This article is about a mistake we made on a project we are otherwise proud of. It is the cheapest mistake in digital transformation to avoid and one of the few that is genuinely impossible to correct afterwards.
The Sentence We Cannot Write
Gutta på Haugen is an Oslo delicatessen, trading since 1994, several branches, national shipping. Until we rebuilt it, the website was a form. A customer described what they wanted, an employee telephoned them back, completed the purchase verbally, took payment offline and arranged shipping. Every order consumed a member of staff.
The most compelling sentence that project could own would be something like: we replaced a process that cost X per order and capped throughput at Y orders a day.
We cannot write it. Asked how many orders a day the phone process could handle, or what it consumed in staff time, the honest answer on the record was:
“I don’t have insights on that.”
Nobody counted. There was no reason to — it was just how the shop worked. And now the process is gone, so the number cannot be recovered. Not with a bigger budget, not with better analytics, not ever. The system that would have generated the data has been deliberately destroyed, by us, as the whole point of the engagement.
That absence shows up in every artefact the project produced. The whitepaper has to cite market benchmarks for the cost of manual ordering — $30–80 per order against $1–5 integrated — clearly labelled as market data rather than as this client’s result. The results page has to state plainly that no pre-rebuild baseline figure exists, and warn readers not to read 3.1x year-over-year growth as growth from zero, because the shop was already making sales.
Everything we published is defensible. It is just weaker than it needed to be, permanently, for want of one week’s work at the start.
Why This Happens To Competent Teams
It is not carelessness. Four structural reasons, all of them reasonable in the moment.
The old process is not the interesting part. The energy at kickoff is on the future state. Nobody’s enthusiasm goes into documenting the thing everybody has agreed is broken.
It looks obviously bad already. When the current state is a form and a phone call, measuring it feels like proving that water is wet. The problem is that “obviously bad” is not a number, and in eighteen months a number is exactly what somebody will ask for.
The people who know are busy doing it. The staff who could tell you orders per day are the staff processing the orders. Asking them for a week of counting feels like an imposition on the exact bottleneck you are trying to relieve.
Nobody owns the before-state. The agency’s scope starts at design. The client’s scope starts at requirements. The measurement of a process that is about to be deleted belongs to nobody, so it does not happen.
Every other missing number in a project can be recovered later. This is the only one with an expiry date, and the expiry date is your own launch.
The Week of Work
It is genuinely a week, not a research programme. Five things, in descending order of value.
Orders per day, including the peak. Count for four weeks if you can, one if you cannot. Then find the busiest week of last year and count that from whatever records exist — inbox volume, invoice numbers, delivery notes. The peak figure is the important one, because that is where the throughput ceiling binds, and the ceiling is the argument for the whole project.
Minutes per order, timed. Not estimated by a manager. Timed, by the person doing it, from enquiry received to order entered. Ten orders is enough for a usable range. This single number multiplied by loaded hourly cost gives you your cost per order, which is the line that ends most internal debates about budget.
Error and re-ship rate. Count wrong items, wrong quantities, wrong addresses and credit notes for a month. Published benchmarks put manual entry error rates at 8–15%, and knowing where you actually sit against that is worth more than the benchmark.
Enquiries that never converted. The number nobody has. How many form submissions or calls did not become orders, and — if anyone remembers — why. This is the closest thing a manual business has to an abandoned-cart metric, and it is usually the largest single number in the exercise.
Time to response. Average and worst case, from enquiry to callback. Include nights, weekends and holidays honestly, because the customer experienced them.
One adjacent number worth capturing in the same week, for a different reason: how many gift cards and how much store credit are outstanding, and in what formats. That is a balance-sheet liability and a migration scope in the same figure, and it is the fastest way to discover that gift cards are an integration project rather than a feature.
If you only have one day rather than one week, do minutes per order and peak orders per day. Those two produce a cost per order and a throughput ceiling, which between them carry most of the business case.
What To Set Up on the Other Side
The mirror-image failure is a launch with no measurement plan, which produces a different unanswerable question: did the thing we built work?
Decide before launch which figures you will publish afterwards, and make sure something is recording them from day one:
- Cost per order in the new process, so the comparison to the manual figure is like-for-like
- Conversion rate by device and by traffic source
- Average order value, separately for the surfaces you built to raise it — shoppable recipes and bundles most obviously
- Repeat purchase rate and subscriber count, which is the only way to evidence replenishment work
- Blended ROAS with the spend base attached, for the reasons set out in what 4.24x blended ROAS actually looks like
- Support contact volume by reason, which is the quietest and most useful of the set — it tells you which of the counter’s jobs your interfaces did not manage to absorb
That last one deserves emphasis. In a business where staff used to answer questions in person, the support inbox after launch is a direct readout of which questions the new pages failed to answer. It is the continuation of asking what customers phoned up about, running permanently, for free.
Who Should Own It
Put it in the contract. Whether it sits with the agency or the client matters less than that it sits with somebody, has a named deliverable, and is dated before development starts. A one-page baseline document, signed off in week two, is the cheapest insurance in the project.
We would now argue for it as a precondition rather than a recommendation, and we say so in the whitepaper — a paper that only reports its successes is not worth reading.
The Honest Summary
The rebuild worked. The store replaced the phone call, the results are real, and the reasoning behind every design decision is documented screen by screen. What is missing is the ability to say how much better in the terms that would matter most to the next retailer facing the same decision — because the process we replaced was never counted, and it no longer exists to count.
If you are about to delete a manual process, spend a week measuring it first. It is the cheapest work in the project, the only work that becomes impossible later, and the difference between a business case and a strong hunch.
For the framework this sits inside — the eight commercial jobs a counter performs, and how to rebuild each one as an interface — start with the counter problem or read the whitepaper in full.
