Case study • US Direct-to-Consumer Ecommerce
We found the leak. He fixed it himself. That was the point.
Karim Dahmani asked us to find where AI and automation would pay back in his US direct-to-consumer ecommerce business. Seven years of his own order data showed a structural revenue leak he did not know he had. Within weeks, he had built the two highest-priority fixes himself.
The assessment at a glance
Every figure on this page is traced to the client's own order history, or labelled an estimate. That rule shaped the engagement, and it shapes this page.
Why Karim brought us in
He is a technically capable founder who can build most of what his business needs. What he could not get from inside the business was a defensible ranking of what to build first.
Revenue had fallen sharply
Karim put the fall down to a supply problem he had already solved, and expected volume to return. Whether the operation could carry that return was a different question, and nobody had asked it with data.
Support had slipped
The business was built on replies within minutes. First responses were now taking hours, and responsiveness complaints had reached public reviews for the first time in the company's history.
A backlog with no ranking
He had a long list of things he could build and no way to rank them. Every week spent building the wrong thing would be a week the leak stayed open.
What we did
Five phases, from public-footprint research to a scored Blueprint. Each phase narrows what the next one has to test.
Operational research
The public footprint, the technology stack and the operating constraints of the business, mapped before anyone was interviewed, so the interviews could test hypotheses instead of collecting first impressions.
Stakeholder interviews
Five structured interviews across the founder and the person running daily support and fulfilment. End-to-end workflows and root causes, mapped without leading the respondent.
Quantitative analysis
Seven years of orders, all 23,351 of them, reconstructed order by order: conversion trajectories, payment latency, stall rates by payment option, and how each moved over time.
Opportunity scoring
Seven opportunities scored on impact, feasibility, confidence and strategic alignment, then assigned to an implementation tier. The ranking is the deliverable, not a brainstorm.
Quality assurance and the Blueprint
Every qualitative claim cross-referenced against the quantitative record. ROI calculated on fixed formulas anchored to observed order history, with every assumption stated.
Two decisions worth naming
We started from the unfiltered order history
Cancelled and abandoned orders included, alongside the completed ones. A completed-orders report cannot show you what you are losing, because the losses are the rows it leaves out.
We modelled as if revenue never recovers
Karim told us volume would return, and we believed him. We modelled on the most recent measured quarter anyway, so that no recommendation depended on optimism. If volume recovers, the numbers improve. If it does not, they stand.
The cost from Karim's side: one data pull and the interviews. His words: "It was three meetings, and probably less than a week of my time."
What we found
One structural finding that changed what to build first, and two operational ceilings sitting underneath it.
One order in three never completed
Across the full history, 32.6% of mature orders entered the funnel and never converted. Karim suspected as much: before seeing any analysis, he had put the failure rate at 23 to 30% himself. The corroboration mattered, but it was not the finding.
The finding was in the trend line. The leak ran at roughly 28% through the strongest trading period the business ever had. It was structural, not a symptom of the downturn, and that reframes the economics: as order volume recovers, the same rate takes a larger absolute amount with it. Fixing it was worth more during a recovery, not less.
Across seven years, the orders that never converted came to roughly $1.7 million in gross intended value, a figure the report states plainly is not a recoverable balance and never was; the recoverable share was modelled separately and far more conservatively. A number that large is easy to put on a slide and hard to stand behind. We did the second thing.
Payment conversion diagnostics
Reconstructed from seven years of order-level payment behaviour. Options are unnamed on purpose: the fix adds no vendor, it closes the gap to the best-performing option already in the checkout.
never completed
Held at ~28% through 2022 to 2024, the strongest trading period the business has had.
Median time for a payment to clear
Manual instructions against automated confirmation
Manual runs roughly 4.7× slower on the median, and the tail is worse: a payment that needed reconciling over email could take past ten hours to clear, and 16% of manually paid orders took more than a day.
Stall rate across payment options already in use
Share of orders that never complete, by option
The spread inside his own checkout is the recovery ceiling: closing the gap to the best-performing option is a 34% reduction in stalled orders. We modelled the central case at a 20% reduction.
A throughput ceiling the business never used to have
A degraded integration between the store and the shipping side forced order-by-order manual handling, capping daily throughput roughly a third below the peak the business had already proven it could process. Not a ceiling it had historically lived within: a new one, created by the degradation, and it would bite exactly when volume returned.
Support had slipped from minutes to hours
The near-immediate response standard the reputation was built on had slipped to first responses measured in hours. The knowledge needed to answer sat with two people, written down nowhere, which makes the standard impossible to restore by effort alone.
What happened next
What got built from the Blueprint, and who built it.
He built the first two. We designed the third.
The Blueprint ranked four systems. For the two most urgent, the payment fix and the fulfilment integration, the fastest path was Karim building from our specifications, and both were live within weeks. Specs that a client can ship from that quickly are the quality bar the whole engagement is written to.
An assessment has to be honest about who is best placed to build each recommendation. Sometimes that is us. For these two it was the founder, because the engineering was already worked out on paper and he knows his own stack cold.
The third system, an AI support layer that answers routine customer questions in minutes from live order data, he handed to us to design in full, down to the policy engine that decides what it may handle on its own.
The Blueprint
Findings with the evidence behind them, seven scored opportunities, an ROI model anchored to the order history, an eighteen-month roadmap in three layers, and four systems sized and sequenced.
The project proposal
For the support system he chose: scope and exclusions in writing, a phased implementation plan with an acceptance checkpoint at the end of every phase, and a fixed investment.
The technical design specification
The working engineering document: architecture, data model, environments, integration contracts and acceptance checkpoints. His to keep, and his to hand to any build team.
In Karim's words
Pulled from the recorded testimonial at the top of this page.
"When I saw the report and what was being left on the table, I almost fell off my chair. I came in with just the expectation of the AI chat."
The moment it clicked
"The payment process was very confusing, it was very support intensive. Once I implemented these changes, it is literally almost zero."
After building the payment fixes
"You guys are not just selling smoke, you have real solutions. I've been working in IT for many years, and this is not something that is usual."
To anyone on the fence
Karim Dahmani, Founder. Quotes lightly trimmed for length, meaning untouched.
What an assessment would get you
The Blueprint with findings you can challenge line by line, a scored opportunity list, an ROI model anchored to your own numbers rather than industry averages, and the full solution architecture for one system of your choice, included. Four weeks end to end, and from your side a data pull and a few hours of interviews.