What Building an In-House CRM Taught Us About Our Own Sales Funnel

We did not set out to learn anything about our sales funnel. We set out to fix a reporting problem.

For years our pipeline lived where most agencies' pipelines live: in a spreadsheet that one person owned, a few inboxes, a shared calendar, and the collective memory of whoever happened to be on the call. It worked, in the way that things work right up until they don't. When someone asked "how many leads did we close last quarter and where did the rest go," the honest answer was a shrug dressed up as a number. We could tell you how many leads came in. We could not tell you, with any confidence, where they went after that.

So we built our own CRM. Not because the market lacks good ones - it doesn't - but because we wanted to control exactly how every stage, field, and transition was defined, and we wanted the data to live in one place we could query. We expected to come out the other side with cleaner reports. What we actually came out with was an uncomfortable, useful education about a funnel we thought we already understood.

This is that education. The software is beside the point - you do not need to build anything to take these lessons. What you need is the discipline that building forced on us: defining every stage by evidence, timestamping every move, and being willing to read what the data says instead of what you wish it said.

The short answer

Building an in-house CRM taught us five things about our own sales funnel that our spreadsheet had been hiding: our funnel had stages we had never named and stages we tracked that did not really exist; our biggest leak was slow lead response in the middle, not lost proposals at the bottom; "qualified" was a story we told ourselves rather than a defined condition; attribution only got honest once a single system owned the whole timeline; and the funnel behaves like a conversation cadence, not a static pipeline. You can capture all five lessons in any decent CRM - or even a disciplined spreadsheet - because the value was never the code. It was being forced to define things we had been guessing at.

Why we built it at all

A quick word on the "build your own CRM" decision, because it deserves honesty: for almost everyone reading this, the answer is don't. Off-the-shelf tools are cheaper, faster, and better-maintained than anything you will build in a quarter. We are a B2B services business with specific reporting needs and engineers in the building, and even for us it was a close call.

What pushed us over the line was not a feature gap. It was that every off-the-shelf setup we tried let us be vague. We could drag a card from one column to another without anyone agreeing what that move meant. We could mark a lead "qualified" because it felt right. Building our own forced us to write down the rules - and writing down the rules turned out to be where all the value was. If you take one thing from this piece, take that: most of what we learned, you can learn by setting up a standard CRM as if every definition mattered, because it does.

Lesson 1: Our funnel had stages that didn't exist - and stages we never named

The first thing the build exposed was that our mental model of the funnel and our actual funnel were different shapes.

On the whiteboard, our funnel was clean: lead, qualified, proposal, won. Four tidy stages. The moment we tried to make every real lead pass through those four stages, the model fell apart. Leads kept getting stuck in places that had no name. There was a limbo between "lead came in" and "someone qualified it" where a huge number of leads simply sat - not rejected, not progressed, just waiting. We had never drawn that stage because, in our heads, qualification happened instantly. In reality it happened whenever someone got around to it.

We also had a phantom stage. "Proposal" on our whiteboard implied a single event. In practice "proposal" covered three different things - a verbal scope on a call, a written estimate, and a formal proposal document - and they had wildly different close rates. Collapsing them into one stage meant our "proposal-to-won" number was an average of three things that should never have been averaged.

The lesson generalises well beyond us. The gap between the marketing funnel and the sales funnel - between when a stranger raises their hand and when a salesperson treats them as real - is almost always an unnamed stage. And unnamed stages are where leads go to die quietly, because nobody owns a stage that does not exist.

The fix was not clever. We renamed the funnel to match reality: new lead, contacted, qualified, scoped, proposed, won or lost. Six stages, each with a written definition of what gets a lead in and what gets one out. Boring. Transformative.

Lesson 2: The biggest leak wasn't where we thought

Ask any agency where it loses deals and you will hear the same answer: at the bottom. We lose on price. We lose to a competitor. We lose because the timing was wrong. The bottom of the funnel feels like where the drama is, so that is where we assumed our leak was.

Once every lead lived on one timeline with timestamps, the data said something we did not want to hear. Our biggest leak was in the middle, and it had a single dominant cause: how long it took a human to respond after a lead raised their hand.

Leads that were contacted within minutes converted at a dramatically higher rate than leads contacted hours later, and the leads contacted the next day might as well not have existed. This is not a new finding - the five-minute lead response rule is well documented across the industry - but there is a difference between knowing something from a study and seeing it in your own pipeline with your own logo on it. We had been losing a meaningful share of every month's leads not to competitors and not to price, but to the simple fact that nobody got to them while they were still warm.

The reason this leak hides so well is that it leaves no evidence in a spreadsheet. A lead that was never properly followed up does not show up as a loss - it shows up as a row that quietly stopped updating. You only see it when a system is timestamping every transition and you can ask "how long did leads sit before first contact, and what happened to the slow ones." The answer reframed our entire idea of where to spend effort.

The best part: it is the cheapest leak in the funnel to fix. It needs no more budget and no more leads - just a routing rule and the discipline to answer fast. We rebuilt our intake so that a new lead pages a human immediately and the clock is visible to everyone. That single change did more for our close rate than any campaign we ran that quarter.

Lesson 3: "Qualified" was a story we told ourselves

Here is the most humbling one. When we forced ourselves to write down what "qualified" actually meant, we discovered we had never agreed on it.

To one person, qualified meant "they have budget." To another, it meant "they replied to my email." To a third, it meant "I have a good feeling about this one." All three were marking leads as qualified, into the same funnel, with no shared definition. Our qualified-lead count - a number we reported, planned around, and felt good about - was an average of three different people's optimism.

This is the quiet rot in most funnels. Stages get defined by mood, not evidence. And mood-defined stages produce reports that everyone reads differently and nobody can act on, because the number does not mean a consistent thing from one week to the next.

We replaced the feeling with a checklist. A lead is qualified when three observable conditions are true: the problem is confirmed in their words, the decision-maker is identified, and there is a real budget range on the table - not a guess, a range they have said out loud. Optimism is not on the list. The instant we did this, our qualified count dropped by roughly a third, which felt like bad news for about a day and then turned out to be the best thing that happened to our forecasting. A smaller, true number beats a bigger, fictional one every time. If you do nothing else after reading this, write down what each of your stages actually requires and make everyone use the same definition.

Lesson 4: Attribution got honest only when one system owned the whole timeline

We thought we knew which channels produced our best clients. We were running content, search, referral, and outbound, and we had a confident internal story about which one was carrying the business.

That story was built on first-touch memory. "Oh, they came from a referral" was something the salesperson remembered, not something the data recorded - and human memory is generous to whichever channel the person happens to like. Once every lead carried its real source and its full timeline in one system, the story changed. Some channels produced a lot of leads that never closed. Others produced few leads that closed fast and stayed for years. Our lead-to-customer rate by source looked nothing like our lead-volume-by-source chart, and we had been making budget decisions off the wrong one.

This is the deeper truth about marketing attribution: it is not really a modelling problem, it is a data-ownership problem. As long as the customer journey is scattered across an ad platform, an analytics tool, a salesperson's memory, and a spreadsheet, every attribution model is just a different way of guessing. The moment one system owns the whole timeline from first touch to closed deal, you stop guessing. You also stop over-investing in channels that produce impressive lead counts and disappointing customers - which, in our case, quietly reshaped where we put money. The same principle drives why we push clients toward first-party data and enhanced conversions: own the signal, or model around its absence forever.

Lesson 5: The funnel is a conversation cadence, not a pipeline

The last lesson is more philosophical, but it changed how we run the whole thing. We had always pictured the funnel as a pipeline - a tube that leads flow down, losing some volume at each joint. That image is comforting and wrong. It treats leads as water, passive and uniform, moving by gravity.

What the timeline data actually showed is that the funnel is a cadence of conversations, and leads move because someone moves them. The stages are not pipe segments; they are prompts for a specific next action within a specific window. A lead in "contacted" is not waiting to flow downward - it is waiting for a follow-up, and if that follow-up does not happen inside its window, the lead does not back up in the pipe, it evaporates.

Once we held the cadence model instead of the pipeline model, our metrics changed. We stopped obsessing over how many leads sat in each stage and started watching whether the next action happened inside its window. Time-in-stage became as important as conversion rate. A lead stuck in "scoped" for ten days was not a healthy lead in our pipe - it was a stalled conversation we needed to either revive or kill. This is the same shift we push clients toward when we run their lead generation and B2B lead generation programmes: a lead is not an asset you bank, it is a conversation with a half-life.

What this means for your funnel - even if you never build anything

You do not need engineers or a custom build to capture any of this. You need to treat your funnel setup as a strategy exercise instead of a data-entry chore. Here is the transferable checklist we wish we had started with:

1. Map your real stages, not your whiteboard stages. Watch ten actual leads move through your process and name every place they actually pause - including the ugly, unnamed ones between marketing and sales. Those gaps are where your leaks hide.
2. Define every stage by an observable condition. Write down exactly what gets a lead into a stage and what gets them out. "Qualified" should be a checklist, never a feeling.
3. Timestamp every transition. You cannot find time-based leaks without time data. The single most valuable field in any CRM is "when did this move," because it surfaces both slow response and stalled deals.
4. Instrument first contact above everything. Measure the gap between a lead arriving and a human reaching them, make it visible, and route new leads to a person immediately. It is the cheapest, highest-return fix in the whole funnel.
5. Carry true source from first touch to closed deal. One system should own the lead's whole timeline so attribution stops being a memory exercise. Look at lead-to-customer rate by source, not lead-count by source.
6. Watch time-in-stage as closely as conversion rate. A stalled conversation is a loss that has not been recorded yet. Set a target window for each stage and treat anything past it as a lead that needs reviving or killing.
7. Report the smaller true number. When your honest definitions shrink your pipeline, that is the system working. A forecast built on fiction is worse than a smaller forecast built on evidence.

Every one of these is achievable in an off-the-shelf CRM, a well-built spreadsheet, or a notebook with enough discipline. The tool did not teach us these lessons. Being forced to define things did. You can give yourself that same forcing function on purpose.

The mistakes we'd warn you about

A few things we got wrong so you do not have to:

• We over-engineered the stages first. Our first design had nine stages with sub-statuses. Nobody used them honestly because the friction was too high. Fewer, well-defined stages beat many precise ones that people route around. Start with five or six.
• We measured volume before we measured movement. For the first month we proudly tracked lead counts, which told us nothing about where deals died. The diagnostic metrics - stage conversion, time-in-stage, response time - are the ones worth wiring up first. We learned this the hard way, the same way a lot of teams track blog traffic instead of revenue signals.
• We let definitions drift. Within weeks, "qualified" started quietly meaning "feels good" again, because old habits are strong. Definitions need an owner and a periodic audit, or they rot back into mood.
• We treated it as a sales tool when it was a marketing tool too. The funnel data told us as much about which marketing channels produced real customers as it did about how sales closed them. Keep both teams reading the same timeline.

What we measure now

Here is the short version of what we actually watch, and why each one earns its place:

Total leads and total revenue still sit on the dashboard, but we treat them as outcomes, not diagnostics. They tell us whether things are working. The five metrics above tell us what to fix.

The bottom line

We built a CRM to get better reports and accidentally got a better business, because the build forced us to stop describing our funnel from memory and start reading it from evidence. The lessons had nothing to do with software. Our funnel had stages we had never named. Our biggest leak was a response-time problem hiding in the middle, not a price problem at the bottom. "Qualified" meant whatever the person saying it wanted it to mean. Our attribution was memory dressed as data. And the funnel was never a pipe - it was a set of conversations with half-lives, moving only when someone moved them.

You can have all of that without writing a line of code. Set up whatever tool you already use as if every definition mattered, timestamp everything, and instrument first contact above all else. Then read what it tells you, especially the parts you would rather not hear. That is the whole lesson - the discipline, not the database.

If you want a second set of eyes on where your own funnel is leaking, or help building the measurement so the leaks stop hiding, tell us what you are trying to fix and we will tell you honestly where we would start.