Measuring Loss Across Hundreds of Sites
Shrinkage in quick commerce is a network measurement problem before it is a counting problem.
A chain of several hundred micro-fulfilment sites cannot be counted the way a warehouse is counted, because sending a team to every site on one date costs more than much of the loss being measured. What has to be complete is the high-value and high-theft population at every site, since those lines carry most of the money and are where deliberate loss concentrates. Everything else can be sampled, drawn so that each site is represented and each category is covered across the network.
The distinction that matters is between the site view and the network view. One site's variance says little on its own: a site with an unusual result may simply have had an unusual week. The same variance appearing at forty sites in one city, or at sites sharing a manager, a supplier or a shift pattern, is a finding.
Shrinkage here means the gap between the stock the system expects and the stock physically present, whatever the cause. Theft is only one cause, and often not the largest one.
Why Site-by-Site Counting Fails at Scale
Counting every site the way a warehouse is counted stops being workable long before a network reaches full size, and the reasons are arithmetic rather than organisational.
Reason one
Cost per site barely fallsEvery site needs a team, a date, access, a system extract and a reconciliation. That overhead hardly drops as sites are added, so total cost grows almost in step with the network while the stock value at each site stays small.
Reason two
Counts on different dates do not compareA network figure built from sites counted over several weeks describes no single moment, and every site-to-site comparison is distorted by the different trading each site saw. Stock turns fast enough in this format that a fortnight between counts makes two sites hard to compare.
Reason three
Travel outweighs countingSites are spread across a city or a country, and the time spent moving between them can exceed the time spent counting in them. The cost of each rupee of stock verified is far higher than in a conventional warehouse, and it rises as the network spreads.
None of this means sites should go uncounted. It means the count has to be designed for the network, with complete coverage kept for the lines where it pays and a deliberate sample for the rest.
Sampling a Network
Once complete coverage is impractical, the sample has to be built rather than drawn casually. Three criteria do most of the work:
- Volume. Weight the sample toward the sites that hold and move the most stock, so it describes where the money is.
A small-site sample describes the least of the network - Age. Include recently opened sites, where processes are unsettled, staff are new and accuracy is usually still improving.
Leaving them out understates the problem - Prior variance. Revisit sites that counted badly last time, to see whether the corrective action worked.
A revisit tests the fix, not the staff
Simultaneous counting matters wherever comparability does. Sites that will be compared with each other, or added into one network figure, are counted on the same date, so differences between them are differences in performance rather than in timing. Between those dated snapshots, each site's routine cycle count keeps the system close to the shelf; how to run one without stopping fulfilment is a separate subject.
What a sample can conclude is a statement about the network within stated limits. What it cannot do is certify any site that was not visited.
Separating Network Loss From Site Loss
The most useful thing a network programme produces is the line between problems that belong to the operating model and problems that belong to individual sites. Reading it wrongly wastes a great deal of management effort.
Reading a site result against the network
Where a result sits in the spread decides whether the answer is a process change or a site visit.
Systemic loss
- A similar level at most sites
- Points at picking, returns posting, packaging or system behaviour
- Fixed by changing the process, not by managing sites
Outlier sites
- Materially worse than the spread
- Points at supervision, layout, staffing or genuine loss
- Worth a visit rather than more analysis
Category effects
- A site holding more high-loss categories
- Looks worse in total without performing worse
- Seen only by comparing like categories
Systemic shrinkage is the hardest to accept because no single site is at fault. If every site runs the same picking routine, posts returns the same way and receives the same packaging, every site will show a similar loss, and replacing a store manager changes nothing. Pick accuracy and the speed at which returns are posted are usually where such a pattern starts.
Category effects are the trap in the middle. A site carrying an unusually high share of a high-loss category, such as fresh produce or small high-value items, will show worse total shrinkage without performing worse at all. Compare like categories across sites before comparing site totals.
Evidence That Supports a Network Number
A network shrinkage figure is only as good as the comparability of the counts behind it. Four things decide what it can be used for:
- Count dates are stated. Where one common date is impossible, the dates are listed and the result is presented as a figure for the period, not a position on a day: a weaker claim, but an honest one.
- Coverage is expressed as a share of network stock value, not as a number of sites. Many small sites covered can still leave most of the value unverified, so a high share of sites beside a low share of value is the pattern to watch for.
- The sample design is written down: how sites and lines were chosen, so a reader can judge what the figure represents.
- Verified and estimated figures are kept apart in the report, never added together into one number.
What an auditor will certify follows from all four. Verified positions at named sites on named dates can be certified. A network-wide shrinkage percentage projected from a sample is an estimate with stated assumptions, and presenting it as a verified figure asks the evidence to carry weight it does not have.
Running a Programme Across the Network
Choose deliberately between rolling coverage and periodic snapshots, because they answer different questions. Rolling coverage reaches every site over a period and is the only affordable way to touch a large network, but it produces no single dated position. Periodic snapshots count a smaller sample of sites on a common date, which supports comparison and aggregation but leaves most of the estate unvisited. Most networks need rolling coverage for control and a snapshot at the reporting date for the accounts.
Staffing and technology follow from that choice. Rolling coverage across hundreds of sites needs either a large field team or scanning tools that let site staff count reliably under supervision, and the second is only credible where those counts are checked independently on a sample.
An outside team is worth bringing in for the reporting-date snapshot, for any site where internal counts produce results nobody believes, and wherever the figure will be published or relied on by a lender or investor. A dark store stock audit planned across the network can then be scoped from the network down rather than one site at a time.
