The most expensive problems in a business are the ones that survive the good times. They hide inside averages, they never appear in the reports you already read, and by the time they surface on their own they have been compounding for years. We recently found one that had been running for seven: roughly one in every three orders in a US ecommerce business reached the checkout and never completed.
The founder is technically capable, knows his own operation cold, and had even guessed the failure rate almost exactly. He had never measured it, because nothing in his day-to-day reporting ever asked him to. This post is about how a leak like that stays invisible, how we found it, and how to run the same checks on your own business. The numbers come from a real engagement; the full story is in the case study, and the founder tells his version of it on camera there too.
What the founder asked for, and what the data said
Karim Dahmani runs a US direct-to-consumer ecommerce business. When he came to us, his list of problems looked like most founders' lists: revenue had fallen sharply (he put it down to a supply constraint he had already fixed), support had slipped from replies-in-minutes to replies-in-hours, and he had a long backlog of things he could build with no defensible way to rank them.
What he wanted was an AI support agent. That was the brief.
We run assessments in a specific order precisely because the brief is often not the problem. Before recommending anything, we pulled his complete order history: 23,351 orders across seven years, September 2019 to June 2026, every status included. We interviewed him and the person who runs daily support and fulfilment, five structured interviews in total, and then cross-referenced what people believed against what the orders recorded.
The support agent made the roadmap. It came third.
Ahead of it sat a structural payment leak: 32.6% of mature orders, orders old enough that their outcome is final, had entered the funnel and never converted. Not during a bad patch. Across the life of the business.
| Metric | Result |
|---|---|
| Orders analysed | 23,351 |
| History covered | 7 years |
| Mature orders never completed | 32.6% |
| Rate through the strongest trading period | ~28% |
| Founder's own pre-analysis estimate | 23 to 30% |
| Median time for a manual payment to clear | 2.35 hours |
| Median on the automated path | 0.5 hours |
| Stall rate spread across payment options in use | 18.5% to 54% |
| Engagement length | 4 weeks, a few hours of the founder's time |
Why nobody saw it for seven years
A 32.6% failure rate sounds like something you would notice. Three things kept it invisible, and none of them are specific to this business.
The reports you already read cannot show it
Standard reporting is built on completed orders: revenue by month, average order value, top products, repeat rate. Every one of those reports starts by filtering out the rows where the money never arrived. The leak lived entirely inside the filtered-out rows. You cannot see what your reports are designed to exclude, and no dashboard flags the absence of money that was never booked.
This is why the first request in our assessments is the unfiltered order history, cancelled and abandoned orders included. It is a heavier pull and a messier dataset, and it is the only version of the data that contains the losses.
It survived the best trading period the business ever had
The instinct that protects most operational problems is "we were fine when we were busy." If the business had its best year with the current setup, the setup cannot be the problem. The data said otherwise: through 2022 to 2024, the strongest trading period in the company's history, the leak ran at roughly 28%. It was not a symptom of the downturn. It was structural, and a structural rate has an uncomfortable property: when volume recovers, the same percentage takes a larger absolute amount with it. Fixing it is worth more during a recovery, not less.
That distinction changed the entire ranking of what to build first. A leak caused by the downturn heals when the downturn does. A structural leak compounds.
The mechanism hid inside ordinary-looking friction
Payment in this business runs on manual instructions: the customer places an order, waits for a human to send payment details, pays through a separate channel, and confirms by email. Every step looks reasonable in isolation. Measured across seven years, the friction is stark: a median of 2.35 hours for a manually instructed payment to clear against 0.5 hours on the automated path, and when a payment needed reconciling over email, the tail stretched past ten hours. 16% of manually paid orders took more than a day. Orders age out, stock holds expire, and customers cool off, one order at a time, for seven years.
The clearest signal was already sitting inside the checkout: across the payment options the business was actively using, the stall rate ranged from 18.5% on the best option to 54% on the worst. Nearly a threefold difference, in production, in the same store. Closing the gap to the option that already performs is a 34% reduction in stalled orders with no new vendor involved.
We have written before about what response latency does to conversion in a completely different context, in the five-minute rule. This is the same physics on a different surface: time between intent and completion is where money quietly leaves.
The $1.7 million figure we refused to headline
Across seven years, the orders that never converted came to roughly $1.7 million in gross intended value, and that figure is not a recoverable balance, was never a recoverable balance, and we said so in the report in exactly those words. Some of those customers would never have paid. Some orders were replacements or tests. Recovering all of it would require a time machine.
We separated the impressive number from the honest one deliberately. The recoverable share was modelled on the most recent measured quarter, with no assumed return to historical volume: a ceiling anchored to the business's own data (the 18.5% option already running in production), a floor taken from conservative published recovery benchmarks, and a central case at a 20% reduction in stalled orders. Every input traces to the order history or is labelled an estimate.
If an analyst shows you a seven-figure headline and the model underneath assumes your best year comes back, the headline is doing the work the evidence cannot. The test of a serious diagnostic is whether the conservative number still justifies the fix. Here it did, and the founder could check every step of the arithmetic against his own export.
How to check your own business for a leak like this
The method transfers to any business with transactional history. None of it requires our involvement, and most of it does not require AI. It requires the discipline to look at the rows your reporting throws away.
-
Pull the unfiltered history. Every order or transaction, every status, as far back as the records go. Raw export, not dashboard summaries. Dashboards inherit the completed-only filter you are trying to escape.
-
Measure completion against mature orders only. Recent orders have not had time to resolve, and counting them understates the leak. Pick a maturity window (we used fourteen days), take orders older than that, and measure what share never completed. Write your gut estimate down before you run the query. The gap between guess and measurement tells you how much your reporting has been hiding. Karim guessed 23 to 30%. The export said 32.6%.
-
Run the structural test. Recompute the same rate through your strongest trading period. If the leak collapses when times were good, it is circumstantial and it will shrink on its own. If it holds, it is structural, and it will scale right back up with your recovery.
-
Measure latency, not just failure. For every completed transaction, compute time from intent to completion, and look at the median and the tail separately. Averages flatter you: a 2.35-hour median hid a tail past ten hours here. The orders in the tail are the ones that die.
-
Use your own spread as the ceiling. Segment the failure rate by payment path, channel, or product line. The gap between your best-performing segment and your blended average is recovery that requires no new vendor, because one version of your own funnel already achieves it. That framing keeps the model honest and the fix concrete.
-
Corroborate with the people in the workflow, then rank. The data tells you what is happening; the people closest to the work usually know why. We ran five interviews and scored seven opportunities on the same four dimensions (impact, feasibility, confidence, strategic alignment) so the roadmap ranked things a spreadsheet could defend, rather than things that felt urgent that week.
If you carry one thing out of this list, carry step three. Most operational analysis never separates structural from circumstantial, and it is the difference between a problem worth fixing now and a problem the recovery fixes for you.
What happened after the report
Within weeks of the Blueprint, Karim had shipped the two highest-priority fixes himself, working from our specifications: the payment fix and the fulfilment integration. He is a founder who can build, the specs were written to be built from, and the fastest path to closing the leak was exactly that.
The third system on the roadmap, the AI support layer he had originally come for, he handed back to us to design in full: a project proposal and a technical design specification with the architecture, data model, environments and acceptance checkpoints, written so any competent build team can start from it. The assessment's job was never to sell a build. It was to make the ranking defensible, and the ranking said the support agent mattered, just not first.
That order of operations is the quiet lesson of the whole engagement. The system he asked for was real. The system the data put ahead of it was worth more, and it had been for seven years.
FAQ
What is an AI Strategic Assessment?
A paid, fixed-scope diagnostic. We map workflows, interview the people closest to the work, analyse the operation's own data, score the opportunities on impact, feasibility, confidence and strategic fit, and deliver a Blueprint: findings with evidence, an ROI model anchored to your numbers, and a sequenced roadmap. The assessment page covers scope, timeline and cost. This engagement ran four weeks and cost the founder a few hours of his own time.
How much history do I need for the checks to work?
Enough to include at least one strong trading period, because the structural test in step three needs contrast: you are comparing the leak during good times against the leak during bad ones. For most established businesses that means two or more years. Seven years, as in this case, makes the structural signal unambiguous.
My platform's reports do not show cancelled orders. Is the data gone?
Almost certainly not. Most commerce platforms and billing systems store every order regardless of status; the dashboards simply exclude the incomplete ones by default. Export the raw order table with all statuses rather than a sales report, and the rows will be there, usually back to the first day of trading.
Where does AI actually come into this?
Honestly, the finding itself is statistics and discipline, not AI. Where AI earns its place is on both ends of the analysis: it compresses the cross-referencing of interview claims against thousands of order records from weeks into days, and it powers the fixes the analysis justifies, in this case automated payment instruction dispatch with an AI compliance gate, and a support layer grounded in live order data. AI is the how of the fix. Evidence is the why of the ranking.
Do I have to build with the firm that runs the assessment?
Not with us. The Blueprint and the system designs are the client's to keep, and they are written so any competent team can build from them, including the client's own. In this engagement the founder built the first two systems himself and had us design the third. We think that is what a diagnostic is for, and we priced the assessment so it stands on its own rather than as a foot in the door.
Where to take this
The full story, with the founder telling it in his own words on camera, is in the ecommerce revenue leak case study. The wider set of systems we have shipped, across support, outreach and full platform builds, is on the case studies page. If the checks in this post have never been run on your business, and you would rather find a leak before it surfaces on its own, the AI Strategic Assessment is the same process, run on your data with the people who ran it here. And if you are a coach rather than a store, the same evidence-first thinking applied to your operation lives in the AI automation playbook for coaches.
