. One AE believed named account EQLs were fully creditable regardless of contribution. The model did not say that. Nobody had said explicitly that it did not.

Two additional all-hands sessions in month two, plus a one-page decision tree for split-credit scenarios. Disputes went to zero. The time cost was a few hours. The alternative was resentment that would have corroded adoption across both roles for months. Compensation confusion is not a minor operational friction. It is the thing most likely to poison a system that is otherwise working.

Building This in Your Organisation

Four phases. The sequencing is not optional. Most of the mistakes in this case came from doing the right things in the wrong order.

Appendix A: EQL Definition Checklist

Before building any pipeline, use this checklist to validate that your EQL definition is complete. Every unchecked item is a gap that will surface as an adoption problem or a data quality issue after launch.

Appendix B: Expansion Coverage Diagnostic

Run this analysis before building your business case. It produces the number that justifies the investment and reveals the specific customer segments where the expansion gap is largest.